Three quick recall questions from last Tuesday's workshop on RTF (Role, Task, Format). Tap each question to reveal the answer, then tap it again to hide. This is just for you, no scoring, no submission.
Answer: Role, Task, Format. Role tells AI who it is being. Task names the verb, object, and audience. Format specifies length, structure, and tone.
Example of use
Role: You are a friendly on-site property manager. Task: Draft a 5-line email replying to a resident who complained about the pool being closed for maintenance. Format: Warm but firm tone, no promises about a reopen date, ends with an invitation to stop by the office.
Answer: Format. Skipping it makes AI default to about 300 words of average corporate prose. Naming the shape is the biggest single lever on the answer you get back.
Example of use
Without Format, "draft a resident welcome email" gets you 400 words of generic corporate-speak. Add "Format: 3 short paragraphs, warm and casual, no sales language, ends with an invitation to stop by the office" and the answer arrives close to what you actually wanted.
Answer: Bracket them. "[Resident name]" and "[Balance]" go into the RTF prompt. AI drafts the template. You fill the sensitive slots in by hand. Structure never changes what data is allowed in.
Example of use
You need a payment-plan email for a resident in apartment 214 who owes $2,847. Instead of pasting the real name and balance into a public AI tool, you send: "Draft a payment-plan email to [Resident name] in [Apartment] for a past-due balance of [Balance]. Warm tone. Offer three options." AI drafts the template. You paste the real details into your document afterward, in your own PMS or your inbox.
Missed last week?
Open Week 14 · RTF before you start this one. Password on that page is 1452. This week builds on it.
The Workshop · Opening
Show your work.
Last week we shaped the output. RTF told AI what the answer should look like. This week we shape the reasoning. Chain of Thought tells AI to walk you through how it got there before it gets there.
This matters because AI is very good at producing an answer that sounds right. Confident, well-written, delivered in the exact shape you asked for. And sometimes wrong. When the stakes are high enough that "sounds right" is not good enough, Chain of Thought is how you check its work.
This week in one line
Ask AI to show its work before it gives you the answer. Then check the work.
Read this workshop as
Tap the role closest to yours. A few examples throughout the workshop 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. Come back later if you have to. Your notes and progress save to this device automatically.
1
What Chain of Thought is
Slides 4 and 5. Show your work in one line.
~4 min
2
When to use it and when to skip it
Slides 6 and 7. Not every task earns the extra tokens.
~4 min
3
CoT stacked with RTF
Slides 8 through 10. Two structures, one prompt.
~4 min
4
Power Preview: two examples
Slides 12 through 15. Vendor bid comparison and a concession-versus-turnover call.
~6 min
5
Reference and reinforcement
Slides 16 through 19. Vocab, CoT reference card, standards, Spot the Tier.
~4 min
6
Quiz, downloads, and close
Slides 20 through 22. Five questions, two downloads, Week 16 teaser.
~3 min
No timer runs. The minute estimates are honest averages that count reading and interaction time. Slower is fine. Reading twice is better.
The Workshop · Why You Should Care
When a wrong-sounding-right answer costs you.
The most dangerous AI answer is not the one that sounds wrong. It is the one that sounds right and is not. Fluent. Confident. Delivered in exactly the shape you asked for. And built on a mistake in the reasoning three steps back that you cannot see.
Scenario 1
The vendor bid that looks cheaper
Two janitorial bids. Vendor A: $2,400 per month. Vendor B: $2,650 per month. You ask AI which is cheaper. It says Vendor A. Correct math. Wrong answer. Vendor A billed monthly with a 12-month commitment; Vendor B billed quarterly with no commitment and includes floor waxing that Vendor A charges $400 extra for. Real cost of Vendor A is higher. AI never surfaced any of that because you never asked it to show its work.
Scenario 2
The concession that felt generous
A good long-term resident asks for $75 off to renew. You ask AI whether the concession is worth it. It says yes, keep the resident. Sounds obvious. Then a real calculation: $75 times 12 months is $900. The turnover would have cost about $2,800. Correct answer. But AI could just as easily have said no on a hunch that a $900 write-down was too high, without doing that math. When AI is guessing instead of calculating, you cannot tell.
Here is the thing
AI does not tell you when it is guessing. It sounds the same either way. Chain of Thought is how you make it separate the two. When it shows the work, you can see whether the work exists.
The Workshop · Part 1 of 6
Chain of Thought, in one line.
Chain of Thought is one added instruction. You tell AI to show its reasoning step by step before it gives you the final answer. That is the whole trick. Nothing else about the prompt changes.
Without CoT
"Which vendor bid is the better deal for a 200-apartment community?"
