AI Tuesdays · Week 17 · RTF: Role, Task, Format

Warmup

The Workshop

Power Preview

Resources & Tools

Take Action

Week 17 of 52 · Self-Guided · ~25 min

0 starred 0 notes

Pre-Show · Warmup

Welcome back. Let us warm up first.

Three quick recall questions from last Tuesday's workshop on Few-Shot. Tap each to expand, tap again to close. Just for you, no scoring.

Answer: Two is the floor. Three is the sweet spot. Four or five is fine. Past six you are usually wasting time you spent finding them. One example is imitation, not pattern-matching, and AI often just rewrites the single example instead of learning the pattern behind it.

Example of use

Renewal outreach for [Property]. You paste three sanitized past emails that landed well. AI writes the fourth. Two examples would have worked. Six would have made AI mimic only the sixth.

Answer: Agree, Cover, Label, Sanitize. Every example points at the same pattern (agree). Show the variety AI will face (cover). Same format every time (label). No real resident data (sanitize).

Example of use

Teaching AI to classify service requests. Two examples labeled EMERGENCY (leak, no heat in winter). One labeled ROUTINE (dishwasher slow to drain). All labeled the same way. Names and apartment numbers stripped. AI now sorts new requests the way your team sorts them.

Answer: Examples with real resident details are still real resident data. Sanitize before you show. Your best past correspondence stays in your PMS. Brackets go in the Few-Shot examples.

Example of use

Wrong: pasting three past emails to Sarah Martinez in apartment 214 with balance $2,847. Right: replacing every real detail with [Resident], [Apartment], [Balance]. Same pattern. No data leak.

Missed last week?

Open Week 16 · Few-Shot before you start this one. Password on that page is 1652. This week builds on it.

The Workshop · Opening

Tell AI what NOT to do.

Three weeks in a row we told AI what to do. RTF told it the shape. Chain of Thought told it to show its work. Few-Shot showed it the pattern. This week we build the fence: the "do not" lines that stop AI from inventing statistics, hedging with "approximately," or drifting into Fair Housing traps you did not authorize.

This is Guardrails. Not a governance-team-approves-a-tool kind of guardrail. A prompt-line kind of guardrail. The negative instruction that keeps AI from doing the specific thing you just watched it do wrong.

This week in one line

Every prompt structure so far tells AI what to do. This one draws the line AI cannot cross.

Read this workshop as

Tap the role closest to yours. A few examples throughout will lean toward that context. Change your mind anytime.

The Workshop · Agenda

Six parts. About 25 minutes. No timer.

Self-guided. Pause whenever you need. Your notes and progress save to this device automatically.

What guardrails are

Slides 4 and 5. The definition and the anatomy of a good guardrail.

~4 min

When to reach for them and when to skip

Slides 6 and 7. High-stakes and outward-facing content: reach. Personal notes and brainstorming: skip.

~3 min

Four structures, one prompt

Slides 8 through 10. RTF plus CoT plus Few-Shot plus Guardrails, stacked. The bright line and the default rule.

~4 min

Power Preview: two examples

Slides 12 through 15. Renewal outreach without and with guardrails. Fair Housing guardrails on a property social caption.

~7 min

Reference and reinforcement

Slides 16 through 19. Vocab (46 terms), Guardrails reference card, standards, Spot the Tier.

~4 min

Quiz, downloads, and close

Slides 20 through 22. Five questions, two downloads, Week 18 teaser.

~3 min

No timer runs. The minute estimates are honest averages. Slower is fine.

The Workshop · Why You Should Care

Some AI output looks fine. Right up until it goes out.

Two things happen every week in this industry that guardrails prevent. Neither one looks like an emergency in the moment. Both are avoidable with one added line to your prompt.

Scenario 1

The stat AI made up

A leasing pro asks AI to draft renewal outreach for a resident whose rent is going up 4 percent. AI writes: "This is still approximately 17 percent under current market for comparable apartments in your area." The number is invented. AI has no market data. Someone had to catch it before the email went. Sometimes nobody catches it. One added guardrail: "Do not include any statistic or percentage you cannot cite from the source data I provide. If you do not have the number, say I do not have that data."

Scenario 2

The Fair Housing line AI crossed

A marketing associate asks AI to write a social caption for the property's new playground. AI writes: "A perfect place for young families to make memories." That is familial-status language, and it is a Fair Housing violation waiting to be screenshot. One added guardrail: "Do not use language that suggests a preference for or against any age group, family structure, religion, or other protected class. Describe features and amenities, not who they are for."

Here is the thing

Neither scenario is a hallucination story. Both are AI doing exactly what it is trained to do: sound helpful and human. Guardrails are how you make AI stop sounding helpful about the specific things you do not want it helpful about.

The Workshop · Teaching A

Guardrails in one line.

The definition

Guardrails are the "do not" lines you add to a prompt. Not what AI should do. What AI must not do.

Every prompt structure so far tells AI what to do. RTF, Chain of Thought, Few-Shot. Guardrails do the opposite. They tell AI what to leave out, what to refuse to do, and what to say when it does not know. Positive instruction has a ceiling. Negative instruction is the only way to hit certain kinds of quality: no invented statistics, no hedge words, no Fair Housing traps, no accidental price commitments.

Without guardrails

Prompt

You are a leasing pro. Draft a renewal outreach email for a resident on a 4 percent rent increase. Warm but direct.

What AI writes

"Your renewal is coming up. Even with the modest increase, you are still approximately 17 percent under current market. We would love to lock you in for another year at this competitive rate..."

Problem: AI invented "17 percent under market." It has no market data. AI hedged with "approximately." It committed you to a "competitive rate" you did not authorize.

With guardrails

Prompt (same task, four added lines)

You are a leasing pro. Draft a renewal outreach email for a resident on a 4 percent rent increase. Warm but direct. Do not include statistics you cannot cite from data I gave you. Do not use hedging words like approximately or roughly. Do not commit to any price, discount, or promotion. Say "let me confirm" when the answer is not on this page.

What AI writes

"Your renewal offer is ready to review. Rent moves from $[Current] to $[New], a 4 percent adjustment. If you have questions about anything on the offer or want to walk through it in person, let me confirm a time that works..."

Result: No invented number. No hedge. No commitment you cannot honor. Same warm tone. Ready to send.

Pay attention

The word "guardrail" already lives in your governance vocabulary: a rule inside your prompt or your workspace that limits what AI can do. This week is about the prompt-line side of that. Same word. Same idea. Just the piece of it you can write yourself, right now, without waiting for a policy update.

The Workshop · Teaching B

Anatomy of a guardrail.

Not every "do not" line works equally well. Four rules separate a guardrail AI actually follows from one AI silently ignores.

Rule 1 · Be specific

Weak: "Be accurate."

Strong: "Do not include any statistic, percentage, or dollar figure you cannot cite from the source data I provided."

The specific version names the exact behavior AI will otherwise default to. "Be accurate" is a wish. The specific version is a rule.

Rule 2 · Pair positive with negative

Weak: "Do not guess."

Strong: "Say 'I do not have that information' instead of guessing."

Negative-only leaves AI without a fallback lane. Positive-plus-negative gives AI something to do instead. This is the single most reliable guardrail pattern.

Rule 3 · Keep the list short

Weak: Nine guardrails in one prompt.

Strong: Three at a time. Ceiling.

Past three, AI starts hedging everything to avoid tripping any of them. Output goes to mush. Pick the three highest-leverage rules for this task. Save the rest for next time.

Rule 4 · Layer them over time

Weak: Guess at every possible guardrail you might need.

Strong: Start with one or two. When AI produces bad output, add the rule that would have prevented it.

Your guardrail list should grow the way your team's brand voice guide grew: from real mistakes, not from theory. Save the working versions in a folder. Reuse them by task type.

Trust me

Rule 2 is the one everyone skips and the one that matters most. "Do not X" alone gives AI a fence with no gate. "Do not X. Say Y instead." gives AI a fence and a gate. AI follows the version with the gate almost every time.

The Workshop · Teaching C

Five times to reach for guardrails.

Anywhere AI output has a downstream reader who will judge you by it, add guardrails.

Anything going to a resident

Renewal outreach, delinquency notes, move-in welcomes, community newsletters. If the recipient will save the email, forward the email, or quote it back to you in three months, guardrails.

Anything with numbers you cannot verify on the spot

Market comparisons, occupancy references, turnover benchmarks, submarket rent averages. If AI does not have the data and you cannot check it in ten seconds, guardrail it: "Do not include statistics you cannot cite. Say 'I do not have that data.'"

Any Fair Housing-sensitive content

Social captions, property descriptions, amenity write-ups. Anywhere language could suggest a preference by age, family structure, religion, disability, or any protected class. Guardrail once, use forever.

Vendor, legal, or leadership communications

Anywhere hedging is a liability. "We can probably start next week" and "this should reduce turnover by roughly 15 percent" both create expectations you may not be able to meet. Guardrail: "Do not use hedge words. State what you know and mark what you do not."

Anything going into a system without human review

Automated workflows, embedded chat responses, generated FAQs, batch content. When the output does not stop at a human before it hits the reader, guardrails are the only check between AI and your resident. Layer them thick.

The Workshop · Teaching D

Four times to skip guardrails.

Guardrails narrow AI output. Sometimes narrow is the enemy of what you actually need.

Personal notes to yourself

You are asking AI to sketch a game plan for your Monday. Nobody but you will read it. Adding guardrails just slows you down. Skip.

Drafts you will review anyway

You are going to read every line before it goes out. You will catch the invented stat yourself. Guardrails still help, but they are not doing critical work. Skip if you are pressed. Add if you are training a habit.

