AI Prompts for Appraisers: Narratives, Neighbourhoods, Reconciliation

Prompts that keep judgement with you split each request into a facts block you have verified and a conclusion slot you fill yourself. The assistant turns facts into plain sentences for the market-conditions, neighbourhood, reconciliation and reviewer-reply sections, and leaves every conclusion as a marked placeholder. Below are five prompts, each with sample output and a judgement check.

Two habits make this safe. Use placeholders in place of client data, so names, addresses and loan details never reach the tool. And finish every draft with an audit prompt that lists any sentence your facts do not support. All five prompts share one illustrative assignment, so you can follow the same numbers from the market paragraph to the reviewer reply.

Follow me on Instagram@sagnikteaches

Three rules behind every prompt in the pack

The exact wording of a prompt matters less than the structure it enforces, and three rules do most of the work.

Connect on LinkedInSagnik Bhattacharya

Split facts from conclusions. The facts block holds only what you have verified: figures from your own market analysis, boundaries you drew, adjustments you calculated. Conclusions are slots the assistant must leave empty and label as yours. That structure is the whole answer to drafting without outsourcing judgement. The tool never gets the chance to decide the trend, the adjustment or the weight, because those words are missing from its material.

Subscribe on YouTube@sagnikteaches

Use placeholders in place of client data. Swap the client, borrower, subject address, loan number and the lender's instructions for labels such as [CLIENT], [SUBJECT] and [COMP 1], and put the real details back in your report software, not in the chat. The assistant needs the shape of your facts, not who they belong to.

Finish every draft with a check. Each prompt below is followed by the checks I run on its output, and prompt 5 turns the audit into a prompt of its own.

The split between the two jobs looks like this:

Keep the pack in a Project. ChatGPT Projects and Claude Projects both hold standing instructions and files, so each new assignment starts from the same three rules. The facts block is the only part you paste fresh. The assignment used in the examples is a three-bedroom detached house of 1,450 square feet, built in 1998, in an average condition, in a market area labelled [MARKET AREA]. Every figure is invented for illustration.

Prompt 1: market conditions from your own data

The market-conditions paragraph is where invented statistics do the most damage, so this prompt has the tightest rules. It takes the indicators you have already pulled from your data source and writes them up. It does not comment on the market. For a deeper treatment of this single section, see drafting market-conditions sections with ChatGPT.

You are drafting the market-conditions paragraph of an appraisal report.
Use ONLY the facts block. Do not add commentary, forecasts or statistics
from memory.

FACTS BLOCK ([MARKET AREA], 12 months to [DATE] against the prior 12 months):
- Closed sales: 168 vs 154
- Median sale price: $412,500 vs $398,000
- Median days on market: 29 vs 21
- Active listings at period end: 43 vs 28
- Months of supply: 3.1 vs 2.2
- Average sale-to-list price ratio: 98.4% vs 99.6%

Rules:
1. One paragraph of plain sentences. Every number exactly as given.
2. Describe direction with neutral verbs (rose, fell, eased). Do not use
   words that classify the market (strong, soft, balanced, seller's market)
   or predict it (will, likely to continue).
3. Give the percentage change for closed sales and median price, with the
   working on a separate line after the paragraph.
4. End with these two placeholders, exactly as written:
   [APPRAISER CONCLUSION: market trend for this property type, with reasoning]
   [APPRAISER CONCLUSION: whether a time adjustment is supported, with evidence]

Illustrative output:

In the 12 months to [DATE], 168 sales closed in [MARKET AREA], up from 154 in the prior 12 months (9.1%). The median sale price was $412,500, compared with $398,000 (3.6%). Median days on market rose from 21 to 29. Active listings at the end of the period stood at 43 against 28, which equates to 3.1 months of supply against 2.2. The average sale-to-list price ratio eased from 99.6% to 98.4%. [APPRAISER CONCLUSION: market trend for this property type, with reasoning] [APPRAISER CONCLUSION: whether a time adjustment is supported, with evidence]

Working: (168 - 154) / 154 = 9.1%. (412,500 - 398,000) / 398,000 = 3.6%.

