Last week you learned the mechanism: AI does not know, AI predicts. It hallucinates because it does not naturally admit uncertainty. It flatters because its training taught it that agreeing felt helpful. This week we go one level down. If AI predicts based on what you give it, then everything hinges on what you give it. That is the prompt.
New here?
Coming to AI Tuesdays for the first time? Read Week 10 first. This workshop assumes you already know why AI hallucinates and flatters. Week 10 teaches that in 25 minutes. Come back here after.
Question 1 of 3
Week 10 recap: The core mechanism
Fill in the blank. "AI does not know. AI ____."
Predicts.
One word at a time. Each one picked based on the pattern of your prompt. Which means your prompt is doing more work than most people realize. That is what Week 11 is about.
Question 2 of 3
Week 10 recap: Why AI hallucinates
When AI produces a confident but fabricated fact (like a rent statistic it could not know), what is the mechanism?
Prediction does not naturally admit uncertainty.
Without an instruction that tells it to hedge, AI produces a plausible-sounding answer. That is why the verification prompt from Week 10 works: it installs the exit ramp AI does not build in by default.
Question 3 of 3
Bridge into Week 11
If AI predicts the next word based on the pattern of your prompt, what does that make the prompt?
The only material AI has to work with.
Plus a hidden system prompt (the tool's built-in instructions) and any earlier conversation history. That is it. If it is not in one of those three places, AI does not have it. That is what Week 11 unpacks, and it is why weak prompts produce weak output.
Multifamily NEXT · AI Tuesdays · Week 11
The Anatomy of a Prompt
AI does not read minds. AI reads prompts.
Week 10 opened the hood on the mechanism: AI predicts the next word. Week 11 opens the hood on the input that mechanism runs on. What you type. And a couple of things you did not type but are in there anyway: the tool's hidden system promptSYSTEM PROMPTA hidden instruction the AI tool sends to the model before your prompt. It usually tells the model what role to play, what to avoid, what tone to use, and what format to respond in. You do not write it and you cannot see it directly, but it shapes every response. Multifamily context: a leasing chatbot's system prompt might say "you are a friendly leasing assistant. Never quote specific rent numbers. Always suggest scheduling a tour." That is why every response steers toward a tour. Anchor: RETTC Principle 3 (Uphold Transparency & Explainability). and the ongoing conversation historyCONVERSATION HISTORYEvery message earlier in the current chat, from both you and AI. AI reads the whole history every time it responds. This is how AI "remembers" things you said earlier. It does not have real memory: it has a re-reading habit. When the history gets too long for the context window, the oldest messages drop off. Multifamily context: if you told AI on message 3 that your property allows pets, and then on message 40 you ask about a pet application, AI may have already lost that detail. Anchor: RETTC Principle 3.. Together, that is the entire universe AI is working from.
When you close this workshop, you will build every prompt like a container you are packing on purpose, not a wish you are throwing at the machine.
Last Tuesday · Week 10
How a Large Language Model Thinks
AI does not know, it predicts. That mechanism explained hallucination and sycophancy. This week: what the mechanism actually gets to see.
Localize · Which of these is you?
Which role are you reading this in? Your answer sets up the anchor scenarios you will see next.
Beginner-friendly · Read at your own pace · Star slides, add notes, print, or email yourself anytime
The Workshop · Agenda
Six parts. About 25 minutes. No timer.
Pause whenever you need. Star anything you want to come back to. Add your own notes as you go. This is designed to teach you how to build a prompt on purpose, not to race you through it.
Part 1 · ~3 min
What AI actually receives when you hit send
Your prompt + the tool's hidden system prompt + the conversation so far. Everything else is invisible to AI.
Part 2 · ~4 min
The five parts of a well-formed prompt
Instruction, context, input data, output format, examples. Not all five every time, but knowing them changes how you write.
Part 3 · ~3 min
The context window (with real model specs)
The ceiling on how much text AI can see at once. What gets dropped when it fills up. Current sizes for ChatGPT, Claude, Gemini, and Copilot.
Part 4 · ~3 min
Why prompts fail
Buried instruction, missing context, no output format. Three failures every beginner makes and how to spot them in your own prompts.
Part 5 · ~7 min
Power Preview: side-by-side prompts
Same task. Weak prompt. Well-formed prompt. Read what AI produces from each. The gap is the point.
Part 6 · ~5 min
Try This Today · Reference · Session Complete
One prompt-anatomy exercise you can do this afternoon. Updated vocabulary cheat sheet. Board paragraph. Handoff to the workspace where you can practice this safely.
Time estimate is honest: reading time plus interaction time plus a little reflection. Move at your pace.
Why You Should Care
Two scenarios you might recognize.
Two composite situations built from the pattern we keep seeing. Both are the kind of thing a smart professional does when they think they are being efficient. In both, the AI tool got blamed for a problem that was actually in the prompt.
Scenario one
The generic renewal letter
A leasing professional needs a renewal letter for a resident who has lived in her property for four years and is renewing at $47 over current rent. She has 15 minutes before her next tour. She opens ChatGPT and types: "write a renewal letter."
ChatGPT produces a perfectly grammatical, entirely generic renewal letter. Nothing about her resident. Nothing about the four years. Nothing about the $47 increase. Nothing about the property. She copies it, personalizes the greeting, sends it. The resident replies asking why the letter feels like it could be from any property in the country.
She blames the AI. But AI did exactly what her prompt asked for: a renewal letter. Generic in, generic out. Everything specific to her resident was not in the prompt, so AI did not have it.
Scenario two
The forgotten budget number
A property manager is planning a resident-appreciation event with AI. On message 3, she says her budget is $2,400. Over the next 40 messages, they hash out ideas: food truck options, entertainment, timing, signup workflow. On message 42, she asks: "now go back and use the budget number I gave you to spec the final food-truck package."
AI produces a package. The number does not match what she said on message 3. She asks again. AI produces a different number. She gets frustrated: "why can you not remember what I said?"