AI gives you a single sentence answer. "Vendor A is the better deal." No breakdown. No math. No visible trail. If it is wrong, you cannot see where it went wrong.
With CoT
"Which vendor bid is the better deal for a 200-apartment community? Think through it step by step before you give me your final answer."
AI walks you through the annual cost of each bid, the commitment terms, the scope differences, and what those add up to. Then the recommendation. Now you can check the math on the way to the conclusion.
Why it works
When you ask for the final answer directly, AI reaches for the shortest path to a plausible-sounding sentence. It is very good at plausible. It is less good at correct.
When you ask for the reasoning first, you change the shape of the response. AI now produces the intermediate steps out loud. Two useful things happen at once: the extra steps often catch errors before the final answer is written, and you get to see the logic so you can catch anything AI missed.
Chain of Thought is not a special mode. It is one sentence you add. That is it.
The Workshop · Part 2 of 6
The trigger phrases.
You do not need magic words. Any phrasing that clearly asks AI to show its work before the conclusion will do it. These are the ones that hold up across every AI tool.
Six phrases that work
1. "Think step by step."
2. "Show your reasoning before you give me the answer."
3. "Walk through your logic first, then conclude."
4. "Explain how you got there, then give me your recommendation."
5. "Break this down before you answer."
6. "List the assumptions you are making, then work through it."
The sixth one is worth a second look. Asking AI to name its assumptions before the reasoning starts catches a specific kind of mistake: the answer that is technically correct given assumptions AI made up on the spot. If the assumptions are visible, you can correct them. If they are hidden inside a confident answer, you cannot.
Try one now
Think of a decision you would ask AI about this week. Write out how you would add a Chain of Thought trigger to that prompt. Pick one of the six above. No answer to send anywhere, just get the sentence into your head.
The Workshop · Part 3 of 6
When to use it.
Chain of Thought earns its keep on tasks where the reasoning matters and a wrong-sounding-right answer costs you something. Four situations where it is the right call:
Four times to reach for CoT
1. Any task with real math. Concession math. Vendor cost comparisons. Renewal versus turnover. NOI variance. Deposit disputes. If a number matters at the end, ask AI to show the arithmetic. You catch the errors AI would have hidden inside a clean-looking recommendation.
2. Comparisons with tradeoffs. Two options that are not clearly better or worse without breaking them down. Two vendors, two hires, two capex projects, two pricing strategies. CoT forces AI to list what each option gives up, not just what it offers.
3. Multi-step decisions. When getting to the answer requires three or more steps, each of which has to be right. Fair Housing edge cases. Escalation calls. Anything where the sequence matters.
4. "Why" questions. When you need to know the reason, not just the answer. "Why is delinquency up this month?" "Why did that resident leave a bad review?" The reasoning IS the answer.
Match it to your role
Name one recurring task in your role that would benefit from CoT. Which of the four situations above does it fit? Write both.
The Workshop · Part 4 of 6
When to skip it.
CoT is not free. It makes AI produce more text, which takes longer to read, longer to review, and longer to work with. On the wrong task, you have made yourself slower for no benefit. Four cases where a plain RTF prompt is better:
Four times to skip CoT
1. Writing tasks with no reasoning to check. Draft this email. Write these three subject lines. Rewrite this paragraph. The output is the point. There is no reasoning chain that needs to be visible.
2. Quick edits or short generations. "Make this shorter." "List five options." "What is another way to say this?" The task is small, the risk is low, the reasoning is not interesting.
3. When you already know the answer. Sometimes you are asking AI to draft the writeup of a decision you have already made. In that case you do not want its reasoning, you want its words. Skip CoT.
4. When the reasoning would waste your time. If you are asking a factual question with a short answer, do not force AI to lay out three paragraphs of thinking before it gives it to you. "What is the standard notice period in this state?" needs an answer, not a walkthrough.
Rule of thumb: if you are going to check the answer against real numbers or a real policy before you act on it, use CoT. If you are just going to send what comes back, do not.
The Workshop · Part 5 of 6
CoT plus RTF, stacked.
RTF from last week and Chain of Thought from this week are not two competing methods. They stack. RTF sets the role, the task, and the shape of the answer. CoT adds one instruction that changes how AI gets there.
Sample stacked prompt
Role: You are a regional manager weighing a vendor decision for one property in the portfolio.
Task: Compare two janitorial bids for a 200-apartment community and recommend one. Bid A is $[XX,XXX]/year, monthly billing, 12-month commitment, standard scope. Bid B is $[XX,XXX]/year, quarterly billing, no commitment, includes floor waxing.
Format: A short recommendation with a one-paragraph rationale.