Brainstorming and idea generation

You want ten angles for a resident event, five subject line options, twelve possible names for an internal contest. Guardrails narrow output. Brainstorming needs the opposite. Skip.

Simple factual lookups

"What is the capital of Texas" does not need guardrails. If the task is short and the answer is verifiable in a second, skip. Guardrails are for the tasks where you cannot verify.

Here is the thing

The temptation with guardrails is to add them to everything once you learn them. Do not. They cost tokens. They narrow output. Save them for the tasks where the downside of AI going wrong is bigger than the effort of adding four lines to your prompt.

The Workshop · Teaching E

Four structures, one prompt.

RTF gave you the shape. Chain of Thought gave you the reasoning. Few-Shot gave you the pattern. Guardrails give you the constraints. When all four stack, you get an answer that is the right shape, arrived at by the right logic, matching the pattern your team uses, without the specific things you did not want.

Task: renewal outreach for a resident on a 4 percent rent increase

Role: You are a leasing pro at [Property] with three years of tenure with this resident.

Task: Draft a renewal outreach email. Rent moves from $[Current] to $[New]. Offer is valid through [Date].

Format: Under 150 words. Warm sign-off.

Chain of Thought: Before you write, think step by step about what would matter to a three-year resident learning about a 4 percent increase. State what you are assuming about the relationship, then write.

Few-Shot: Match the pattern of these three renewal emails our team sent last year:

Example 1: [sanitized past email]

Example 2: [sanitized past email]

Example 3: [sanitized past email]

Guardrails: Do not include any statistic or percentage you cannot cite from data I gave you. Do not use hedge words like approximately, roughly, or about. Do not commit to any price, discount, or promotion. When something is uncertain, say "let me confirm."

Notice what each structure contributed. RTF told AI who to be and what to write. Chain of Thought made it walk its logic first. Few-Shot fed it your team's voice. Guardrails kept the invented statistics, the hedge words, and the accidental price commitments out. Four passes at the same task, each closing a specific failure mode. That is what stacking is.

The Workshop · Teaching F

The bright line: guardrails do not sanitize.

Every week has one line that must not blur. This one is about what guardrails do not do.

Wrong

The confidence trap

A leasing pro pastes into a general AI tool: "Sarah Martinez in apartment 214 is 60 days delinquent on $2,847. Draft an outreach email. Guardrail: do not use language that suggests we are threatening eviction."

What went wrong: The guardrail is fine. The real resident name, apartment number, and balance are still Tier 3 data that just went into a general AI tool. The guardrail constrained what AI wrote. It did not sanitize what was pasted in.

Right

The two-step

Same task, brackets in place: "[Resident] in [Apartment] is [X days] delinquent on $[Balance]. Draft an outreach email. Guardrail: do not use language that suggests we are threatening eviction. Say 'work with you on options' when discussing next steps."

What went right: Same guardrail, same output quality. Real details never left the property management system. Guardrails do their job (shape what AI writes). Sanitization does its job (protect what AI reads).

The bright line

Guardrails constrain what AI writes. They do not sanitize what you paste in. Real resident data still stays out.

Read that line twice. This week's topic sounds like safety. It sounds like the thing that protects you from putting the wrong data in. It does not. Guardrails work on AI output. Sanitization works on AI input. Two different jobs, both required, neither replaces the other. Bracketed data. Guardrailed prompts. Human review. That is the full stack.

The Workshop · Default Rule

Shape, Reason, Pattern, Constraints.

Four Phase 3 structures. Four different jobs. Do not choose between them. Use the ones the task needs.

Shape

RTF

Role, Task, Format. The workhorse. Every prompt starts here.

Reason

Chain of Thought

Add when the task has real math, comparisons, or judgment. Ask AI to think step by step.

Pattern

Few-Shot

Add when you want AI to match your team's voice, format, or category rule. Paste two to four sanitized examples.

Constraints

Guardrails (this week)

Add when the output goes to residents, leadership, or straight into a system. Set the "do not" lines.

Use RTF alone for a personal note. Use RTF plus Guardrails for a resident email. Use all four when the task is high-stakes and outward-facing. The stack is additive. You are not choosing. You are layering.

The Workshop · Q & A

Six questions people ask about guardrails.

Tap any question to expand the answer. Tap again to close.

Answer: Not refuse, but hedge everything. AI starts adding "pending confirmation" and "subject to review" to every sentence to avoid tripping any rule. Output goes to mush. That is why the ceiling is three guardrails per prompt. Above that, you lose the value of the ones you actually need.

Example of use

A renewal email with three guardrails (no invented stats, no hedge words, no price commitments) reads clean. The same email with nine guardrails reads like a legal disclosure. Pick the three that matter most for the task.

Answer: No. Build a small library, task by task. Renewal outreach guardrails. Delinquency guardrails. Social caption guardrails. Save each set in a note or a folder called "Guardrails." Paste them in when needed. Add to the set when a new bad output shows a gap.

Example of use

Your "Social caption guardrails" note has three lines: no age or family-status language, no religious references, no assumptions about who lives here. You paste that block into every caption prompt. Consistency and Fair Housing coverage in seven seconds.

Answer: Chain of Thought is positive: "Show your reasoning." Guardrails are negative: "Do not include invented statistics." CoT shapes how AI thinks. Guardrails shape what AI leaves out. Different jobs, and they stack in the same prompt without conflict.

Example of use

Vendor comparison prompt uses both: "Think step by step through the annual cost of each bid" (CoT). "Do not include recommendations you cannot support from the numbers I gave you" (Guardrail). Reasoning is transparent. Output is bounded.

Answer: In practice, the pattern that transfers most reliably across the mainstream tools is the one that follows the four rules on Slide 5: specific, positive-paired-with-negative, short list, layered. I am not certain any single guardrail behaves identically across ChatGPT, Claude, Gemini, and Copilot on every task, and behavior can change with tool updates. So the honest answer is: test every specific guardrail in the tool your team actually uses, and keep the pattern portable even when the wording is not.

Example of use

"Do not use hedge words like approximately, roughly, or about" works the same in every mainstream tool. "Be careful" does not. Specificity is the transferability.

Answer: Nine times out of ten, adding the "say Y instead" part. Bare "do not X" is where guardrails fail. Give AI something to do instead of the thing you banned. That is the positive-negative pairing on Slide 5, Rule 2.

Example of use

"Do not guess" alone gets ignored. AI still guesses in soft language. "Do not guess. Say 'I do not have that information' instead." gets followed. Same rule, gate added, works.

Answer: No. Guardrails reduce risk. They do not eliminate it. AI still makes mistakes inside the fence. RETTC calls for a human review anywhere AI output reaches a resident. Guardrails plus sanitized input plus human review. All three, always, for outward-facing work.

Example of use

A renewal email drafted with all four structures still passes across a leasing pro's screen before it sends. AI writes most of it right. The human catches the small share no guardrail could have caught: the resident just lost a family member, and the warm tone needs to be gentler still.

Power Preview · Setup

Two runs. Same prompt each time. First without guardrails, then with.

You are about to see two side-by-side examples. Each shows AI output before and after adding a small set of guardrails. Watch what changes.

Run 1

Renewal Outreach With Guardrails

The invented statistic. The hedge word. The accidental price commitment. Watch three guardrails eliminate all three in one pass.

Run 2

Fair Housing Guardrails on a Property Social Caption

A family-friendly amenity caption AI wrote three ways. All three sound fine. Two are Fair Housing violations. Guardrails know the difference.

Read both, then Try This Today on slide 15 walks you through building guardrails for a real task in your job.

Power Preview · Run 1

Renewal outreach: without, then with.

Same resident. Same rent increase. Same request. Only the guardrails are new.

The zero-guardrail prompt (Run 1a)

"You are a leasing pro. Draft a renewal outreach email. Rent is going up 4 percent to $1,560 for [Resident] on a 12-month renewal. Warm but direct."

What AI writes back (excerpt)

Hi [Resident], your renewal is coming up. Great news: even with this modest adjustment, you are still approximately 17 percent below current market for comparable apartments in your submarket. We would love to lock you in for another year at this competitive rate, and we may be able to offer a small concession if you sign early. Let us know if you have any questions...

Three problems: the 17 percent number is invented (AI has no market data). "Approximately" hedges a stat you did not authorize. "A small concession if you sign early" is a promise your leasing pro cannot make. Any one of these could ship. Sometimes all three do.

Same prompt plus three guardrails (Run 1b)

"You are a leasing pro. Draft a renewal outreach email. Rent is going up 4 percent to $1,560 for [Resident] on a 12-month renewal. Warm but direct.
Guardrails: Do not include any market statistic, percentage, or comparison you cannot cite from data I gave you. Do not use hedge words like approximately, roughly, or about. Do not offer or imply any concession, discount, or promotion. Say 'let me confirm' when the answer is not on this page."

What AI writes back (excerpt)

Hi [Resident], your renewal offer is ready to review. Rent moves from $[Current] to $1,560 for a 12-month renewal, a 4 percent adjustment. If you have questions about anything on the offer or want to walk through it in person, let me confirm a time that works. Looking forward to another year...

What is gone: no invented statistic, no hedge word, no unauthorized concession. What stayed: the warm tone, the real information the resident actually needs. Same tool. Same task. Four added lines.

Power Preview · Run 2

Fair Housing guardrails on a social caption.

A marketing associate asks AI for three social captions for the property's new playground. Read all three. Two of them are Fair Housing violations. Guardrails know which.

The zero-guardrail prompt (Run 2a)

"Write three social media captions for [Property]'s new playground. Warm, community-focused, short."