AI does not have memory. It has a re-reading habit. Every message re-reads the conversation history. When the history got long enough to hit the context window ceiling, the earliest messages, including message 3 with her budget, dropped off. AI is not forgetting on purpose. AI just does not see that far back anymore. (Slide 6 explains the mechanism.)
Both are prompt-anatomy problems. Both are avoidable. Neither is AI's fault, but both look like AI's fault to the person who does not understand what AI actually saw.
Part 1 · Teaching A
What AI actually receives when you hit send.
Most people picture their prompt going alone to AI. It does not. AI receives three things every time, and only three things.
The three pieces AI sees.
Piece 1 · Visible to you
Your promptPROMPTWhat you type to AI. Your input. The one part of the container you fully control. Every word matters because AI predicts based on the pattern of your prompt. Full treatment: this workshop. Anchor: RETTC Principle 3 (Uphold Transparency & Explainability).
What you typed. The only piece you fully control. Everything you did not say is invisible to AI.
Piece 2 · Hidden
The system prompt
The tool's built-in instructions. You do not write it, cannot see it, but it shapes every response. That is why the "same model" behaves differently in different tools. Practical example: your property's leasing chatbot always steers conversations toward "let me schedule a tour," even when the resident is asking about maintenance. That is the system prompt at work.
Piece 3 · Grows as you chat
The conversation history
Every earlier message in this chat, yours and AI's. AI re-reads all of it every time it responds. When it gets too long, the oldest messages drop off.
Everything else is invisible.
AI cannot see
Your calendar. Your PMS. Your resident's file. The document sitting open on your other monitor. What you said out loud before typing. What you meant but did not write. Your property's name unless you type it. Your policies unless you paste them. Your resident's history unless you include it.
If it is not in your prompt, in the system prompt, or in the conversation history, it does not exist as far as AI is concerned.
Advanced note · What about tools with connectors?
If you do not use connected AI tools, you can skip this note. Some AI products can now connect to your Gmail, calendar, Google Drive, or other systems. In those cases, information from those connected sources gets pulled INTO the prompt (or into the context AI sees) before AI responds. It is still true that AI only sees what is IN the container. The container just got bigger because your tool loaded it for you.
Gut check · 15 seconds
You have been chatting with AI about a renewal draft for 20 messages. You now open a new browser tab, start a fresh chat, and ask AI "continue the renewal draft." What happens?
AI has no idea what "the renewal draft" refers to.
A new chat starts a new container. Nothing from the old chat carries over. If you want AI to work with material from a previous session, you have to paste that material into the new chat. Some tools have "memory" features that carry limited context across chats, but the default is: new chat = blank container.
Part 2 · Teaching B
The five parts of a well-formed prompt anatomyPROMPT ANATOMYThe named parts of a well-formed prompt: instruction, context, input data, output format, and (optional) examples. Not every prompt needs every part. Knowing the parts changes how you write because you can spot immediately which one you left out. Different guides use different labels and slightly different splits; this workshop teaches five. Anchor: RETTC Principle 3 (Uphold Transparency & Explainability)..
Not every prompt needs all five parts. But knowing the five changes how you write, because you can spot immediately which one you left out.
Quick disambiguation
Context is descriptive: who is this for, what is the situation, what tone, what background matters. Input data is content: documents you paste, actual records, text AI should work with. When in doubt: if it describes the situation, it is context; if AI should read it and use it, it is input data.
Part 1
Instruction · What do you want AI to do?
The verb. Write, draft, summarize, list, rewrite, compare, explain. This should be the first thing AI reads, or close to it. Buried instructions are the number-one reason prompts fail. Multifamily example: "Draft a renewal letter" is the instruction. "I need something for a renewal" is not.
Part 2
Context · What does AI need to know to do this well?
Who is the audience? What is the situation? What tone? What background matters? Multifamily example: "For a first-year resident renewing at 3% over current rent, in a mid-tier suburban property, in a market that is softening." That is context. AI cannot guess any of that.
Part 3
Input data · What material should AI work with?
The actual content AI should reference: a document you paste, a table, a resident's stated concern, a policy excerpt, a previous email thread. Multifamily example: pasting the resident's ledger and their last two work-order histories, then asking AI to draft a retention outreach based on both.
Before you paste · Data classification check
Week 1 taught the three data tiers. A well-formed prompt does not override those tiers. Before you paste resident financial data, health information, application details, or anything else Tier 2 or Tier 3, check which AI tool you are using and whether that tool is approved for that tier. Consumer ChatGPT, Claude, and Gemini free tiers are Tier 1 only. Pasting Tier 2 or Tier 3 data into them is a violation, even if it makes your prompt better.
Part 4
Output formatOUTPUT FORMATThe shape you want AI's response in: length, structure, tone, what to include or omit. Named at the end of your prompt. If you do not specify, AI picks a default (usually longer than you want, with headers or bullets you may not need). Multifamily example: "Respond as two short paragraphs, 60-80 words each, no header, warm but professional." Anchor: RETTC Principle 3 (Uphold Transparency & Explainability). · What should the response look like?
A three-paragraph email. A five-bullet list. A table with these columns. A short professional response. If you do not specify, AI picks. You may not like what AI picks. Multifamily example: "Respond as two short paragraphs (60-80 words each), no bullet points, no header, warm but professional."
Part 5
Examples · What does "good" look like?
Sometimes it is much faster to show AI one or two examples of the kind of output you want than to describe it. Multifamily example: paste one renewal letter your team already loves, then say "write another one in this exact voice for [new resident details]." This is called few-shot prompting and it is taught in depth in the Phase 3 prompt structures weeks (Weeks 14-20).
A note on the "five parts"
These are ingredients, not a required order or complete set. Some prompts use only one or two. Some prompts include additional structural elements (personas, chains of reasoning, guardrails). Phase 3 (Weeks 14-20) teaches seven specific prompt structures that combine these ingredients differently for different tasks. What Week 11 does is give you the vocabulary.
A well-formed prompt using all five.
[Part 1: Instruction] Draft a renewal letter.