Chain of Thought: Before you recommend, walk through the true annual cost of each bid including the scope differences, then compare. State any assumptions you are making about waxing frequency or scope equivalence.
The order matters. RTF first, CoT last. RTF frames the task; CoT tells AI how to work through it before answering. Every part of RTF still does its job (role sets the voice, task names the deliverable, format sets the shape). The Chain of Thought line rides on top.
The two habits, together
Every real work task: use RTF. Role, task, format.
Any real work task with math or judgment or comparisons: add CoT on the end. Four extra words. Very different answer.
The Workshop · Where AI Goes Wrong
The bright line, again.
Two weeks ago on Slide 9 of Week 13, we drew a line: prompt structure improves the words, not the data rules. Same line last week. Same line this week. Same line every week of Phase 3.
The bright line
Making AI show its reasoning does not change what data is allowed in.
A CoT prompt is easier to check. It is not a safer place to paste sensitive data. If a resident's real name and balance should not go into a public AI tool as an RTF prompt, they should not go in as a CoT prompt either. Structure is structure. Data is data.
Right way
Bracketed CoT prompt
RTF + CoT with brackets for every real number and every real name. AI walks through the logic using placeholders. You substitute in the real values when you apply the recommendation.
Wrong way
"Show me your math" is not a permission slip
Same prompt, but you pasted the real resident's name, real balance, real unit number, and asked AI to "think step by step." The structure got smarter. The data still exposed a real resident to a general AI tool. Rule unchanged.
Full data classification treatment: Week 1. Practice on Slide 19 of this workshop.
The Workshop · Default Rule
The default: reach for CoT when it matters.
Not every prompt earns Chain of Thought. Most of your day is short writing tasks and quick edits where RTF alone is enough. But the moments when the reasoning matters are also the moments when getting it wrong costs the most. Those are the CoT moments.
The Week 15 default rule
1. Start every real work prompt as RTF. Role, task, format.
2. If real math, real judgment, real comparisons, or a real "why" is involved, add a CoT line at the end. Four extra words.
3. When the reasoning comes back, actually read it. If the walk-through would take longer to check than doing the work yourself, that task did not need CoT.
4. Sensitive data still uses brackets. CoT plus brackets is the safe combination.
Trust me
The first time you use CoT on a decision that had been making you nervous, you will notice something specific. Reading the reasoning is where you actually decide whether to act on the answer. That is the point. AI drafted the thinking. You are still the person doing the deciding.
Guess Green, get seen. Choose Yellow, stay mellow. Red? Use your head.
Asking AI to show its reasoning does not change the tier of your data. If you are unsure, treat it as Yellow.
The Workshop · Q & A
Questions people ask about Chain of Thought.
Tap any question to expand the answer. These are the honest questions that come up when CoT stops being a slide and starts being something you use.
Answer: It works, and it is not a superstition. When you ask AI to produce reasoning first, it produces more intermediate content, and that intermediate content changes what the final answer looks like. On tasks with real logic (math, comparisons, multi-step reasoning), the accuracy of the final answer measurably improves. Google DeepMind and other AI research groups have documented this. You do not need to trust the research.
Example of use
Ask AI which of two vendor bids is better without CoT. Then ask the same question with "Think step by step" on the end. Watch the second answer walk the annual cost, the scope difference, and the commitment terms before it picks. Watch the first one just pick. That is the difference.
Answer: Newer models do it more often on their own than older ones, especially for math problems. But "more often" is not "always," and the reasoning you get without asking is not always visible to you (some models show it, some hide it). Asking explicitly guarantees you get the walkthrough on the page where you can read it. Cheap habit, high return.
Example of use
Same concession math question, sent to the newest model without a CoT trigger, may give you the reasoning in a hidden "thought" block that never appears on your screen, so you cannot check it. Add "Think step by step. Show the math." and the walkthrough always renders. Now you can verify $75 × 12 = $900 matches your local vacancy loss estimate.
Answer: Yes, this is the real risk. Wrong reasoning that sounds right is the trap. Two things help: the assumptions-first trigger (question 6 on Slide 5) forces AI to state what it is taking for granted, so you can catch the wrong ones. And treating the reasoning as a draft to check, not a conclusion to accept, keeps you in the loop. If you are not going to check it, CoT is the wrong tool for that task.
Example of use
AI says: "Step 1: standard vacancy loss is 45 days. Step 2: at $1,650 per month, that is roughly $2,475 in lost rent." Sounds right. But your market runs closer to 25 days, not 45. Because AI stated the assumption, you can catch it and correct: "Actually assume 25 days." The recalculated answer might change the decision. Without the stated assumption, you would have accepted a wrong number wrapped in confident math.