What AI writes back

Caption 1: "A perfect place for young families to make memories. Our new playground is open!"

Caption 2: "New at [Property]: a playground the kids will love. Come see what makes this place feel like home for families like yours."

Caption 3: "Open now: a bright new play area next to the pool. Shaded seating for guardians. Rubberized surface. All ages welcome."

Two out of three are violations. Caption 1 says "young families": familial status. Caption 2 says "the kids will love" and "families like yours": familial status. Only Caption 3 describes the amenity without steering by protected class. Which caption does the marketing associate pick? Whichever sounds warmest, unless something stops them.

Same prompt plus Fair Housing guardrails (Run 2b)

"Write three social media captions for [Property]'s new playground. Warm, community-focused, short.
Guardrails: Do not use language that suggests a preference for or against any age, family structure, religion, disability, or other protected class. Describe the amenity and its features, not who it is for. Say 'all residents welcome' if you need a signal about openness."

What AI writes back

Caption 1: "Open now: our new playground next to the pool. Rubberized surface, shaded seating, and lots of room to move. All residents welcome."

Caption 2: "Fresh at [Property]: bright new play equipment, soft-fall surfacing, and benches under shade. A new spot on property, all residents welcome."

Caption 3: "Something new at [Property]: a play area with modern equipment, safety surfacing, and seating for anyone who wants to hang out. Open every day."

All three are usable. No steering language. No familial framing. The amenity is described. The audience is not. Same warmth. Same AI. One added guardrail line, saved once, pasted into every caption prompt from here on.

Trust me

Save this guardrail block once. Title the note "Fair Housing caption guardrails." Paste it into every marketing prompt from now on. This is how compliance stops being a review step and starts being a default.

Power Preview · Try This Today

Try This Today: 12 minutes to your first guardrail set.

Not next week. Today, before you close this tab. Twelve minutes and you will have a saved guardrail block you can reuse forever on one specific task.

Pick a task you sent to AI in the last week that ended up wrong (3 min)

Not the worst one. The most typical one. A renewal email, a delinquency note, a social caption, a vendor inquiry. Something you do more than once a month.

Read the output. Circle 2 or 3 things AI did wrong (3 min)

Invented a number. Used "approximately." Committed to a discount. Used age or family-status language. Named a policy you never wrote. Whatever it was, three specific things you would have caught if you were paying attention.

Write 2 guardrails in the "do not X. Say Y instead" pattern (3 min)

One for each thing AI did wrong. Specific. Positive-paired. Under 25 words each. Look at the four rules on Slide 5 while you write. If you have three things you want to fix, pick the two most important. Save the third for next week.

Run the same task with your new guardrails added (3 min)

Same tool, same prompt, guardrails at the bottom. Read the output. Did AI stay inside the fence? If yes, save the guardrail block in a note called "Guardrails · [Task Type]." Paste it into every future prompt for that task. That is the whole practice.

Good job

Twelve minutes and you have a reusable guardrail set for one task. Next week, do this for a second task. In two months, you will have a library of guardrail sets that runs quietly in the background of every prompt you send.

Reference · Bookmark this slide

Vocabulary cheat sheet: 46 core terms.

Every AI term this curriculum has introduced so far, in one place. Plain-language definitions. This is the same term list on the PDF download at the bottom of this slide and on Slide 21.

Foundational (from Weeks 8-10)

Generative AI

AI that creates new content (text, images, audio, code) based on training data patterns.

Large Language Model (LLM)

The kind of AI behind ChatGPT, Claude, Gemini, and Copilot. Trained on vast text and predicts one word (token) at a time.

Token

A small chunk of text an LLM reads and writes. Roughly 0.75 words per token; 1,000 tokens is about 750 words.

Training Data

The vast collection of text an LLM learned patterns from. Not a searchable database, a compressed pattern archive.

Prediction

How AI decides what to say next: the most likely word based on the pattern of what came before, one word at a time.

Hallucination

AI producing a confident, fluent answer that is factually made up. Especially common when it does not know something you did not tell it.

Sycophancy

AI agreeing with what you just said instead of pushing back. Feels helpful, is often wrong.

Prompt anatomy (from Week 11)

Prompt

What you type to AI. Your input. The one part of the container you fully control.

System Prompt

A hidden instruction the AI tool sends the model before your prompt. Shapes tone, role, and constraints. You do not write it and cannot see it.

Conversation History

Every earlier message in the current chat, yours and AI's. AI re-reads all of it every turn. This is how AI "remembers."

Context Window

The ceiling on how much text AI can read in one conversation. When it fills up, the oldest content silently drops off.

Prompt Anatomy

The five parts of a well-formed prompt: instruction, context, input data, output format, examples.

Output Format

The shape you want the answer in: length, structure, tone. One line in your prompt saves you rewriting the whole reply.

Prompt quality (from Week 12)

Specificity

The single biggest lever on prompt quality. Naming the audience, the format, and the constraints so AI is not guessing.

Ambiguity

What makes AI fall back on average defaults. The parts of a prompt AI has to guess. Iteration cannot rescue what specificity should have prevented.

Iteration (from Week 13)

Iteration

Refining a prompt after reading the answer, in the same chat. Your first prompt is a draft. Your second, written on purpose, is almost always better.

Refinement

The act of sending a targeted follow-up: add what was missing, correct what was wrong, or show an example of right. One deliberate change beats a full rewrite.

The Three Moves

ADD (give it the context it did not have), CORRECT (tell it what to stop doing), SHOW (give it an example of right). Every refinement is one of these three.

Diminishing Returns

When each new round of refinement barely changes the answer. Your signal to stop prompting and finish the piece by hand.

Loop

The Ask → Read → Refine → Stop cycle. The four-step rhythm that turns first-answer disappointment into a usable draft.

RTF (from Week 14)

Prompt Structure

A repeatable pattern for writing prompts. RTF is a prompt structure. Chain of Thought is another. Seven of them are taught in Phase 3 (Weeks 14 through 20), each for a different kind of job.

RTF

Role, Task, Format. The workhorse prompt structure. Three parts in one line that make the first prompt land closer to what you actually needed.

Role (in RTF)

The first line of an RTF prompt. Tells AI who it is being right now. "You are a friendly on-site property manager." Changes what AI draws on and how it talks.

Task (in RTF)

The second line. What AI is actually doing. A verb, an object, and enough detail that two people could not interpret it differently. Not a topic.

Format (in RTF)

The third line. What shape the answer should take. Length, structure, tone. The part almost everyone skips, and the reason so many AI answers feel "close but not quite."

Chain of Thought (from Week 15)

Chain of Thought (CoT)

A prompt structure where you tell AI to show its reasoning step by step before it gives the final answer. Improves accuracy on tasks with real math, comparisons, or judgment. Costs a few extra seconds and a few extra tokens.

Reasoning

The steps between the question and the answer. AI can produce reasoning that sounds correct and is not, which is why CoT works: you get to check the steps against reality before you act on the conclusion.

Step-by-Step

The most common CoT trigger phrase. "Think step by step" added to any RTF prompt turns on the reasoning walk-through. Simple, portable across every AI tool.

Show Your Work

The everyday-English version of CoT. What a teacher asks a student. When you ask AI the same thing, you get the same benefit: visible arithmetic, visible assumptions, catchable errors.

Few-Shot (from Week 16)

Few-Shot Prompting

A prompt structure where you give AI a small set of examples of the answer you want (usually two to four), then ask AI to produce a new one that matches the same pattern. The examples do the teaching. Faster than describing a voice, format, or category rule in words.

Zero-Shot Prompting

The default: you describe the task and give no examples. What most people do without thinking about it. Fine for factual questions and quick edits. Weak when you want AI to match a specific voice or pattern.

Example (in prompting)

A concrete sample of the answer you want. In Few-Shot, the examples together form the pattern AI will match. Consistent, sanitized, and labeled the same way every time. Also called a "shot."

Pattern Matching

What AI does when you give it examples: it looks for what stays constant across them and reproduces that constancy in its new answer. This is the mechanism Few-Shot relies on. If your examples do not share a pattern, there is nothing to match.

This week · Guardrails (Week 17) *NEW

Negative Instruction

A "do not" line in a prompt. The core building block of a guardrail. Tells AI what to leave out, refuse to do, or fall back on when it does not know. The opposite of positive instruction, which tells AI what to do.

Positive/Negative Pairing

The most reliable guardrail pattern: "Do not X. Say Y instead." The negative side sets the boundary. The positive side gives AI a lane to move into. Negative-only leaves AI without a fallback and often gets ignored.

Hedge Word

Softener words AI uses when it is uncertain: approximately, roughly, about, kind of, sort of, probably. Fine in personal notes. A liability in resident, leadership, or vendor communications, because they turn into commitments the reader remembers as facts.

Governance and data (running through the curriculum)

Data Classification

Sorting a piece of information by which AI tool it is allowed to enter. Tier 1 (safe for any AI), Tier 2 (enterprise only), Tier 3 (never in AI). Full treatment: Week 1.

Tier 1 · Safe

Public or fully de-identified information. Any approved AI tool is fine.

Tier 2 · Enterprise Only

Sensitive but useful information. Only in an approved enterprise AI workspace with data protections in place.

Tier 3 · Never

Information that never enters an AI tool. Real names tied to sensitive details, financial identifiers, protected-class data, and screening or eviction information.

Use Case

A specific way you plan to use AI: who is doing it, what task, what data goes in, what comes out. The Council approves use cases, not tools.