[Part 2: Context] The resident is Maria Alvarez, a first-year resident in apartment 214 at Riverside on Broad. Rent is currently $1,485, renewing at $1,530 (3% increase). She has never submitted a work order. She has been quiet and reliable.
[Part 3: Input data] Attached is our standard renewal script and this month's new-resident survey shows Maria rated her move-in experience 5 of 5.
[Part 4: Output format] Two short paragraphs, 60-80 words each. No bullet points. Warm but professional. Do not mention the specific dollar amount of the increase.
[Part 5: Example] Here is one renewal letter our team liked last quarter: [paste example].
Gut check · 15 seconds
Which part of a well-formed prompt is most commonly missing in beginner prompts?
Context.
Instruction is usually there (people say "write a letter"). Output format is often skipped but you can iterate. Context is what beginners consistently under-supply, and it is the part that makes the difference between generic and useful.
Part 3 · Teaching C
The context window.
The context windowCONTEXT WINDOWThe maximum amount of text AI can process in a single request. It includes everything: your prompt, the hidden system prompt, the conversation history, any documents you pasted, and the space AI needs for its response. Measured in tokens (~0.75 words per token). When the total exceeds the window, the oldest content drops off. Multifamily context: this is what makes AI "forget" earlier parts of long chats. Anchor: RETTC Principle 3 (Uphold Transparency & Explainability). is the ceiling on how much text AI can see in one request. All three pieces from Slide 4, plus room for the response, all counted together. Measured in tokensTOKENThe small chunks of text an LLM reads and writes. Roughly 0.75 words per token in English, so 1,000 tokens is about 750 words and 1 million tokens is about 750,000 words. Multifamily context: a five-page vendor contract might be around 3,000-4,000 tokens. A 200-page compliance manual might be around 100,000 tokens. Full treatment: Week 08. Anchor: RETTC Principle 3. (about 0.75 words per token).
What actually gets exposed to you (as of July 2026).
This is where it gets confusing. Behind every consumer AI product is an underlying model with a large context window (often 1 million tokens or more). But the consumer PRODUCT (the app you use) exposes only a slice of that. How much depends on your subscription tier. These specs change every few months. Verify current numbers before you commit to a workflow.
Product
Plan
Approximate context window
ChatGPT (OpenAI)
Free
Approximately 8,000-16,000 tokens (about 6,000-12,000 words)
ChatGPT
Plus ($20/mo)
Approximately 32,000 tokens (about 24,000 words)
ChatGPT
Pro ($200/mo)
Approximately 128,000 tokens (about 96,000 words)
Claude (Anthropic)
Free
Limited (not publicly documented)
Claude
Pro ($20/mo)
Approximately 200,000 tokens (about 150,000 words)
Claude
Team / Enterprise
Approximately 500,000 tokens (about 375,000 words)
Gemini (Google)
Free (in Google apps)
Up to 1,000,000 tokens (about 750,000 words) · the largest free-tier window in this list
Gemini
Google One AI Premium ($20/mo)
Up to 1,000,000+ tokens (about 750,000+ words)
Microsoft Copilot
Microsoft 365 subscription
Approximately 128,000 tokens (about 96,000 words) when using the default GPT-based model. In Word, Excel, PowerPoint, and Outlook, Copilot uses retrieval instead of loading full documents into the window, which changes the effective context calculation.
Numbers verified July 2026. Consumer AI tiers change every few months. Always check your tool's current documentation before you build a workflow around a specific number.
Most affordable large context
If context window matters (long documents, long conversations, lots of pasted material) and budget matters, Gemini's free tier gives you the largest window at zero cost. That is the current state and it may change. Verify before you build a workflow around it.
Important nuance · Advertised is not effective
A 1,000,000-token window does not mean AI reliably uses every word in a 750,000-word document. Independent research (NVIDIA's RULER benchmark and Stanford's "Lost in the Middle" study by Liu et al., 2023) has documented that models often miss information buried in the middle of very long prompts. Practical takeaway: if a specific detail matters, put it near the beginning or end of your prompt, not buried in the middle.
Gut check · 15 seconds
Which of these is true about the context window?
Consumer product exposes less. Model uses only part of what is exposed.
Two separate reasons the number you see is less than the number the model can theoretically handle. Both matter. Put important information near the beginning or end of your prompt, not buried in the middle where models often lose it.
Part 4 · Teaching D
Why prompts fail.
Three failure patterns account for the vast majority of "AI gave me a bad answer" complaints. All three are prompt problems, not AI problems.
Failure 1
Buried instruction
What it looks like: The instruction is somewhere in your prompt, but the reader has to hunt for it.
Bad prompt: "So we have this resident, Maria, and she has been here four years now, actually five in November, and her rent is going up 3%, well, $47 to be exact, and I think we should probably write her a letter about the renewal so she knows we appreciate her and can you help me with that?"
Why it fails: AI predicts based on the pattern of what came before. If the first 60 words of your prompt are context wrapped around a hedged instruction ("probably write her a letter"), AI reads that as your dominant signal. Front-load the instruction. Then supply context.
Failure 2
Missing context
What it looks like: Clear instruction, but AI has no way to make the response specific to your situation.
Bad prompt: "Write a renewal letter."
Why it fails: AI does not know who the resident is, how long they have been there, what the increase is, what tone your property uses, or anything else. It fills in with generic averages, which is why the letter feels like it could come from any property in the country. Context is what beginners consistently under-supply. It is the single biggest lever on prompt quality.
Failure 3
No output format specified
What it looks like: Good instruction, good context, but AI picks the wrong shape for the response.
Bad prompt: "Draft a renewal letter for Maria [with good context]."
Why it fails: AI defaults to what its training saw most often for "draft a letter," which is often longer than what you want, with headers you did not need, or bullet points where you wanted paragraphs. Tell AI the shape. Two paragraphs, 60-80 words each, no header, warm tone.
A fourth failure that combines the first three
The lazy one-liner. "Help me with the renewal." No verb, no context, no output format. This is the most common prompt in the wild. It fails all three ways at once.
Gut check · 15 seconds
Of the three failures, which is the single biggest lever on prompt quality?