Answer: Yes, a little. AI has to produce more words. On a modern chat AI, we are talking about a few extra seconds, not minutes. If you are running many CoT prompts against an API on a schedule, the token cost matters and you may want to be more selective. For a person at a desk making one decision at a time, the extra seconds are trivial compared with the check.
Example of use
A bare vendor comparison prompt might come back in 3 seconds with one paragraph. The same prompt with CoT might come back in 8 seconds with five numbered steps and the recommendation. Five extra seconds. In exchange, the difference between signing the wrong 12-month janitorial contract and catching it.
Answer: You can, and sometimes it is a fine iteration move. But asking after is not as reliable as asking before. When AI has already committed to an answer, the "reasoning" you get back is often a justification for that answer rather than the actual chain that produced it. Ask before, and you get the walk-through that leads to the answer. Ask after, and you get a story that supports it. Both are useful; only the first is real Chain of Thought.
Example of use
You ask: "Which of these two apartment amenities upgrades would improve retention more, a gym refresh or upgraded package lockers?" AI says "package lockers." You then ask "Why?" and get three plausible reasons. But if you had asked "Walk through the retention drivers step by step first, then recommend" from the start, AI might have surfaced that resident survey data on your portfolio actually rates fitness higher than package handling, and the answer flips.
Answer: No. Slide 9 said it and it is the answer here too. Chain of Thought makes the prompt easier to check. It does not change what data is allowed to enter a general AI tool. Real resident names, real balances, real unit numbers still stay out. Use brackets. Get the reasoning. Fill in the real values yourself.
Example of use
Instead of "Sarah Martinez in unit 214 owes $2,847, walk through step by step whether to file for eviction," send: "Resident with [X months] tenure, $[Balance] past due, [payment history summary], walk through step by step how to weigh eviction vs a payment plan. State assumptions." AI walks the reasoning. You apply it to Sarah's real file in your PMS.
Power Preview · Setup
Watch CoT do the work.
The next two slides each show the same task twice. Bare prompt first. RTF plus Chain of Thought second. Same AI, same session, only the prompt changes. Both are decisions where the reasoning is the whole point.
Slide 13
Example one: vendor bid comparison
Two janitorial bids for a 200-apartment community. Different prices, different terms, different scopes. The bare prompt tells you which is cheaper. The CoT prompt tells you which is actually a better deal, which is not the same thing.
Slide 14
Example two: concession versus turnover
A long-term resident asks for a small concession to renew. Bare AI shrugs and picks. CoT walks through the true cost of losing them, then the answer becomes obvious. And sometimes the obvious answer is the opposite of the shrug.
Reminder before you scroll
Both examples use bracketed placeholders for dollar amounts and no real resident or vendor names. Nothing sensitive was pasted. The bright line from Slide 9 holds here too.
Power Preview · Example 1 · Vendor Bid Comparison
Which bid is actually the better deal?
Two janitorial vendors have bid on the same 200-apartment community. Bid A quotes $2,400 per month, monthly billing, 12-month commitment, standard scope. Bid B quotes $2,650 per month, quarterly billing, no commitment, and includes floor waxing quarterly. Standard scope on Bid A does not include waxing; that would add roughly $400 per quarter as an add-on.
Bare prompt · the whole thing you sent
Bid A is $2,400/month, Bid B is $2,650/month. Which one is the better deal?
What AI sent back
Bid A is the better deal. At $2,400 per month versus $2,650 per month, Bid A saves you $250 per month, or $3,000 per year. I would recommend going with Bid A.
Confident. Wrong. AI compared the sticker prices and ignored everything else because you did not ask about anything else. The scope difference alone flips the math. You would have signed with A.
RTF + CoT prompt · the whole thing you sent
Role: You are a regional manager weighing a janitorial vendor decision for one 200-apartment community.
Task: Compare Bid A ($2,400/month, monthly billing, 12-month commitment, standard scope with waxing available as a $400/quarter add-on) with Bid B ($2,650/month, quarterly billing, no commitment, includes quarterly waxing). Recommend one.
Format: One-paragraph recommendation with the true annual cost of each shown.
Chain of Thought: Before you recommend, work through the true annual cost of each bid including scope differences and commitment terms. State any assumptions you are making.
What AI sent back
Step 1. Bid A base cost. $2,400 × 12 = $28,800 per year for standard scope.
Step 2. Bid B base cost. $2,650 × 12 = $31,800 per year, waxing included.
Step 3. Scope adjustment. To make Bid A comparable, add quarterly waxing at $400 × 4 = $1,600 per year. Adjusted Bid A: $28,800 + $1,600 = $30,400.
Step 4. True annual comparison. Adjusted Bid A $30,400 versus Bid B $31,800. Difference is $1,400 per year, or about $117 per month.