Guardrail

A rule that limits what AI can do. Two forms: (1) governance guardrails are rules built into your workspace or Council-approved use case; (2) prompt guardrails are the "do not" lines you add inside a single prompt (Week 17). Same idea, different scope. The prompt-line version is what a leasing pro can write today without a policy update.

Human-in-the-Loop

A person reviews AI output before it goes to a resident or a decision. The RETTC principle behind almost every safe use case.

AI Council

The group in your organization that approves use cases, reviews new tools, and owns the response when something goes wrong. Full treatment: Week 7.

NIST AI RMF

National Institute of Standards and Technology AI Risk Management Framework. The four functions: GOVERN, MAP, MEASURE, MANAGE.

RETTC AI Governance Framework

Real Estate Technology and Transformation Center. Eight principles specific to rental housing. Detailed on Slide 18.

💾 Download AI Vocabulary Cheat Sheet (PDF)

Reference · Bookmark this slide

Guardrails, on one page.

Read the left column when you are writing a prompt. Read the right when you want to see it in the wild.

The template

Write your RTF prompt first. Add CoT or Few-Shot if the task needs them.

Add up to three guardrails at the end. Use the "do not X. Say Y instead" pattern.

"Guardrails: Do not [specific behavior]. Say [alternative] instead. Do not [second specific]. Say [second alternative] instead."

The four rules for a good guardrail are on Slide 5: be specific, pair positive with negative, keep the list short, layer over time.

Filled in

Role: You are a leasing pro at [Property].

Task: Draft a renewal outreach email. Rent moves from $[Current] to $[New].

Format: Under 150 words. Warm sign-off.

Guardrails:

- Do not include statistics or percentages you cannot cite from data I gave you. Say "let me confirm" instead.

- Do not use hedge words (approximately, roughly, about). State what you know exactly.

- Do not offer or imply any concession, discount, or promotion. Say "let me confirm what the offer includes" instead.

When to reach for it

Use Guardrails for any output going to a resident, leadership, vendor, or straight into a system. Anywhere Fair Housing language is a risk. Anywhere numbers matter. Anywhere hedging is a liability.

Skip Guardrails for personal notes, drafts you will review anyway, brainstorming (which needs the opposite), and simple factual lookups.

Watch for these traps

Thinking guardrails sanitize input. They do not. A guardrail-heavy prompt still leaks resident data if you paste unbracketed names, apartment numbers, or balances. Guardrails work on output. Sanitization works on input. See the bright line, Slide 9.

Stacking more than three. Nine guardrails make AI hedge everything. Output goes to mush. Ceiling is three per prompt. Save the rest for next time.

Negative-only guardrails. "Do not use bullet points" without a positive alternative gets ignored. Pair every "do not X" with "say Y instead" and AI follows almost every time.

The bright line, always

Guardrails constrain what AI writes. They do not sanitize what you paste in. Real resident data still stays out. See Slide 9.

💾 Download Guardrails Quick Reference (PDF)

Permanent · Standards Library

NIST AI Risk Management FrameworkNIST AI RMF (NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY, AI RISK MANAGEMENT FRAMEWORK)
The US federal framework for managing risk in AI systems. Four functions: GOVERN, MAP, MEASURE, MANAGE. Voluntary but widely adopted.

Multifamily example: A regional manager wants prompt guardrails standardized across every property's renewal-outreach workflow. The AI Council walks NIST's four functions: GOVERN (who approves the guardrail block for that task type), MAP (which failure modes matter most: invented stats, hedge words, unauthorized concessions), MEASURE (spot-check outbound emails against the guardrails for a month), MANAGE (revise the block when a new failure mode shows up). Same framework as Weeks 1 through 16. Different anchor. Guardrails now sit inside a governed workflow.

Full document at nist.gov/itl/ai-risk-management-framework & RETTC 8 PrinciplesRETTC AI GOVERNANCE FRAMEWORK (REAL ESTATE TECHNOLOGY AND TRANSFORMATION CENTER)
Eight principles for AI governance in rental housing. Anchored to Fair Housing, resident privacy, and human accountability.

Multifamily example: A property adds a Fair Housing guardrail block to every marketing prompt. RETTC's Promote Fairness & Consumer Protection principle says the operator is responsible for outputs that comply with the Fair Housing Act. The guardrail block is one specific way to operationalize that responsibility inside the workflow, not just inside the policy binder. Belt and suspenders: guardrail in the prompt, human review after, both required.

Published at rettc.org/frameworks/ai-governance.

Two frameworks. The NIST AI Risk Management Framework (AI RMF) from the National Institute of Standards and Technology, and the AI Governance Framework from RETTC (Real Estate Technology and Transformation Center, at rettc.org). US-anchored. Together they cover the risk side (NIST) and the operating side (RETTC) of governing AI in rental housing. Every Week 17 concept about Guardrails lands somewhere in one or both.

The "do not" lines you add to a prompt

GOVERN function: policy defines which negative instructions belong in prompts for which task types. A sanctioned guardrail block is a governed asset, not a personal preference.

Establish an Organizational Philosophy on AI Use and Oversight: guardrails that reflect team standards are a lever for the philosophy, not a workaround around it.

Preventing invented statistics and hedge words

MEASURE function: track whether AI output stayed inside the fence you set. If a guardrail is being ignored, you cannot fix it without measuring how often it is being ignored.

Uphold Transparency & Explainability: outputs going to residents should reflect what the operator actually knows and has authorized. Guardrails keep AI from filling gaps with confident-sounding invention.

Fair Housing guardrails on marketing content

MAP function: identifying the failure modes that matter for this task. Familial-status language and other protected-class steering are the specific risks marketing prompts create.

Promote Fairness & Consumer Protection: operators are responsible for outputs that comply with the Fair Housing Act. A guardrail block is one specific way to operationalize that responsibility inside the prompt itself.

Saved guardrail blocks reused by task type

GOVERN function: which guardrail block becomes the standard for which task type is a governance decision, not a personal preference. Whose blocks get elevated matters.

Protect Privacy & Promote Data Integrity: standardized guardrail workflows reduce the surface area where sensitive data could leak, and keep AI output aligned with data-minimization principles.

The bright line (guardrails do not sanitize)

MANAGE function: the incident response plan covers what happens when someone assumes a guardrail also sanitized the input, and pastes real resident data. Rehearse it before it happens.

Maintain Outcome-Based Accountability & Human Involvement: the person who chose to paste the input owns the outcome. Guardrails plus review, not either alone.

Establish an Organizational Philosophy on AI Use and Oversight. Adopt a clear philosophy that defines how AI will be used, governed, and monitored across operations.

Promote Fairness & Consumer Protection. Ensure AI systems comply with the Fair Housing Act, Fair Credit Reporting Act, Americans with Disabilities Act, and state and local law.

Uphold Transparency & Explainability. Promote a general understanding of when AI influences key decisions, and provide clear explanations to renters and applicants of decisions affecting them as required by law.

Protect Privacy & Promote Data Integrity. Adhere to federal, state, and local consumer privacy laws, with an emphasis on data minimization, security, and informed consent.

Maintain Outcome-Based Accountability & Human Involvement. Engage humans in significant decisions, establish processes to review model outcomes, provide escalation channels for complaints and appeals, and educate staff on AI use and limitations.

Promote Positive Renter Experiences and Outcomes. Elevate technologies that enhance affordability, access, and housing stability, and use AI to improve resident experiences.

Strive for Innovation with Responsibility. Encourage development of AI tools that benefit residents, prospects, guests, and operators, without compromising trust.

Encourage Third-Party Partner Accountability and Due Diligence. Incorporate robust due diligence and risk management controls for third-party AI tools. Invite transparency from vendors regarding data sources, model limitations, and known risks.

Read the source

NIST AI Risk Management Framework: nist.gov/itl/ai-risk-management-framework
RETTC AI Governance Framework: rettc.org/frameworks/ai-governance
RETTC launch press release: rettc.org/press-release/rettc-unveils-first-of-its-kind-ai-governance-framework-for-rental-housing-and-technology-partners

Reinforcement · Spot the Tier

Spot the Tier: 8 Guardrails scenarios.

For each scenario, tap the tier you think it belongs to. Feedback shows immediately with the correct tier and the reason. Reset the whole set with the button at the bottom.

Scenario 1

You paste sanitized team emails into an AI tool with three guardrails: no invented stats, no hedge words, no unauthorized concessions. AI drafts a renewal email that reads clean.

Green. Sanitized examples plus a bounded guardrail block. This is exactly the RTF+Few-Shot+Guardrails stack the workshop teaches. Safe use in any approved tool.

Not quite. Everything sensitive is stripped from the examples, and guardrails constrain the output. Nothing here needs an enterprise workspace. Green.

Scenario 2

You paste unsanitized past emails to Sarah Martinez in apartment 214 with balance $2,847 into ChatGPT, and add a guardrail: "Do not use language that threatens eviction."

Red. This is the bright line failure. The guardrail shaped what AI would write. It did not sanitize what was pasted. Real name, real apartment, real balance, all now in a general AI tool. Guardrails do not sanitize.

Not quite. The guardrail is fine. The input is the problem. Real resident name plus apartment number plus a delinquency balance in a general AI tool is Tier 3. Red.

Scenario 3

A marketing associate saves a "Fair Housing caption guardrails" block: no age/family-status language, no religious references, describe amenities not audience. They paste it into every social prompt.

Green. This is the practice the workshop pushes: save the block once, reuse it forever, and let the guardrail run in the background of every marketing prompt. No sensitive data in play.

Not quite. There is no resident data anywhere in this workflow. The guardrail block itself is public content. Green.