Context.
All three matter. But if you fix only one, fix context. AI can guess at format and figure out from a buried instruction. AI cannot guess at your specific resident, your specific market, your specific tone. Context is what beginners under-supply consistently and it is what turns generic into useful.
Part 4 · Teaching E
Three fixes that work starting today.
One fix per failure. Do these three things and your prompts jump a level, even before you learn the specific frameworks in Weeks 14-20.
Fix 1
Front-load the instruction
Rule: Verb + object in the first eight words. Everything else after.
Before: "I have this resident renewing and I was thinking, could you help me draft a letter for her?"
After: "Draft a renewal letter. Here are the details: [resident name, tenure, increase, tone]."
Fix 2
Give the context AI cannot guess
Rule: Answer four questions in one or two sentences: Who is the audience? What is the situation? What tone? What background matters?
Before: "Draft a renewal letter for Maria."
After: "Draft a renewal letter for Maria Alvarez, a first-year resident at Riverside on Broad, renewing at 3% over current, tone warm-professional. She has been quiet and reliable. Her move-in survey scored us 5 of 5."
Fix 3
Name your output format
Rule: One line at the end telling AI the shape you want.
Before: [instruction and context, then nothing more]
After: Add "Respond as two short paragraphs, 60-80 words each. No bullet points. No headers. Do not include the specific dollar amount of the increase."
Combined
Front-load the instruction. Give the context. Name the format. Three sentences, three moves, and your prompt is doing 80% of what a professionally engineered prompt does. The remaining 20% is the frameworks and iteration patterns coming in Weeks 14-20.
What about at volume?
A leasing pro writing 15 renewal letters a week is not going to type 200 words of context every time. Neither is anyone else. Three ways to scale a well-formed prompt without retyping:
Path 1
Saved custom instructions
Most consumer AI tools let you save standing context (your property, your voice, your role) at the account level. Set it once. Every chat starts with it loaded. ChatGPT calls this Custom Instructions. Claude calls this a Project system prompt. Gemini calls this a Gem.
Path 2
Saved prompt templates
Copy your well-formed prompt into a doc. Fill in the resident-specific details each time. You are not retyping context, you are swapping placeholders. Ten seconds per prompt instead of two minutes.
Path 3
Enterprise workspace
For high-volume work with real resident data, an enterprise AI workspace can pre-load context from your PMS or CRM automatically. Data classification and human review are built in. That is what mPro Digital Edge does.
The point
Context is a re-usable investment, not per-prompt overhead. You write it once, and it saves you time on every prompt after.
Part 4 · Teaching F
The bright line for every prompt.
Everything in this workshop points to one rule. Say it out loud. Write it on a sticky note. Attach it to your monitor.
The bright line
If it is not in the prompt, AI does not have it.
Not in the prompt, not in the system prompt, not in the conversation history. That is the entire universe AI is working from. Your property name. Your resident's history. Your budget number from message 3. Your tone preferences. Your policy about pets. Your last quarter's data. Your rent card. Your ledger. If it is not in one of those three places, AI does not have it.
This is not a warning about AI being unreliable. AI is reliable at working with what you give it. This is a warning about assuming AI knows what you did not say. Two different things.
The operational discipline
Before you send, ask yourself: what can AI not see about this task? Then put it in the prompt.
Next up: the default rule that turns this bright line into a habit. And on Slide 12, the well-formed prompt template you can start using this afternoon.
The Workshop · Default Rule
The rule of Week 11.
If you take one thing from this workshop, take this.
The default rule
Every prompt is a container. Choose what goes in it deliberately.
Not a wish list. Not a hint. Not "you know what I mean." A container that you are packing on purpose with the instruction, the context, the input data, the output format, and, when it helps, an example. Everything else stays outside the container, which means it stays outside what AI can see.
Your synthesis.
Write down three things you will do differently in your next AI conversation, based on what you just learned. Not aspirational. Actual. Specific to your role.
Commitment card · Three things
The Workshop · Q&A
Questions readers are already asking.
These are the kinds of questions multifamily professionals ask while working through this material. Tap to expand.
No. You have three shortcuts. First, most AI tools let you save custom instructions or a system prompt on your account. You put the standing context there once (your property, your voice, your policies) and it applies to every chat. Second, you can save a well-formed prompt template and swap in the resident-specific details each time. Third, for high-volume work, an enterprise workspace (like mPro Digital Edge) can pre-load the context automatically. What you cannot skip is the fact that AI still needs the context. You are just moving where it comes from.
Not by time. By length. The context window resets when you start a NEW chat, not when time passes. A chat that goes 40 messages will bump into the window ceiling whether you did it in an hour or across two days. Some tools now have "memory" features that carry limited context across chats: ChatGPT's Memory feature (which stores facts you tell it to remember), Claude's Projects (which share files and instructions across chats within one project), and Gemini's Gems (custom instances with saved context). These carry preferences and named facts, not full histories. Read your specific tool's documentation before you rely on memory to remember something important.
Usually no. Two reasons. First, you will burn a lot of the context window on the parts of the policy AI does not need for this specific question. Second, remember advertised context is not effective context: buried information often gets missed. Better approach: paste only the relevant section (or ask AI to summarize just the retention policy first, then use that summary in later prompts). For repeated use, an enterprise workspace that indexes your policy manual is a better answer than pasting every time.
No. Front-loading the instruction is the strongest single rule. After that, order matters less than presence. Some professionals put context first, then instruction, then input data, then output format. Some put input data (pasted content) first with the instruction right after. What matters is that AI can find the instruction quickly and that the context, input data, and output format are all in the prompt.
Depends on the tool. Some tools will refuse to send the message and show you a warning. Some tools will silently truncate (cut off part of your prompt without telling you). Some tools will trim the earliest parts of the conversation history to make room. That third option is why long chats forget: your recent messages get preserved, but the earliest ones drop off silently. Watch for signs AI has "forgotten" something you said earlier.
Your open question.
What is one prompt-anatomy question you still have after Week 11? Type it here. Your answer is saved locally to you.