Step 5. Non-price terms. Bid A locks you in for 12 months. Bid B has no commitment. If service quality is unproven, that flexibility has real value.
Assumptions. Waxing is needed at Bid B's quarterly cadence at both properties. Both bids cover the same square footage and hours. No hidden fees on either.
Recommendation: On price alone, adjusted Bid A is $1,400 cheaper per year. But that gap is small enough that Bid B's no-commitment terms are probably worth it for the first year while you evaluate service quality. Start with Bid B. Revisit at renewal.
Now you can see the math. Now you can see the assumptions. Now the recommendation makes sense, and if you disagree with the assumption about waxing frequency, you know exactly where to push back. That is what CoT gives you.
Power Preview · Example 2 · Concession vs Turnover
Is the concession cheaper than losing them?
A resident has been on-site three years, always pays on time, never a complaint. Their renewal is coming up. They asked for $75 off per month to sign the next 12. You are trying to decide whether to say yes. All numbers below are sample only, all bracketed.
Bare prompt · the whole thing you sent
Good resident wants $75 off to renew. Should I do it?
What AI sent back
Yes, retaining a good resident is almost always worth a modest concession. A $75 monthly reduction over 12 months is $900, which is generally less than the cost of turnover. Approve the concession.
The answer happens to be right. The reasoning is a shrug. "Generally less than" is not a calculation. Next time the numbers are different, this pattern gives you a confident wrong answer.
RTF + CoT prompt · the whole thing you sent
Role: You are an on-site property manager thinking about resident retention.
Task: A resident who has been on-site three years, always pays on time, no complaints, has asked for $[75] off per month to sign a 12-month renewal. Their current rent is $[1,650]. Recommend whether to approve the concession.
Format: A recommendation with the dollar math shown.
Chain of Thought: Before you recommend, calculate: total 12-month cost of the concession, plus estimated cost of turnover if they leave (vacancy days, make-ready, marketing, leasing time). Compare the two. State your assumptions about vacancy loss and make-ready cost.
What AI sent back
Step 1. Concession cost. $75/month × 12 months = $900 over the renewal term.
Step 2. Vacancy loss if they leave. Assuming 30 days average vacancy on a $1,650/month apartment, that is roughly $1,650 in lost rent.
Step 3. Make-ready cost. Assuming a standard turn (paint, carpet clean, punch list) for a well-kept 3-year resident who has caused no damage, roughly $[600] to $[900]. Use $750 midpoint.
Step 4. Marketing plus leasing time. Roughly $[200] in marketing spend plus about 8 hours of leasing time to show, screen, and sign. At an estimated cost of $[400], that is roughly $600 total.
Step 5. Total turnover cost. $1,650 + $750 + $600 = about $3,000.
Step 6. Compare. Concession $900 versus turnover $3,000. The concession is roughly one third the cost of the alternative.
Assumptions. 30 days vacancy is a normal-market estimate, not slow-market. Make-ready assumes no damage. Marketing spend and leasing time are your local averages, which you should verify.
Recommendation: Approve the $75/month concession. Even under conservative assumptions, retaining this resident is roughly $2,100 cheaper than turning the apartment. If your market vacancy is running longer than 30 days, the gap widens further.
Same recommendation as the bare prompt. Completely different confidence in it. And now if your make-ready number or your local vacancy assumption is different, you know exactly which step to correct without redoing the whole prompt.
Power Preview · Try This Today
One CoT drill you can run this afternoon.
Fifteen minutes. One real decision. Bare prompt first, then RTF plus CoT. You will feel the difference the moment the reasoning starts scrolling.
The drill
1. Pick a real decision you are weighing. Vendor pick. A concession call. A staffing headcount question. Anything with a number or a tradeoff. Nothing with sensitive resident data.
2. Send the bare version first. One sentence. What you would have sent last week. Read the answer. Notice how much of the reasoning is missing.
3. Send the RTF plus CoT version second. Same task, in a fresh chat. Add "Think step by step. State your assumptions before you recommend" on the end.
4. Compare. Not just the two answers, the two explanations. Which one gave you something you could actually push back on? Which one would you defend to your regional?
Good job
The first time CoT changes an answer you were about to accept, you will feel it in your gut. That is the whole reason we are teaching this. You are still the person deciding. AI is now the person showing you the math.
Where to practice safely
If your team is ready to run this drill against real portfolio data instead of bracketed placeholders, do it inside mPro Digital Edge. Multifamily-specific AI, AI Council rules travel with the tool, sensitive data stays inside your organization. Anywhere else, brackets stay in the prompt.
Log the drill
Once you run the two prompts, come back and note: What decision did you ask about? Did the CoT version change your answer, or just your confidence in it?