Scenario 4

You paste an ADA accommodation request that includes the resident's condition (mobility impairment) and provider's name into a general AI tool, with guardrails to keep the tone respectful.

Red. Disability status and provider names are Tier 3 the moment they enter a general AI tool. The guardrail is fine. The input is not. Same bright line: guardrails do not sanitize.

Not quite. Medical or disability data plus a named provider is Tier 3 regardless of how well-written the guardrail is. Red.

Scenario 5

An enterprise AI workspace with SOC 2 is set up so every marketing prompt automatically includes the Fair Housing guardrail block from the Council-approved template.

Green. Enterprise workspace, Council-approved template, and guardrails baked into the workflow. This is what "guardrails plus governance" looks like at scale. Both forms working together.

Not quite. The workspace has SOC 2, the template is Council-approved, and there is no resident data anywhere. This is the fully-governed model. Green.

Scenario 6

You paste a bracketed delinquency scenario ([Resident] on [Apartment] with [X days] past due on $[Balance]) into a general AI tool with three guardrails on tone and language.

Green. Every sensitive slot is bracketed. Guardrails constrain the output tone. No real data ever left the property management system. This is the two-step: sanitize the input, guardrail the output.

Not quite. Every real detail is a placeholder. The guardrail shapes the tone. Nothing sensitive is in play. Green.

Scenario 7

A regional pastes four residents' full eviction filings with names, apartment numbers, and financial history into a general AI tool with the guardrail "Do not summarize identifying information."

Red. Four sets of real Tier 3 data in a general AI tool. The guardrail asking AI not to summarize identifiers does not un-paste them. The data is already in the tool. Guardrails do not sanitize.

Not quite. The moment those filings hit the tool, the data was exposed. What AI does with it afterward is downstream of the failure. Red.

Scenario 8

A property manager writes a guardrail block for vendor comparison prompts: no invented specs, no unauthorized promises about timing, say "I do not have that data" when the answer is not in the bid documents.

Green. Writing a reusable guardrail block for a common task is exactly what the workshop teaches. No resident data. Nothing sensitive. Just cleaner AI output on vendor work.

Not quite. There is no resident data anywhere. The guardrail is on tone and hallucination control for a business comparison task. Green.

Take Action · Anonymous Quiz

Five questions. No name, no email attached.

Tap A, B, C, or D for each question. Read the feedback that appears. Nothing is stored with your name. Nothing you type here identifies you. The whole class's results roll up so Tami can spot where the sticky points are for Week 18.

Question 1 of 5

Which of these describes the four rules for a good prompt guardrail?

Answer: B. Be specific. Pair positive with negative. Keep the list short. Layer over time.

Slide 5 walks through all four. Specific beats general ("do not include statistics you cannot cite" beats "be accurate"). Pairing gives AI a lane to move into ("do not X, say Y instead"). Short list keeps output from going to mush (three per prompt ceiling). Layering means you add new rules when real bad output shows the gap, not from theory. A and C are made up. D is wrong on almost every count.

Example of use

You are drafting a renewal email. Start with two specific guardrails (no invented stats, no hedge words). AI produces one email that also over-promises on timing. Next time, add a third: "Do not commit to any timeline you cannot confirm from the offer document. Say 'let me confirm' instead." That is Layer over time in action.

Question 2 of 5

Your guardrail is "Do not use hedge words like approximately or roughly." AI keeps using them. What usually fixes it?

Answer: C. Add the positive alternative: "State the exact number, or say 'let me confirm' instead."

This is Rule 2 from Slide 5, and it is the single most reliable fix when a guardrail is ignored. Negative-only guardrails leave AI without a fallback lane. AI still wants to be helpful, so it hedges. Adding "say Y instead" gives AI a lane. A is throwing out a working rule. B makes the mush problem worse. D does not change behavior. Nine times out of ten, the fix is the positive pairing.

Example of use

"Do not guess" alone gets ignored in the wild. "Do not guess. Say 'I do not have that information' instead." gets followed. Same rule, gate added, works.

Question 3 of 5

How many guardrails should you stack in a single prompt?

Answer: A. Three, at most.

Rule 3 from Slide 5. Past three, AI starts hedging everything to avoid tripping any of them. Output goes to mush. The three-guardrail ceiling forces you to pick the highest-leverage rules for this specific task. Save the rest for next time. B is invented; there is no "seven is optimal" number in the workshop. C is what people do the first month after learning guardrails and it makes their output worse. D is wrong (guardrails do not cancel; they combine).

Example of use

Renewal outreach with three guardrails (no invented stats, no hedge words, no unauthorized concessions) reads clean. Same email with nine guardrails reads like a legal disclosure. Pick the three that matter most for the task.

Question 4 of 5

What does the Week 17 bright line say about what guardrails do?

Answer: B. Guardrails constrain what AI writes. They do not sanitize what you paste in. Real resident data still stays out.

This is Slide 9 word for word. Every week has one line that must not blur, and this is Week 17's. Guardrails work on AI output. Sanitization works on AI input. Two different jobs, both required, neither replaces the other. Bracketed data, guardrailed prompts, human review. That is the full stack. A is exactly the confidence trap the bright line warns against. C is dangerous (guardrails reduce risk; they do not replace review). D is invented.

Example of use

A delinquency prompt with the guardrail "do not use language that threatens eviction" is fine. Same prompt with real names, apartment numbers, and balance amounts still exposes Tier 3 data. The guardrail shaped the output. It did not un-paste the input. Bracket, then guardrail. Not or.

Question 5 of 5

Which situation is the clearest signal to reach for guardrails?

Answer: C. Any output going to a resident, leadership, or straight into an automated system.

Slide 6 lists five times to reach for guardrails; this is the through-line. Anywhere AI output has a downstream reader who will judge you by it, guardrails. A is a Skip case (brainstorming needs the opposite of narrow). B is a Skip case (personal notes, no downstream reader). D is a Skip case (fast, verifiable, no risk).

Example of use

A resident newsletter draft that will be sent to 400 households, a vendor email that quotes numbers, an automated chat response embedded on the property site: all three are the reach-for-guardrails case. A one-off internal note to yourself is not.

Take Action · Downloads

Two files to take with you.

The Quick Reference is Slide 17 on one page. The Vocabulary Cheat Sheet is every AI term the curriculum has introduced through Week 17. Print them, save them to your phone, share them with the team.

Download 1

Guardrails Quick Reference

The template. The filled-in version. When to reach for it. When to skip it. The three traps to watch for. The bright line at the bottom. Slide 17, in your pocket.

💾 Download (PDF)

Download 2

AI Vocabulary Cheat Sheet

Every AI term the curriculum has introduced through Week 17. 46 terms in plain language. Foundational, prompt anatomy, prompt quality, iteration, RTF, Chain of Thought, Few-Shot, Guardrails, governance, and data. Same list as Slide 16.

💾 Download (PDF)

Send this to your supervisor

If your supervisor has not seen this curriculum yet, drop them a note. Every week people find out their peers were doing this while they were not. This week's note writes itself: guardrails are the guardrail that keeps AI output from becoming a compliance problem downstream. Enter your supervisor's email below (nothing is stored, nothing is sent to anyone else) and the button opens a pre-written email in your mail app.

Take Action · Session complete

Session complete.

What you walked away with

• The definition of a prompt guardrail: the "do not" line you add to a prompt.
• The four rules for an effective guardrail (specific, positive-paired, short, layered).
• Five times to reach for guardrails; four times to skip them.
• The four-structure stacked prompt: RTF + CoT + Few-Shot + Guardrails.
• The Week 17 bright line: guardrails constrain what AI writes. They do not sanitize what you paste in. Real resident data still stays out.
• The default rule for picking which structure to use when.

Good job. Four weeks ago you had RTF. Then Chain of Thought. Then Few-Shot. This week you added Guardrails. Four of the seven prompt structures for Phase 3 are in your hands. Three left. The curriculum board is right here, and this is the map for the next three weeks.

Phase 3 curriculum board

Week 14 · RTF (Role, Task, Format): the shape
Week 15 · Chain of Thought: the reasoning
Week 16 · Few-Shot: the pattern
Week 17 · Guardrails: the constraints (this week)
Week 18 · Persona: sustained character across a session (next week)
Week 19 · Structured Output: forcing the format
Week 20 · Iterative: refining across turns

Next Tuesday · Week 18

Persona: sustained character across a session.

Role in RTF is one line. Persona is the sustained character AI holds across every message in a session: a specific leasing pro voice, an experienced maintenance dispatcher, a compliance-first regional manager. Same idea, longer arc. When you want AI to sound like the same person across ten follow-ups, not ten different helpers, this is the structure.

Bookmark: multifamilynext.com/52-ai-tuesdays/week-18-of-52 (goes live next Tuesday).

Where the platform picks up

Everything you learned this week runs inside mPro Digital Edge. The platform's Chat Assistant lets you save guardrail blocks by task type, so you paste them once and reuse them every time. Fair Housing guardrails on marketing prompts. Statistic guardrails on renewal prompts. Timeline guardrails on vendor prompts. Compliance stops being a review step and starts being a default.

See you Tuesday.

Tami Siewruk
CEO, Multifamily NEXT and mPro Digital Edge

PE html> AI Tuesdays · Week 17 · RTF: Role, Task, Format

Warmup

The Workshop

Power Preview

Resources & Tools

Take Action

Week 17 of 52 · Self-Guided · ~25 min

0 starred 0 notes

Pre-Show · Warmup

Welcome back. Let us warm up first.

Three quick recall questions from last Tuesday's workshop on Few-Shot. Tap each to expand, tap again to close. Just for you, no scoring.