Power Preview · The Prompt Template
The well-formed prompt template.
This is the template you can copy and use starting this afternoon. Five parts. Fill in each section. Every renewal letter, every marketing draft, every retention outreach, every operations question can use this shape.
The template · Copy this
[Instruction] [What you want AI to do. Verb + object. First eight words.]
[Context] [Who is this for? What is the situation? What tone? What background matters? One or two sentences.]
[Input data] [Any specific content AI should work with. Paste it or reference it here.]
[Output format] [What shape do you want the response? Length, structure, tone. One line.]
[Optional: Example] [One or two examples of the kind of output you want.]
One real example we will run in the next two slides.
Same task: draft a renewal letter for Maria Alvarez. Two prompts. First is what a beginner types. Second is well-formed. Same AI, same task, dramatically different result. That gap is what the whole workshop is teaching you to close.
Note before the runs
The outputs on the next two slides are composite examples of what these prompts typically produce. Actual AI output will vary run to run (temperature, remember). The point is the gap in quality, not any specific word choice.
Power Preview · Run 1
Run 1: The weak prompt.
The prompt (as a beginner might type it)
Write a renewal letter.
What AI produces (typical)
Dear [Resident Name],
We are pleased to inform you that your lease at [Property Name] is coming up for renewal. As a valued member of our community, we would like to offer you the opportunity to continue calling [Property Name] home.
Please review the enclosed renewal terms and contact our leasing office at your earliest convenience to discuss your options. We appreciate your residency and look forward to another wonderful year together.
Sincerely, [Property Management Team]
What is wrong with this.
Every field is a placeholder. [Resident Name]. [Property Name]. [Property Name] again. That is not AI being lazy. That is AI being honest. It did not know any of that because the prompt did not tell it.
The tone is corporate-generic. "Valued member of our community." "Another wonderful year together." Any property in the country could send this letter. Nothing anchors it to Maria, to Riverside on Broad, to a first-year resident, to a 3% increase.
The structure is default. Four paragraphs, a greeting, a sign-off. AI picked what its training saw most often for "renewal letter." Not what your team would send.
The lesson
This is not AI failing. This is AI succeeding at "write a renewal letter" with no other input. If you want it to look different, the prompt has to look different. Slide 14 shows how much different it looks.
Power Preview · Run 2
Run 2: The well-formed prompt.
The prompt (all five parts)
[Instruction] Draft a renewal letter.
[Context] Resident: Maria Alvarez, apartment 214 at Riverside on Broad. First-year renewal. Current rent $1,485, renewing at $1,530 (3% increase). Quiet resident, no work orders, move-in survey scored our property 5 of 5. Our voice: warm, personal, professional. Not corporate.
[Input data] Below is a renewal letter our team liked from last quarter (paste example letter).
[Output format] Two short paragraphs, 60-80 words each. No header, no bullets. Address Maria by first name. Do NOT mention the specific dollar amount of the increase. Reference her survey response naturally.
What AI produces (typical)
Hi Maria,
Your first year with us at Riverside on Broad has been such a pleasure. When you rated your move-in a perfect 5, our whole team took note. It is exactly the kind of experience we work to give every resident, and knowing we got it right for you means a lot.
We would love to have you stay another year. Your new lease renewal is ready for review, and I have included it below along with a few quiet updates for the coming year. Take a look and let me know what questions you have. I am here whenever you need me.
Warmly, The Riverside on Broad team
What changed. Only the prompt.
Name and property present. Because we put them in the prompt.
Reference to her actual experience. The 5-of-5 survey. Because we put it in the prompt.
Warm, personal tone. Because we specified that in context (and provided an example in input data).
No dollar amount. Because we explicitly asked AI not to include it in output format.
Shorter, no headers. Because we specified format.
Same AI. Same task. The gap is the prompt.
The lesson
A well-formed prompt is not longer for its own sake. It is longer because it includes the information AI needs to do the task well. The extra 90 seconds you spend on the prompt saves you the 15 minutes you would have spent editing a generic letter into something you would actually send.
Power Preview · Try This Today
One exercise you can run this afternoon.
You have 60 minutes today (or 20 if you are pressed). Do this exercise once. It will change how you write every prompt after.
The exercise · Prompt anatomy audit
Step 1. Pull up your last three AI conversations. Any tool.
Step 2. For each opening prompt, mark which of the five parts you included: instruction, context, input data, output format, example.
Step 3. Count what you left out. Most first-time auditors find they included instruction and maybe input data. They almost always missed context and output format.
Step 4. Rewrite one of those prompts using all five parts. Run it. Compare the output to what you got originally.
Step 5. Notice how much less editing the well-formed output needs. That is your permanent ROI on learning prompt anatomy.
A warning about what sounds good.
When you run a well-formed prompt, AI will produce fluent, confident text. Week 10 taught you that fluent and confident does not mean correct. That is still true. Read what AI gives you. Verify anything factual. Especially when your well-formed prompt made AI sound like an expert. That is when the sycophancy risk is highest.
Where to practice safely
The prompt anatomy audit is safer to run in a governed workspace.
If your organization is not sure where your team should practice prompt-writing on real resident data, this is the governed environment that is designed for it. Data classification, human review, and the RETTC principles are built in. Learn more from Multifamily NEXT at tami@mprodigitaledge.com. mPro Digital Edge: the AI workspace built for using AI with governance, not against it.
Practice this on us
You just spent 25 minutes learning to test confident AI claims. Test this one too. Ask us for a demo. Ask what "governance built in" actually looks like in practice. Ask for third-party reviews. Any AI workspace that survives those questions is worth considering. Any that does not, is worth walking away from. This is what Multifamily NEXT wants you practicing on every AI product, including this one.
Share this workshop with a coworker
If Week 11 helped, someone on your team probably needs it too. Send them the link.
Reference · Bookmark this slide
Vocabulary cheat sheet: 29 core terms.
Every AI term this curriculum has introduced so far, in one place. Plain-language definition + a multifamily read for each. Print it. Bookmark it. Come back to it whenever you need to remind yourself what a term actually means.