Reference · Bookmark this slide
Vocabulary cheat sheet: 39 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."
This week · Chain of Thought (Week 15) *NEW
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.
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 inside your prompt or your workspace that limits what AI can do. "Do not invent numbers" is a guardrail. Full treatment: Week 17.
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.
Read the left column when you are writing a prompt. Read the right when you want to see it in the wild.
The template
Start with your RTF prompt. Role, Task, Format. Same as last week. No change to the first three lines.
Add one Chain of Thought line at the end.
"Think step by step. State your assumptions before you give me the recommendation."
Any of the six trigger phrases from Slide 5 will do the same thing. Pick one that fits how you talk.
Filled in
Role: You are a regional weighing a vendor decision.
Task: Compare Bid A ($[XX]/mo, monthly, 12-mo commitment, standard scope) with Bid B ($[XX]/mo, quarterly, no commitment, waxing included). Recommend one.
Format: One-paragraph recommendation with true annual cost of each shown.
CoT: Think step by step. State your assumptions before you recommend.
When to reach for it
Use CoT for real math, comparisons with tradeoffs, multi-step decisions, and "why" questions where the reasoning is the answer.
Skip CoT for writing tasks with no reasoning to check, quick edits, generations you would send as-is, and factual questions with short answers.
Watch for these traps
The confident-but-wrong walkthrough. AI can produce plausible reasoning built on a bad assumption. Ask for assumptions to be stated explicitly, then read them.
Asking after instead of before. "Explain your answer" after AI already answered gives you a justification, not the real chain. Ask before.
Real resident data. Reasoning transparency is not a permission slip. Brackets stay in the prompt.
The bright line, always
Reasoning transparency is not a permission slip for sensitive data. See Slide 9.
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: Your AI Council reviews a new leasing chatbot. They walk NIST's four functions: GOVERN (who owns approval and off-boarding), MAP (Fair Housing risks, resident data exposure), MEASURE (monthly bias audit on chat transcripts), MANAGE (what happens if the bot gives a discriminatory answer). One framework, four questions, one decision on the record.
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 resident is denied a payment plan after AI flagged their file. RETTC's Uphold Transparency & Explainability principle says you must give that resident a clear, plain-language explanation of any AI-influenced factor in the decision, in a form they can understand and challenge. Not “the system said so.”
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 15 concept about Chain of Thought lands somewhere in one or both.
Week 15 · Chain of Thought
NIST AI RMF anchor
RETTC AI Governance anchor
Asking AI to show its reasoning before answering
MEASURE function: you cannot evaluate what you cannot see. CoT makes AI's intermediate steps observable, which is the prerequisite for measuring output quality.
Uphold Transparency & Explainability: people affected by AI-influenced decisions should be able to understand how the answer was reached. CoT is that explanation at the point of use.
Stating assumptions before conclusions
MAP function: understanding the context an AI system is operating in, before you rely on its output. Named assumptions are visible context.
Establish an Organizational Philosophy on AI Use and Oversight: the philosophy includes when human review of AI reasoning is required. CoT makes that review possible.
Catching wrong-sounding-right answers
MANAGE function: actively mitigating AI risks, not just identifying them. Reviewing the chain catches errors the final answer would have hidden.
Promote Fairness & Consumer Protection: a wrong answer that affects a resident, an applicant, or a decision on their file is a fairness risk. Visible reasoning is how you catch it before it propagates.
CoT + RTF stacked for real decisions
GOVERN function: policy defines what kinds of AI-assisted decisions require documented reasoning. High-stakes decisions require the chain, not just the conclusion.
Encourage Third-Party Partner Accountability and Due Diligence: if a vendor's AI tool cannot show how it reached a decision, that is a due-diligence gap. CoT sets the standard from your side.
The bright line (structure does not change data rules)
GOVERN function: data policy sets what may enter AI, independent of prompt sophistication.
Protect Privacy & Promote Data Integrity: no prompt structure, however transparent, changes whether resident PII or protected-class data may enter a general AI tool.
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.
Eight Chain-of-Thought prompts a real person might write this week. Your job: read each one, then tap the tier that matches what the person put into the prompt. Structure never changes the tier. Only the data does.
Tier 1 · Green: general information, templates, no PII. Tier 2 · Yellow: real business context but no resident identifiers or financials. Tier 3 · Red: real resident names, real balances, real medical, real applicant status, real portfolio financials.
Scenario 1
"You are a regional weighing two janitorial bids. Compare Bid A ($[XX,XXX]/year, monthly billing, 12-month commitment, standard scope) with Bid B ($[XX,XXX]/year, quarterly billing, no commitment, includes waxing). Think step by step. State any assumptions."