Example of use

Renewal outreach for [Property]. You paste three sanitized past emails that landed well. AI writes the fourth. Two examples would have worked. Six would have made AI mimic only the sixth.

Answer: Agree, Cover, Label, Sanitize. Every example points at the same pattern (agree). Show the variety AI will face (cover). Same format every time (label). No real resident data (sanitize).

Example of use

Answer: Examples with real resident details are still real resident data. Sanitize before you show. Your best past correspondence stays in your PMS. Brackets go in the Few-Shot examples.

Example of use

Wrong: pasting three past emails to Sarah Martinez in apartment 214 with balance $2,847. Right: replacing every real detail with [Resident], [Unit], [Balance]. Same pattern. No data leak.

Missed last week?

Open Week 16 · Few-Shot before you start this one. Password on that page is 1652. This week builds on it.

The Workshop · Opening

Tell AI what NOT to do.

This week in one line

Every prompt structure so far tells AI what to do. This one draws the line AI cannot cross.

Read this workshop as

Tap the role closest to yours. A few examples throughout will lean toward that context. Change your mind anytime.

The Workshop · Agenda

Six parts. About 25 minutes. No timer.

Self-guided. Pause whenever you need. Your notes and progress save to this device automatically.

What guardrails are

Slides 4 and 5. The definition and the anatomy of a good guardrail.

~4 min

When to reach for them and when to skip

Slides 6 and 7. High-stakes and outward-facing content: reach. Personal notes and brainstorming: skip.

~3 min

Four structures, one prompt

Slides 8 through 10. RTF plus CoT plus Few-Shot plus Guardrails, stacked. The bright line and the default rule.

~4 min

Power Preview: two examples

Slides 12 through 15. Renewal outreach without and with guardrails. Fair Housing guardrails on a property social caption.

~7 min

Reference and reinforcement

Slides 16 through 19. Vocab (46 terms), Guardrails reference card, standards, Spot the Tier.

~4 min

Quiz, downloads, and close

Slides 20 through 22. Five questions, two downloads, Week 18 teaser.

~3 min

No timer runs. The minute estimates are honest averages. Slower is fine.

The Workshop · Why You Should Care

Some AI output looks fine. Right up until it goes out.

Two things happen every week in this industry that guardrails prevent. Neither one looks like an emergency in the moment. Both are avoidable with one added line to your prompt.

Scenario 1

The stat AI made up

Scenario 2

The Fair Housing line AI crossed

Here is the thing

The Workshop · Teaching A

Guardrails in one line.

The definition

Guardrails are the "do not" lines you add to a prompt. Not what AI should do. What AI must not do.

Without guardrails

Prompt

You are a leasing pro. Draft a renewal outreach email for a resident on a 4 percent rent increase. Warm but direct.

What AI writes

"Your renewal is coming up. Even with the modest increase, you are still approximately 17 percent under current market. We would love to lock you in for another year at this competitive rate..."

Problem: AI invented "17 percent under market." It has no market data. AI hedged with "approximately." It committed you to a "competitive rate" you did not authorize.

With guardrails

Prompt (same task, four added lines)

What AI writes

Result: No invented number. No hedge. No commitment you cannot honor. Same warm tone. Ready to send.

Pay attention

The Workshop · Teaching B

Anatomy of a guardrail.

Not every "do not" line works equally well. Four rules separate a guardrail AI actually follows from one AI silently ignores.

Rule 1 · Be specific

Weak: "Be accurate."

Strong: "Do not include any statistic, percentage, or dollar figure you cannot cite from the source data I provided."

The specific version names the exact behavior AI will otherwise default to. "Be accurate" is a wish. The specific version is a rule.

Rule 2 · Pair positive with negative

Weak: "Do not guess."

Strong: "Say 'I do not have that information' instead of guessing."

Negative-only leaves AI without a fallback lane. Positive-plus-negative gives AI something to do instead. This is the single most reliable guardrail pattern.

Rule 3 · Keep the list short

Weak: Nine guardrails in one prompt.

Strong: Three at a time. Ceiling.

Past three, AI starts hedging everything to avoid tripping any of them. Output goes to mush. Pick the three highest-leverage rules for this task. Save the rest for next time.

Rule 4 · Layer them over time

Weak: Guess at every possible guardrail you might need.

Strong: Start with one or two. When AI produces bad output, add the rule that would have prevented it.

Your guardrail list should grow the way your team's brand voice guide grew: from real mistakes, not from theory. Save the working versions in a folder. Reuse them by task type.

Trust me

The Workshop · Teaching C

Five times to reach for guardrails.

Anywhere AI output has a downstream reader who will judge you by it, add guardrails.

Anything going to a resident

Renewal outreach, delinquency notes, move-in welcomes, community newsletters. If the recipient will save the email, forward the email, or quote it back to you in three months, guardrails.

Anything with numbers you cannot verify on the spot

Any Fair Housing-sensitive content

Vendor, legal, or leadership communications

Anything going into a system without human review

The Workshop · Teaching D

Four times to skip guardrails.

Guardrails narrow AI output. Sometimes narrow is the enemy of what you actually need.

Personal notes to yourself

You are asking AI to sketch a game plan for your Monday. Nobody but you will read it. Adding guardrails just slows you down. Skip.

Drafts you will review anyway

Brainstorming and idea generation

You want ten angles for a resident event, five subject line options, twelve possible names for an internal contest. Guardrails narrow output. Brainstorming needs the opposite. Skip.

Simple factual lookups

"What is the capital of Texas" does not need guardrails. If the task is short and the answer is verifiable in a second, skip. Guardrails are for the tasks where you cannot verify.

Here is the thing

The Workshop · Teaching E

Four structures, one prompt.

Task: renewal outreach for a resident on a 4 percent rent increase

Role: You are a leasing pro at [Property] with three years of tenure with this resident.

Task: Draft a renewal outreach email. Rent moves from $[Current] to $[New]. Offer is valid through [Date].

Format: Under 150 words. Warm sign-off.

Few-Shot: Match the pattern of these three renewal emails our team sent last year:

Example 1: [sanitized past email]

Example 2: [sanitized past email]

Example 3: [sanitized past email]

The Workshop · Teaching F

The bright line: guardrails do not sanitize.

Every week has one line that must not blur. This one is about what guardrails do not do.

Wrong

The confidence trap

Right

The two-step

Same task, brackets in place: "[Resident] in [Unit] is [X days] delinquent on $[Balance]. Draft an outreach email. Guardrail: do not use language that suggests we are threatening eviction. Say 'work with you on options' when discussing next steps."

The bright line

Guardrails constrain what AI writes. They do not sanitize what you paste in. Real resident data still stays out.

The Workshop · Default Rule

Shape, Reason, Pattern, Constraints.

Four Phase 3 structures. Four different jobs. Do not choose between them. Use the ones the task needs.

Shape

RTF

Role, Task, Format. The workhorse. Every prompt starts here.

Reason

Chain of Thought

Add when the task has real math, comparisons, or judgment. Ask AI to think step by step.

Pattern

Few-Shot

Add when you want AI to match your team's voice, format, or category rule. Paste two to four sanitized examples.

Constraints

Guardrails (this week)

Add when the output goes to residents, leadership, or straight into a system. Set the "do not" lines.

The Workshop · Q & A

Six questions people ask about guardrails.

Tap any question to expand the answer. Tap again to close.

Example of use

"Do not use hedge words like approximately, roughly, or about" works the same in every mainstream tool. "Be careful" does not. Specificity is the transferability.

Example of use

"Do not guess" alone gets ignored. AI still guesses in soft language. "Do not guess. Say 'I do not have that information' instead." gets followed. Same rule, gate added, works.

Example of use

A renewal email drafted with all four structures still passes across a leasing pro's screen before it sends. AI writes 90 percent of it right. The human catches the 10 percent no guardrail could have caught: the resident just lost a family member, and the warm tone needs to be gentler still.

Power Preview · Setup

Two runs. Same prompt each time. First without guardrails, then with.

You are about to see two side-by-side examples. Each shows AI output before and after adding a small set of guardrails. Watch what changes.

Run 1

Renewal Outreach With Guardrails

The invented statistic. The hedge word. The accidental price commitment. Watch three guardrails eliminate all three in one pass.

Run 2

Fair Housing Guardrails on a Property Social Caption

A family-friendly amenity caption AI wrote three ways. All three sound fine. Two are Fair Housing violations. Guardrails know the difference.

Read both, then Try This Today on slide 15 walks you through building guardrails for a real task in your job.

Power Preview · Run 1

Renewal outreach: without, then with.

Same resident. Same rent increase. Same request. Only the guardrails are new.

The zero-guardrail prompt (Run 1a)

"You are a leasing pro. Draft a renewal outreach email. Rent is going up 4 percent to $1,560 for [Resident] on a 12-month renewal. Warm but direct."

What AI writes back (excerpt)

Same prompt plus three guardrails (Run 1b)

What AI writes back (excerpt)

Power Preview · Run 2

Fair Housing guardrails on a social caption.

A marketing associate asks AI for three social captions for the property's new playground. Read all three. Two of them are Fair Housing violations. Guardrails know which.

The zero-guardrail prompt (Run 2a)

"Write three social media captions for [Property]'s new playground. Warm, community-focused, short."

What AI writes back

Caption 1: "A perfect place for young families to make memories. Our new playground is open!"

Caption 2: "New at [Property]: a playground the kids will love. Come see what makes this place feel like home for families like yours."

Caption 3: "Open now: a bright new play area next to the pool. Shaded seating for guardians. Rubberized surface. All ages welcome."