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, Copilot. Trained on vast text; predicts one word (token) at a time.
Token
Small chunk of text an LLM reads and writes. ~0.75 words per token. 1,000 tokens ~= 750 words.
Training Data
The vast collection of text (hundreds of billions of words, or more) an LLM learned patterns from. Not a searchable database. A compressed pattern archive.
Parameters
The internal numbers an LLM learned during training. Frontier models have hundreds of billions.
Temperature
Controls how random AI's word picks are. 0 = fully deterministic. Higher = more variety. Consumer AI runs somewhere in the 0.5 to 1.0 range by default (exact number rarely published).
Probability Distribution
The numerical likelihoods AI assigns to each possible next word.
Hallucination
When AI produces false information as if it were fact. Mechanism: prediction does not naturally admit uncertainty, so it fills the gap with a plausible-sounding answer.
Sycophancy
AI's trained tendency to agree with the framing of your question. Because agreement often felt more helpful to raters during training.
RLHF
Reinforcement Learning from Human Feedback. The most common training method used to shape modern AI toward feeling helpful. Humans compare responses during training; AI learns to produce responses that get preferred. A leading explanation for sycophancy. Newer training methods exist (Direct Preference Optimization and Constitutional AI are two you may hear named later in the curriculum), but the sycophancy pattern shows up across all of them.
Verification Prompt
A four-part addition to your prompts that forces AI to declare uncertainty, list disagreements, sort verifiable from asserted, and rate confidence. Taught in Week 10.
AI vs Search vs Database
Three different tools. AI generates. Search retrieves. Database stores. Confusing them causes most of the "AI got it wrong" problems in multifamily. Taught in Week 09.
Fair Housing and Governance
HUD
US Department of Housing and Urban Development. Federal agency that enforces Fair Housing law and publishes guidance on assistance animals, reasonable accommodations, and other rental housing rules. Website: hud.gov.
NIST AI RMF
National Institute of Standards and Technology's AI Risk Management Framework. Four functions: GOVERN, MAP, MEASURE, MANAGE. Government-published, voluntary, industry-standard.
RETTC AI Governance
Real Estate Technology and Transformation Center's eight-principle AI Governance Framework, written specifically for rental housing. rettc.org/frameworks/ai-governance.
AI Council
The internal group in your organization that reviews AI use cases, approves tools, and enforces classification rules. Also acceptable: Governance Team.
Data Classification (Tiers)
Tier 1 Safe for Any AI (public info), Tier 2 Enterprise Only (needs AI Council approval), Tier 3 Never (protected data). Taught in Week 1.
Use-Case Risk (Green/Yellow/Red)
Green Approved, Yellow Conditional, Red Prohibited. Independent from data classification. Answers "is this AI job allowed to run."
Systems and infrastructure
PMS (Property Management System)
Database of record for residents, apartments, leases, balances, work orders. Yardi, RealPage, Entrata, ResMan, AppFolio. Source of truth for anything about a specific resident or property.
CRM (Customer Relationship Management)
Database of prospects, applicants, communications. Knock, Anyone Home, LeaseHawk. Source of truth for prospect data.
Chatbot
Conversational AI tool. Consumer chatbots (ChatGPT, Claude) are general-purpose. Multifamily leasing chatbots have narrower system prompts to keep them on-task.
Prompt Framework
A repeatable structure for writing prompts. This curriculum teaches seven specific frameworks starting Week 14.
Retrieval
When an AI tool searches through documents, databases, or the web BEFORE answering, then uses what it found as part of the prompt context. Microsoft Copilot in Word/Excel/PowerPoint uses retrieval instead of loading full documents. Some ChatGPT, Claude, and Gemini modes also retrieve from the web or connected sources. When retrieval is on, the bright line still holds: AI only has what is in the container. Retrieval just changes how the container gets loaded.
New this week · 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 to the model before your prompt. Sets role, tone, format, and restrictions. You do not write it and cannot see it directly.
Conversation History
Every message earlier in the current chat. AI re-reads it every time. When it gets too long, oldest messages drop off. This is how "AI forgets."
Context Window
The maximum text AI can process in a single request. Prompt + system prompt + conversation history + response, all counted. Measured in tokens.
Prompt Anatomy (the 5 parts)
Instruction, context, input data, output format, examples. Not all five every time, but knowing them changes how you write.
Output Format
The shape you want AI's response in: length, structure, tone, what to include or omit. If you do not specify, AI picks. You may not like what AI picks.
The five parts in one column, common mistakes in the other. Read down the left when you write a prompt. Read down the right when a prompt fails.
Include these five
Instruction: Verb + object. First eight words.
Context: Who, situation, tone, background. One or two sentences.
Input data: The content AI should work with. Paste it.
Output format: Length, structure, tone. One line.
Example (optional): Show AI what "good" looks like.
Watch for these mistakes
Buried instruction: Instruction is somewhere in the prompt but not up front.
Missing context: Generic result because AI had nothing specific to work with.
No output format: AI defaults to a shape you did not want.
Assuming memory: AI "forgot" something because a new chat has no history.
Window overflow: Long chat drops earliest messages silently.
Practice this instinct on us, too.
Week 10 taught you to verify AI's confident claims. That instinct applies here too. Apply it to this workshop. We have been confident throughout. That is worth pushing on.
What Week 11 might be wrong about
1. "Five parts of a prompt" is a teaching frame, not a fixed rule. Different prompt-engineering guides split the anatomy differently: some name four parts, some name six, some use different labels. What matters is the shape (instruction + context + material + format), not the exact count.
2. Context window sizes on Slide 6 change every quarter. The numbers were accurate at the time this workshop was written but may already be different. Verify against your tool's current documentation before you plan a workflow around a specific token count.
3. "Most affordable large context" is a snapshot. Gemini's free-tier lead on window size is where things stand today. If ChatGPT or Claude expand their free tiers next quarter, that comparison flips. Check before you commit.