Green. All numbers bracketed, no vendor names, no property named. AI walks the reasoning with placeholders. You substitute real figures when you act on the recommendation.
Not quite. Every sensitive figure is bracketed. Nothing here identifies a real vendor, property, or account. Green.
Scenario 2
"You are an on-site manager. Sarah Martinez in unit 214 is asking for $75/month off to renew. She has been here 3 years. Her current rent is $1,650 and her payment history shows two late payments in 2024. Walk through whether the concession makes sense step by step."
Red. Real name, real apartment, real tenure, real rent, real payment history. Textbook Tier 3 data. Great CoT structure. Wrong prompt to send anywhere but an approved enterprise workspace. Bracket everything, keep the CoT, send again.
Look again. Name plus apartment plus tenure plus rent plus late-payment history all point at one identifiable resident. CoT does not rescue that. Red.
Scenario 3
"You are a leasing pro. Walk me through, step by step, how to evaluate whether a request for a comfort animal is a reasonable accommodation under Fair Housing. Format: a numbered checklist a leasing team could follow, no specific applicant details."
Green. General Fair Housing education, no applicant, no property. This is exactly the kind of task CoT was built for. Note: AI output on Fair Housing is a starting point, not legal advice. Route the checklist through your compliance function.
Not quite. Generic education, no real applicant. Green.
Scenario 4
"You are a portfolio operations analyst. NOI at [Property Name] came in $[XX,XXX] under budget this month. Walk me through the most likely drivers step by step, in order of typical impact size. Use bracketed placeholders."
Green. Everything bracketed. AI produces a reasoning framework you can apply to real numbers later. This is the pattern: get the analytical structure from AI, plug the real figures in yourself.
Look again. Every real number is a bracket. Green.
Scenario 5
"You are a property manager. Resident David Chen has a $2,847 balance across 3 months. Walk through step by step whether we should file for eviction or offer a repayment plan. His employer is Regional Hospital; his prior 18 months were on time."
Red. Name, balance, employer, and payment pattern combined identify a specific resident and put a decision about their housing into a general AI tool. Eviction risk decisions are also a category where AI output can create Fair Housing exposure if applied inconsistently. Bracket everything and route the general framework through your team.
Real name plus real balance plus real employer plus payment history plus a life-affecting decision. CoT cannot save this. Red.
Scenario 6
"You are a maintenance supervisor. Walk me through step by step how a technician should decide whether an HVAC service request is emergency (same-day) or routine (48-hour window). Use bracketed placeholders for any specific details."
Green. Generic SOP with bracketed placeholders. No property, no resident, no work order number. Standard operational documentation is safe for public AI tools with CoT.
Not quite. Generic SOP, no identifiers. Green.
Scenario 7
"You are a marketing coordinator. Walk me through step by step how to allocate our $[XX,XXX] monthly marketing budget across paid search, ILS listings, and local social. Assume our target is [XX] move-ins per month and current cost per lease is roughly $[XXX]."
Green. Budget figures, targets, and cost-per-lease are all bracketed. No property named, no vendor named, no financial account exposed. This is a strategic reasoning task done safely.
Not quite. Everything is bracketed. Green.
Scenario 8
"You are an accountant. Here is our full August rent roll for Sunset Ridge (200 apartments, resident names, unit numbers, balances attached). Walk through step by step which accounts are highest priority for collection outreach."
Red. A whole property's rent roll with names, apartments, and balances is one of the most sensitive documents in the operation. CoT does not change that. Ask for the ranking framework as a general reasoning exercise with bracketed placeholders, then apply it to the real rent roll inside your PMS or an approved enterprise workspace.
A full property rent roll with names, apartments, and balances is the definition of Tier 3. CoT does not rescue this. Red.
Take Action · Anonymous Quiz
Five questions. Anonymous. Just for the pattern.
Nothing here identifies you. Your answers help us see where Week 15 landed and where to reinforce next Tuesday. Tap one answer per question, then submit at the bottom.
Question 1 of 5
Chain of Thought is:
Correct answer: B. Chain of Thought is not a mode or a setting. It is one added sentence at the end of any prompt telling AI to show its reasoning before it concludes.
Example of use
You paste your normal RTF prompt for "which vendor bid is a better deal" and add one line: "Think step by step. State your assumptions before you recommend." That is Chain of Thought. Same AI, same tool, nothing turned on in settings.
Question 2 of 5
Which task is the BEST fit for CoT?
Correct answer: C. Vendor bid comparison has real math, real tradeoffs, and real money on the line. That is exactly the moment CoT earns its keep. A, B, and D are writing tasks or short factual answers where the reasoning would just slow you down.