Same prompt plus Fair Housing guardrails (Run 2b)

What AI writes back

Caption 1: "Open now: our new playground next to the pool. Rubberized surface, shaded seating, and lots of room to move. All residents welcome."

Caption 2: "Fresh at [Property]: bright new play equipment, soft-fall surfacing, and benches under shade. A new spot on property, all residents welcome."

Caption 3: "Something new at [Property]: a play area with modern equipment, safety surfacing, and seating for anyone who wants to hang out. Open every day."

Trust me

Power Preview · Try This Today

Try This Today: 12 minutes to your first guardrail set.

Not next week. Today, before you close this tab. Twelve minutes and you will have a saved guardrail block you can reuse forever on one specific task.

Pick a task you sent to AI in the last week that ended up wrong (3 min)

Not the worst one. The most typical one. A renewal email, a delinquency note, a social caption, a vendor inquiry. Something you do more than once a month.

Read the output. Circle 2 or 3 things AI did wrong (3 min)

Write 2 guardrails in the "do not X. Say Y instead" pattern (3 min)

Run the same task with your new guardrails added (3 min)

Good job

Reference · Bookmark this slide

Vocabulary cheat sheet: 46 core terms.

Every AI term this curriculum has introduced so far, in one place. Plain-language definitions. This is the same term list on the PDF download at the bottom of this slide and on Slide 21.

Foundational (from Weeks 8-10)

Generative AI

AI that creates new content (text, images, audio, code) based on training data patterns.

Large Language Model (LLM)

The kind of AI behind ChatGPT, Claude, Gemini, and Copilot. Trained on vast text and predicts one word (token) at a time.

Token

A small chunk of text an LLM reads and writes. Roughly 0.75 words per token; 1,000 tokens is about 750 words.

Training Data

The vast collection of text an LLM learned patterns from. Not a searchable database, a compressed pattern archive.

Prediction

How AI decides what to say next: the most likely word based on the pattern of what came before, one word at a time.

Hallucination

AI producing a confident, fluent answer that is factually made up. Especially common when it does not know something you did not tell it.

Sycophancy

AI agreeing with what you just said instead of pushing back. Feels helpful, is often wrong.

Prompt anatomy (from Week 11)

Prompt

What you type to AI. Your input. The one part of the container you fully control.

System Prompt

A hidden instruction the AI tool sends the model before your prompt. Shapes tone, role, and constraints. You do not write it and cannot see it.

Conversation History

Every earlier message in the current chat, yours and AI's. AI re-reads all of it every turn. This is how AI "remembers."

Context Window

The ceiling on how much text AI can read in one conversation. When it fills up, the oldest content silently drops off.

Prompt Anatomy

The five parts of a well-formed prompt: instruction, context, input data, output format, examples.

Output Format

The shape you want the answer in: length, structure, tone. One line in your prompt saves you rewriting the whole reply.

Prompt quality (from Week 12)

Specificity

The single biggest lever on prompt quality. Naming the audience, the format, and the constraints so AI is not guessing.

Ambiguity

What makes AI fall back on average defaults. The parts of a prompt AI has to guess. Iteration cannot rescue what specificity should have prevented.

Iteration (from Week 13)

Iteration

Refining a prompt after reading the answer, in the same chat. Your first prompt is a draft. Your second, written on purpose, is almost always better.

Refinement

The act of sending a targeted follow-up: add what was missing, correct what was wrong, or show an example of right. One deliberate change beats a full rewrite.

The Three Moves

ADD (give it the context it did not have), CORRECT (tell it what to stop doing), SHOW (give it an example of right). Every refinement is one of these three.

Diminishing Returns

When each new round of refinement barely changes the answer. Your signal to stop prompting and finish the piece by hand.

Loop

The Ask → Read → Refine → Stop cycle. The four-step rhythm that turns first-answer disappointment into a usable draft.

RTF (from Week 14)

Prompt Structure

A repeatable pattern for writing prompts. RTF is a prompt structure. Chain of Thought is another. Seven of them are taught in Phase 3 (Weeks 14 through 20), each for a different kind of job.

RTF

Role, Task, Format. The workhorse prompt structure. Three parts in one line that make the first prompt land closer to what you actually needed.

Role (in RTF)

The first line of an RTF prompt. Tells AI who it is being right now. "You are a friendly on-site property manager." Changes what AI draws on and how it talks.

Task (in RTF)

The second line. What AI is actually doing. A verb, an object, and enough detail that two people could not interpret it differently. Not a topic.

Format (in RTF)

The third line. What shape the answer should take. Length, structure, tone. The part almost everyone skips, and the reason so many AI answers feel "close but not quite."

Chain of Thought (from Week 15)

Chain of Thought (CoT)

Reasoning

Step-by-Step

The most common CoT trigger phrase. "Think step by step" added to any RTF prompt turns on the reasoning walk-through. Simple, portable across every AI tool.

Show Your Work

The everyday-English version of CoT. What a teacher asks a student. When you ask AI the same thing, you get the same benefit: visible arithmetic, visible assumptions, catchable errors.

Few-Shot (from Week 16)

Few-Shot Prompting

Zero-Shot Prompting

Example (in prompting)

A concrete sample of the answer you want. In Few-Shot, the examples together form the pattern AI will match. Consistent, sanitized, and labeled the same way every time. Also called a "shot."

Pattern Matching

This week · Guardrails (Week 17) *NEW

Negative Instruction

Positive/Negative Pairing

Hedge Word

Governance and data (running through the curriculum)

Data Classification

Sorting a piece of information by which AI tool it is allowed to enter. Tier 1 (safe for any AI), Tier 2 (enterprise only), Tier 3 (never in AI). Full treatment: Week 1.

Tier 1 · Safe

Public or fully de-identified information. Any approved AI tool is fine.

Tier 2 · Enterprise Only

Sensitive but useful information. Only in an approved enterprise AI workspace with data protections in place.

Tier 3 · Never

Information that never enters an AI tool. Real names tied to sensitive details, financial identifiers, protected-class data, and screening or eviction information.

Use Case

A specific way you plan to use AI: who is doing it, what task, what data goes in, what comes out. The Council approves use cases, not tools.

Guardrail

Human-in-the-Loop

A person reviews AI output before it goes to a resident or a decision. The RETTC principle behind almost every safe use case.

AI Council

The group in your organization that approves use cases, reviews new tools, and owns the response when something goes wrong. Full treatment: Week 7.

NIST AI RMF

National Institute of Standards and Technology AI Risk Management Framework. The four functions: GOVERN, MAP, MEASURE, MANAGE.

RETTC AI Governance Framework

Real Estate Technology and Transformation Center. Eight principles specific to rental housing. Detailed on Slide 18.

💾 Download AI Vocabulary Cheat Sheet (PDF)

Reference · Bookmark this slide

Guardrails, on one page.

Read the left column when you are writing a prompt. Read the right when you want to see it in the wild.

The template

Write your RTF prompt first. Add CoT or Few-Shot if the task needs them.

Add up to three guardrails at the end. Use the "do not X. Say Y instead" pattern.

"Guardrails: Do not [specific behavior]. Say [alternative] instead. Do not [second specific]. Say [second alternative] instead."

The four rules for a good guardrail are on Slide 5: be specific, pair positive with negative, keep the list short, layer over time.

Filled in

Role: You are a leasing pro at [Property].

Task: Draft a renewal outreach email. Rent moves from $[Current] to $[New].

Format: Under 150 words. Warm sign-off.

Guardrails:

- Do not include statistics or percentages you cannot cite from data I gave you. Say "let me confirm" instead.

- Do not use hedge words (approximately, roughly, about). State what you know exactly.

- Do not offer or imply any concession, discount, or promotion. Say "let me confirm what the offer includes" instead.

When to reach for it

Skip Guardrails for personal notes, drafts you will review anyway, brainstorming (which needs the opposite), and simple factual lookups.

Watch for these traps

Stacking more than three. Nine guardrails make AI hedge everything. Output goes to mush. Ceiling is three per prompt. Save the rest for next time.

Negative-only guardrails. "Do not use bullet points" without a positive alternative gets ignored. Pair every "do not X" with "say Y instead" and AI follows almost every time.

The bright line, always

Guardrails constrain what AI writes. They do not sanitize what you paste in. Real resident data still stays out. See Slide 9.

💾 Download Guardrails Quick Reference (PDF)

Permanent · Standards Library

NIST AI Risk Management FrameworkNIST AI RMF (NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY, AI RISK MANAGEMENT FRAMEWORK)
The US federal framework for managing risk in AI systems. Four functions: GOVERN, MAP, MEASURE, MANAGE. Voluntary but widely adopted.

Multifamily example: A regional manager wants prompt guardrails standardized across every property's renewal-outreach workflow. The AI Council walks NIST's four functions: GOVERN (who approves the guardrail block for that task type), MAP (which failure modes matter most: invented stats, hedge words, unauthorized concessions), MEASURE (spot-check outbound emails against the guardrails for a month), MANAGE (revise the block when a new failure mode shows up). Same framework as Weeks 1 through 16. Different anchor. Guardrails now sit inside a governed workflow.

Full document at nist.gov/itl/ai-risk-management-framework & RETTC 8 PrinciplesRETTC AI GOVERNANCE FRAMEWORK (REAL ESTATE TECHNOLOGY AND TRANSFORMATION CENTER)
Eight principles for AI governance in rental housing. Anchored to Fair Housing, resident privacy, and human accountability.