4. Retrieval and connectors are changing what "in the prompt" means. When a tool connects to your Google Drive or your email, information from those sources gets loaded into the container automatically. The bright line still holds (AI only has what is in the container), but the container gets loaded differently. This is a moving target.
5. We are confident everywhere. Same as Week 10. That is a feature of workshop writing, not a claim about certainty. If any specific claim struck you as too confident, that instinct is right. Verify it against your tool before you build a policy around it.
The habit we want you to keep
Verify any confident-sounding claim, including this workshop. If you spot something in Week 11 that does not match what your tool actually does, that is Week 10 working. Good.
Permanent · Standards Library
NIST AI Risk Management FrameworkNIST AI RMF (NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY AI RISK MANAGEMENT FRAMEWORK)A voluntary US government framework for managing risks associated with AI systems. Published January 2023 by the National Institute of Standards and Technology, a bureau of the US Department of Commerce. Not a law. Not regulation. A structured way to think about, measure, and manage AI risk that has become the de facto US standard for AI governance. The four functions: GOVERN (culture, policies, and accountability structures for AI risk). MAP (understanding context and categorizing what your AI system is). MEASURE (assessing and tracking AI risks in a specific implementation). MANAGE (prioritizing and responding to AI risks based on impact). Multifamily example: When your AI Council writes a policy that says "no autonomous decisions affect residents," that is GOVERN. When you categorize a leasing chatbot as "LLM with system prompt, Tier 2 data allowed," that is MAP. When you audit AI-drafted renewal letters quarterly for Fair Housing risk, that is MEASURE. When you decide which use cases get approved with guardrails vs. refused, that is MANAGE. Four functions. One rhythm. Link: https://www.nist.gov/itl/ai-risk-management-framework and RETTC AI Governance Framework.
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 11 concept lands somewhere in one or both.
The RETTC AI Governance Framework · all 8 principles
Official titles as published at rettc.org/frameworks/ai-governance. These are the 8 principles the Framework asks every rental housing provider and technology partner to adopt.
#
Principle
1
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.
2
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.
3
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.
4
Protect Privacy & Promote Data Integrity. Adhere to federal, state, and local consumer privacy laws, with an emphasis on data minimization, security, and informed consent.
5
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.
6
Promote Positive Renter Experiences and Outcomes. Elevate technologies that enhance affordability, access, and housing stability, and use AI to improve resident experiences.
7
Strive for Innovation with Responsibility. Encourage development of AI tools that benefit residents, prospects, guests, and operators, without compromising trust.
8
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.
This week bridged to the frameworks
Week 11 topic
NIST AI RMF anchor
RETTC AI Governance anchor
Prompt anatomy (5 parts)
Govern function. A written prompt structure your team uses across roles is a governance artifact. It makes AI use inspectable and consistent.
Principle 1 (Establish an Organizational Philosophy on AI Use and Oversight). How your organization structures prompts is part of your written AI policy.
System prompt (hidden)
Map function. Knowing what a tool's system prompt does is prerequisite to categorizing what the tool is and is not fit for.
Principle 3 (Uphold Transparency & Explainability). If a resident is affected by a chatbot decision, your organization must be able to explain what shaped the response, including the hidden system prompt.
Conversation history and context window
Measure function. Long chats silently lose earlier content. This is a measurable AI reliability limit that belongs in your team's risk assessment.
Principle 3 (Uphold Transparency & Explainability). Explain to a resident why AI "forgot" something they said earlier. The answer is a context-window limit, not a mystery.
The Week 11 bright line (if it is not in the prompt)
Govern function. The bright line is a policy, not a suggestion. It belongs in the AI Council's written standard for how the organization uses AI.
Principle 5 (Maintain Outcome-Based Accountability & Human Involvement). What went into the prompt is the record. The human who chose what went in is accountable.
Context (the missing piece in most prompts)
Manage function. Prioritize training on context-writing for roles that use AI on resident-specific tasks. This is where prompt quality determines output quality.
Principle 2 (Promote Fairness & Consumer Protection). A generic prompt that ignores individual context can produce a response that fails a specific resident. Fair Housing risk lives here.
Output format (making AI responses inspectable)
Govern function. Standard output formats across your team make AI-assisted work reviewable and comparable. Governance depends on legibility.
Principle 3 (Uphold Transparency & Explainability). Format-standardized responses are easier for a supervisor, an AI Council reviewer, or an auditor to check.
Pick the best answer for each. This tests whether Week 11 landed. Reset when you want to redo.
Scenario 1 · Prompt-anatomy failure
A leasing pro types: "So we have Sarah in 305, she has been here two years, and I am thinking maybe we should write her something about the renewal, could you help?" What is the primary problem with this prompt?
Buried instruction.
The instruction is there ("write her something") but it is hedged, softened, and buried in context. Front-load. "Draft a two-year renewal letter for Sarah in apartment 305. Here are the details..."
Scenario 2 · Prompt-anatomy failure
A property manager types: "Write a lease violation notice." AI produces a generic template. Manager is frustrated. What is the primary problem?
Missing context.
Generic prompt, generic output. AI needs to know: who is the resident, what specifically was violated, what tone (firm-but-fair? formal?), what your property's previous notices look like. Without context, AI defaults to average-of-training-data.
Scenario 3 · Context window failure
A regional manager has been planning a resident event with AI for 60 messages. She refers back to a specific catering quote from message 8. AI produces a different number. What likely happened?
Context window overflow.
In long chats, the oldest messages drop off the context window silently. AI is not lying. It just cannot see message 8 anymore. Fix: start a new chat and paste the specific quote at the top, or use a tool with a larger context window.
Scenario 4 · Prompt anatomy misconception
A leasing pro spent an hour last night crafting the perfect renewal template with ChatGPT. This morning she opens ChatGPT (new chat) and types: "continue with the template we made yesterday." What happens?
New chat, no history.
A new chat is a blank container. Nothing from yesterday's chat carries over. She either needs to open yesterday's chat and continue there, or paste the template she saved into today's new chat. Some tools have "memory" features but the default is: new chat = blank.