Example of use
CoT walks you through: Bid A $28,800 base + $1,600 waxing add-on = $30,400 vs Bid B $31,800 waxing included. Difference is $1,400. Now you can defend the decision to your regional. A bare prompt would have said "Pick Bid A" based on sticker price only.
Question 3 of 5
Where does the Chain of Thought instruction go in a stacked RTF prompt?
Correct answer: C. The order matters: Role, Task, Format, then Chain of Thought. RTF frames the task; CoT tells AI how to work through it before answering. Putting CoT earlier scrambles the structure.
Example of use
Role: You are a regional weighing a vendor decision. Task: Compare Bid A and Bid B. Recommend one. Format: One paragraph with the annual cost of each. Then, on a new line: "Think step by step. State your assumptions before you recommend." That last line is CoT. It rides on top of RTF.
Question 4 of 5
A CoT prompt with real resident names and balances pasted in is:
Correct answer: C. Structure is structure. Data is data. Making AI show its reasoning changes how you read the answer, not what data belongs in a general AI tool. A, B, and D are all versions of the same wrong idea: that making the reasoning visible somehow protects the resident. It does not.
Example of use
Wrong: "Sarah Martinez in unit 214 owes $2,847. Walk through step by step whether we should offer a payment plan." Right: "Resident in apartment [Apartment] owes $[Balance]. Walk through step by step whether we should offer a payment plan." Same CoT quality. Real name and balance stay out.
Question 5 of 5
Why is asking for CoT BEFORE the answer more reliable than asking for reasoning AFTER?
Correct answer: B. Once AI has committed to a conclusion, asking for the reasoning tends to produce a justification for that conclusion, not the chain that actually would have produced it. Ask before, and the intermediate steps can change the final answer. Ask after, and they cannot.
Example of use
Before: "Compare these two bids. Think step by step, then recommend." AI walks the math and often lands somewhere different than a gut answer. After: AI recommends Bid A, then you ask "Why?" It gives you three reasons Bid A is best. All three are constructed to defend the answer it already gave, even if the math would have pointed to Bid B.
Take Action · Policy & Download
Take one thing back to your team.
The Chain of Thought Quick Reference is a one-page PDF for your monitor or your break room. Print it. Send it to your supervisor. Talk about CoT at your next team huddle. That is how a workshop becomes a habit.
This week's downloads
Chain of Thought Quick Reference. One page. Template on the left, worked example on the right. Six trigger phrases. When to reach for it, when to skip it. Bracket-first data-safety note at the bottom.
AI Vocabulary Cheat Sheet (39 terms). Every term this curriculum has introduced, in one place. Same list as Slide 16.
Give your supervisor a preview of what you learned this week. One paragraph, one link. Takes ten seconds. Opens the door to a real conversation at your next one-on-one.
Opens your email client with the message pre-filled. You add your supervisor's email address before sending. Nothing is sent for you.
Take Action · Session Complete
Good job. Session Complete.
What you walked away with
1. Chain of Thought in one line: ask AI to show its reasoning before it gives you the answer.
2. Six trigger phrases that work across every AI tool, including "State your assumptions before you recommend."
3. A working rule for when to reach for CoT (math, comparisons, multi-step decisions, "why" questions) and when to skip it (writing tasks, quick edits, short factual answers).
4. Two worked examples showing CoT flipping a vendor decision and confirming a concession decision with math you can defend.
5. The bright line, again: making AI show its reasoning does not change what data is allowed in. Brackets stay in the toolkit.
The one-paragraph board version
This week we introduced Chain of Thought, the second of the seven prompt structures we teach in Phase 3. CoT is one added instruction at the end of an RTF prompt telling AI to walk through its reasoning step by step before giving the final answer. On any decision involving math, comparisons, or judgment, CoT catches wrong-sounding-right answers that a bare prompt would have hidden inside a fluent, confident recommendation. The bright line held: making the reasoning visible does not change what data is allowed to enter a general AI tool. Bracketed placeholders remain the standard for sensitive slots.
Where to practice safely
CoT plus real portfolio data is exactly the moment where a governed workspace stops being optional. mPro Digital Edge is where multifamily teams run CoT prompts against real numbers with AI Council rules built into every tool. If your team is graduating from "brackets in a public AI tool" to "real data with the guardrails on," that is what mPro Digital Edge is for.
Next Tuesday · Week 16
Few-Shot: teach AI with examples.
This week you told AI to show its reasoning. Next week you show AI what "right" looks like. Few-Shot is when you paste in two or three examples of the answer you want, then ask for a new one that matches the pattern. It is the fastest way to get AI to match your voice, your format, and your team's standards without spelling out a fifty-word style guide.