Multifamily example: A property adds a Fair Housing guardrail block to every marketing prompt. RETTC's Promote Fairness & Consumer Protection principle says the operator is responsible for outputs that comply with the Fair Housing Act. The guardrail block is one specific way to operationalize that responsibility inside the workflow, not just inside the policy binder. Belt and suspenders: guardrail in the prompt, human review after, both required.

Published at rettc.org/frameworks/ai-governance.

The "do not" lines you add to a prompt

GOVERN function: policy defines which negative instructions belong in prompts for which task types. A sanctioned guardrail block is a governed asset, not a personal preference.

Establish an Organizational Philosophy on AI Use and Oversight: guardrails that reflect team standards are a lever for the philosophy, not a workaround around it.

Preventing invented statistics and hedge words

MEASURE function: track whether AI output stayed inside the fence you set. If a guardrail is being ignored, you cannot fix it without measuring how often it is being ignored.

Fair Housing guardrails on marketing content

MAP function: identifying the failure modes that matter for this task. Familial-status language and other protected-class steering are the specific risks marketing prompts create.

Saved guardrail blocks reused by task type

GOVERN function: which guardrail block becomes the standard for which task type is a governance decision, not a personal preference. Whose blocks get elevated matters.

Protect Privacy & Promote Data Integrity: standardized guardrail workflows reduce the surface area where sensitive data could leak, and keep AI output aligned with data-minimization principles.

The bright line (guardrails do not sanitize)

MANAGE function: the incident response plan covers what happens when someone assumes a guardrail also sanitized the input, and pastes real resident data. Rehearse it before it happens.

Maintain Outcome-Based Accountability & Human Involvement: the person who chose to paste the input owns the outcome. Guardrails plus review, not either alone.

Establish an Organizational Philosophy on AI Use and Oversight. Adopt a clear philosophy that defines how AI will be used, governed, and monitored across operations.

Promote Fairness & Consumer Protection. Ensure AI systems comply with the Fair Housing Act, Fair Credit Reporting Act, Americans with Disabilities Act, and state and local law.

Protect Privacy & Promote Data Integrity. Adhere to federal, state, and local consumer privacy laws, with an emphasis on data minimization, security, and informed consent.

Promote Positive Renter Experiences and Outcomes. Elevate technologies that enhance affordability, access, and housing stability, and use AI to improve resident experiences.

Strive for Innovation with Responsibility. Encourage development of AI tools that benefit residents, prospects, guests, and operators, without compromising trust.

Read the source

Reinforcement · Spot the Tier

Spot the Tier: 8 Guardrails scenarios.

For each scenario, tap the tier you think it belongs to. Feedback shows immediately with the correct tier and the reason. Reset the whole set with the button at the bottom.

Scenario 1

You paste sanitized team emails into an AI tool with three guardrails: no invented stats, no hedge words, no unauthorized concessions. AI drafts a renewal email that reads clean.

Green. Sanitized examples plus a bounded guardrail block. This is exactly the RTF+Few-Shot+Guardrails stack the workshop teaches. Safe use in any approved tool.

Not quite. Everything sensitive is stripped from the examples, and guardrails constrain the output. Nothing here needs an enterprise workspace. Green.

Scenario 2

You paste unsanitized past emails to Sarah Martinez in apartment 214 with balance $2,847 into ChatGPT, and add a guardrail: "Do not use language that threatens eviction."

Red. This is the bright line failure. The guardrail shaped what AI would write. It did not sanitize what was pasted. Real name, real unit, real balance, all now in a general AI tool. Guardrails do not sanitize.

Not quite. The guardrail is fine. The input is the problem. Real resident name plus apartment number plus a delinquency balance in a general AI tool is Tier 3. Red.

Scenario 3

A marketing associate saves a "Fair Housing caption guardrails" block: no age/family-status language, no religious references, describe amenities not audience. They paste it into every social prompt.

Green. This is the practice the workshop pushes: save the block once, reuse it forever, and let the guardrail run in the background of every marketing prompt. No sensitive data in play.

Not quite. There is no resident data anywhere in this workflow. The guardrail block itself is public content. Green.

Scenario 4

You paste an ADA accommodation request that includes the resident's condition (mobility impairment) and provider's name into a general AI tool, with guardrails to keep the tone respectful.

Red. Disability status and provider names are Tier 3 the moment they enter a general AI tool. The guardrail is fine. The input is not. Same bright line: guardrails do not sanitize.

Not quite. Medical or disability data plus a named provider is Tier 3 regardless of how well-written the guardrail is. Red.

Scenario 5

An enterprise AI workspace with SOC 2 is set up so every marketing prompt automatically includes the Fair Housing guardrail block from the Council-approved template.

Green. Enterprise workspace, Council-approved template, and guardrails baked into the workflow. This is what "guardrails plus governance" looks like at scale. Both forms working together.

Not quite. The workspace has SOC 2, the template is Council-approved, and there is no resident data anywhere. This is the fully-governed model. Green.

Scenario 6

You paste a bracketed delinquency scenario ([Resident] on [Unit] with [X days] past due on $[Balance]) into a general AI tool with three guardrails on tone and language.

Green. Every sensitive slot is bracketed. Guardrails constrain the output tone. No real data ever left the property management system. This is the two-step: sanitize the input, guardrail the output.

Not quite. Every real detail is a placeholder. The guardrail shapes the tone. Nothing sensitive is in play. Green.

Scenario 7

A regional pastes four residents' full eviction filings with names, apartment numbers, and financial history into a general AI tool with the guardrail "Do not summarize identifying information."

Red. Four sets of real Tier 3 data in a general AI tool. The guardrail asking AI not to summarize identifiers does not un-paste them. The data is already in the tool. Guardrails do not sanitize.

Not quite. The moment those filings hit the tool, the data was exposed. What AI does with it afterward is downstream of the failure. Red.

Scenario 8

Green. Writing a reusable guardrail block for a common task is exactly what the workshop teaches. No resident data. Nothing sensitive. Just cleaner AI output on vendor work.

Not quite. There is no resident data anywhere. The guardrail is on tone and hallucination control for a business comparison task. Green.

Take Action · Anonymous Quiz

Five questions. No name, no email attached.

Question 1 of 5

Which of these describes the four rules for a good prompt guardrail?

Answer: B. Be specific. Pair positive with negative. Keep the list short. Layer over time.

Example of use

Question 2 of 5

Your guardrail is "Do not use hedge words like approximately or roughly." AI keeps using them. What usually fixes it?

Answer: C. Add the positive alternative: "State the exact number, or say 'let me confirm' instead."

Example of use

"Do not guess" alone gets ignored in the wild. "Do not guess. Say 'I do not have that information' instead." gets followed. Same rule, gate added, works.

Question 3 of 5

How many guardrails should you stack in a single prompt?

Answer: A. Three, at most.

Example of use

Question 4 of 5

What does the Week 17 bright line say about what guardrails do?

Answer: B. Guardrails constrain what AI writes. They do not sanitize what you paste in. Real resident data still stays out.

Example of use

Question 5 of 5

Which situation is the clearest signal to reach for guardrails?

Answer: C. Any output going to a resident, leadership, or straight into an automated system.

Example of use

Take Action · Downloads

Two files to take with you.

The Quick Reference is Slide 17 on one page. The Vocabulary Cheat Sheet is every AI term the curriculum has introduced through Week 17. Print them, save them to your phone, share them with the team.

Download 1

Guardrails Quick Reference

The template. The filled-in version. When to reach for it. When to skip it. The three traps to watch for. The bright line at the bottom. Slide 17, in your pocket.

💾 Download (PDF)

Download 2

AI Vocabulary Cheat Sheet

💾 Download (PDF)

Send this to your supervisor

Take Action · Session complete

Session complete.

What you walked away with

Phase 3 curriculum board

Next Tuesday · Week 18

Persona: sustained character across a session.

Bookmark: multifamilynext.com/52-ai-tuesdays/week-18-of-52 (goes live next Tuesday).

Where the platform picks up

See you Tuesday.

Tami Siewruk
CEO, Multifamily NEXT and mPro Digital Edge

Week-17-of-52: Guardrails

Welcome back. Let us warm up first.

Tell AI what NOT to do.

Six parts. About 25 minutes. No timer.

Some AI output looks fine. Right up until it goes out.

Guardrails in one line.

Anatomy of a guardrail.

Five times to reach for guardrails.

Four times to skip guardrails.

Four structures, one prompt.

The bright line: guardrails do not sanitize.

Shape, Reason, Pattern, Constraints.

Six questions people ask about guardrails.

Two runs. Same prompt each time. First without guardrails, then with.

Renewal outreach: without, then with.

Fair Housing guardrails on a social caption.

Try This Today: 12 minutes to your first guardrail set.

Vocabulary cheat sheet: 46 core terms.

Guardrails, on one page.

Spot the Tier: 8 Guardrails scenarios.

Five questions. No name, no email attached.

Two files to take with you.

Session complete.

Welcome back. Let us warm up first.

Tell AI what NOT to do.

Six parts. About 25 minutes. No timer.

Some AI output looks fine. Right up until it goes out.

Guardrails in one line.

Anatomy of a guardrail.

Five times to reach for guardrails.

Four times to skip guardrails.

Four structures, one prompt.

The bright line: guardrails do not sanitize.

Shape, Reason, Pattern, Constraints.

Six questions people ask about guardrails.

Two runs. Same prompt each time. First without guardrails, then with.

Renewal outreach: without, then with.

Fair Housing guardrails on a social caption.

Try This Today: 12 minutes to your first guardrail set.

Vocabulary cheat sheet: 46 core terms.

Guardrails, on one page.

Spot the Tier: 8 Guardrails scenarios.

Five questions. No name, no email attached.

Two files to take with you.

Session complete.