Scenario 5 · Context window nuance
A regional manager pastes a 400-page policy manual (roughly 300,000 words) into a chat with a tool that advertises a 1 million-token context window. She asks a question about page 200. AI gives a vague answer. Why?
Advertised context is not effective context.
Even a 1M-token window does not mean AI reliably uses every word in it. NVIDIA's RULER benchmark and Stanford's "Lost in the Middle" paper both documented that models perform worse on information in the middle of very long contexts. Put important details near the beginning or end, not buried in the middle.
Scenario 6 · System prompt discovery
A leasing chatbot on your property's website always steers every conversation toward "let me schedule a tour." A resident with a maintenance question feels unheard. Why does the chatbot behave this way?
System prompt shaping.
Every AI product has a hidden system prompt that tells the model what to prioritize and what to avoid. That is why the same underlying model (say, GPT-5.6) behaves differently in different products. Ask your leasing chatbot vendor to explain the system prompt's governing logic: its priorities, restrictions, tone, and format rules. Reputable vendors will explain the logic even if they keep the exact wording proprietary. Refusal to explain the logic (as opposed to refusing to show exact text) is the governance concern.
Scenario 7 · Well-formed prompt
A maintenance supervisor types: "Draft a resident notification about tomorrow morning's water shutoff. Building A only. 8am to 11am. Warm tone. Two short paragraphs, no bullets. Include a courtesy apology." What is this prompt?
Well-formed prompt.
Instruction is front-loaded ("Draft a resident notification"). Context is present ("Building A, 8am-11am, warm"). Output format is named ("two short paragraphs, no bullets, courtesy apology"). Example is optional and not needed here. This prompt will produce a usable draft on the first try.
Scenario 8 · Input data judgment
An accounting analyst wants AI to help summarize a resident's payment history question. Which is the better approach?
Paste only what is relevant.
Pasting a 5-year ledger burns context window on details AI does not need for this question and increases the risk of information getting lost in the middle. Paste the relevant slice + a sentence of context. Also: check your data classification tier before pasting resident financial data into a consumer AI tool.
Take Action · Anonymous Quiz
Anonymous quiz. Five questions.
No names collected. Answers help Multifamily NEXT calibrate what to teach in Weeks 12-52. Takes 90 seconds.
Take Action · Policy / Action Download
Take one thing back to your team.
Two artifacts. One document. One email. Both ready to send.
Artifact 1 · The Prompt Anatomy Quick-Reference (PDF)
A one-page reference showing the five parts of a well-formed prompt, the three most common failure patterns, and a template your team can copy-paste. Suitable for printing and keeping at every desk.
If your supervisor has not seen Week 11 yet, send them this. The mailto link opens in your email client with a draft already filled in. Add your supervisor's address and send.
Draft (edit before sending)
Subject: Week 11 of AI Tuesdays · The anatomy of a prompt
Hi [supervisor],
I just finished Week 11 of the Multifamily NEXT AI Tuesdays workshop series. This one covers what AI actually receives when you send a message: your prompt, plus a hidden system prompt, plus the conversation history. Everything else is invisible.
The workshop teaches five parts of a well-formed prompt and three fixes that would immediately improve how our team writes prompts to AI tools. It also explains why long chats "forget" things (the context window fills up and drops the oldest content).
Would love to have our team walk through this together. The workshop is free at multifamilynext.com. Password 1152.
Note: The email is drafted for you but the recipient field is left blank on purpose. Add your own supervisor's address. Multifamily NEXT never pre-fills recipients.
Take Action · Session Complete
Good job. Session Complete.
You walked away with
✔ A working mental model of what AI actually sees when you hit send (your prompt + system prompt + conversation history, nothing else).
✔ The five parts of a well-formed prompt: instruction, context, input data, output format, examples.
✔ The bright line for every prompt: if it is not in the prompt, AI does not have it.
✔ A copy-ready well-formed prompt template you can use this afternoon.
✔ Current context window sizes for ChatGPT, Claude, Gemini, and Copilot (with caveats about how fast those change).
✔ Six new vocabulary terms (Prompt, System Prompt, Conversation History, Context Window, Prompt Anatomy, Output Format), plus "Retrieval" as a supporting concept, bringing your cheat sheet to 29 terms.
The board paragraph.
If your CEO asked you at Monday's meeting what our team learned in Week 11, here is what you say:
Our team learned that when AI produces a bad response, the problem is almost always in the prompt, not in AI. AI sees three things when we hit send: what we typed, a hidden system prompt from the tool, and whatever came earlier in the conversation. Everything else, including our residents' files, our property policies, and our own tone preferences, is invisible unless we include it. So we are shifting how our team writes prompts. Every prompt has five parts: instruction up front, context specific to the situation, any material AI should work with, the format we want, and sometimes an example. We are also now aware of the context window (how much AI can see in one conversation) and what happens when it fills up. This means fewer generic drafts, less "why can AI not remember," and better use of the time we already spend with AI tools. Two artifacts are ready to go: a one-page prompt template and an anatomy quick-reference. Next Tuesday, Week 12 covers writing prompts that get you to a usable answer on the first try.
Where to practice this safely.
Governance handoff · mPro Digital Edge
Where to practice this safely.
mPro Digital Edge is the enterprise AI workspace for multifamily: a governed environment where NIST AI Risk Management Framework and the RETTC 8 Principles are built into every tool. Your team can practice prompt anatomy on real material without worrying about which tier of data is in the prompt. Human review is built in. AI Council approvals are tracked. That means the answer to "where can our team try this?" has a real place.
One platform where every person in the organization gets measurably better at their job with every interaction.
Now that you know what goes in a well-formed prompt, Week 12 teaches you how to be specific. Which words matter. Which words hurt. The eight things every prompt should say. How to ask for what you actually want.
See you Tuesday. Trust me, this one levels up your leverage on AI.
Tami Siewruk
CEO, Multifamily NEXT and mPro Digital Edge
Built by Multifamily NEXT. Powered by mPro Digital Edge, the AI workspace for governed AI use in multifamily.