The judgement checks for this prompt:

• Recompute the working yourself. Language models can slip on arithmetic. The prompt asks for the working so that you can verify two percentages in ten seconds with a calculator.
• Check the period and the area against your analysis. A paragraph that quotes the right numbers for the wrong 12 months is still wrong.
• Read for classifying words. "Rose" and "eased" describe. "Cooling" or "tightening" decide. The second kind belongs in your conclusion.
• Fill both slots in your own words, including the evidence: the paired sales or the monthly medians that support, or fail to support, an adjustment.

What I would fix: without rule 2, assistants tend to close this paragraph with a sentence such as "indicating a balanced market". That sentence is a conclusion, and it quietly pre-empts the trend slot, which is why classifying words are banned. If the assistant asks for a fact it was not given, such as the data source or the property type, supply it in the block rather than letting it assume.

Prompt 2: a neighbourhood that describes places, not people

Neighbourhood sections invite adjectives, and adjectives are where a draft starts describing who lives somewhere instead of what is there. This prompt narrows the assistant to land, buildings, services and access, and compares the subject with the area on two measurable points only. I leave schools, crime and demographics out of the facts block so there is nothing to repeat. If your client or standards require comment on amenities, supply neutral, checkable facts such as type and distance.

Draft the neighbourhood description for an appraisal report from the
facts block only.

FACTS BLOCK:
- Name: [MARKET AREA]
- Boundaries: north [BOUNDARY 1], east [BOUNDARY 2], south [BOUNDARY 3],
  west [BOUNDARY 4]
- Built-up: about 95%
- Land use: single-family detached about 85%, townhomes about 10%,
  neighbourhood retail about 5%, mostly along [ROAD B]
- Typical age of homes: built 1985 to 2005
- Typical lot size: 0.18 to 0.30 acre
- Utilities: public water, sewer, electricity and gas
- Access: primary shopping about a five-minute drive; bus route on [ROAD B]
- Subject: built 1998, lot 0.22 acre

Rules:
1. Describe land, buildings, services and access only.
2. Do not describe residents. Do not use desirable, popular, sought-after,
   family-friendly, quiet or similar words.
3. Do not mention schools, crime or demographics.
4. Compare the subject with the area on age and lot size only, using the
   numbers above.
5. End with: [APPRAISER CONCLUSION: effect of the [ROAD B] frontage, if any,
   with evidence]

Illustrative output:

[MARKET AREA] is a built-up suburban neighbourhood, about 95% developed, bounded by [BOUNDARY 1] to the north, [BOUNDARY 2] to the east, [BOUNDARY 3] to the south and [BOUNDARY 4] to the west. Land use is predominantly single-family detached (about 85%), with townhomes (about 10%) and neighbourhood retail (about 5%), the latter mostly along [ROAD B]. Most homes were built between 1985 and 2005 on lots of roughly 0.18 to 0.30 acre. The subject, built in 1998 on 0.22 acre, is typical of the area for age and lot size. Public water, sewer, electricity and gas are available. The nearest primary shopping is about a five-minute drive away, and a bus route runs along [ROAD B]. [APPRAISER CONCLUSION: effect of the [ROAD B] frontage, if any, with evidence]

The judgement checks:

• Every sentence maps to a line in the block. If you cannot point to the line, the sentence goes.
• Search for adjectives. Any word that rates the area rather than measures it is a sign the rule slipped.
• "Typical of the area" is a conclusion in disguise. It is safe here only because the facts block gave both ranges. If you did not supply the ranges, you should not let the phrase stand.
• The road slot is yours. Whether a busy road affects marketability depends on your paired sales or listing evidence, and the draft cannot see either.

A typical slip when rules 2 and 3 are missing: the draft calls the area "a family-friendly suburb popular with young professionals". Neither phrase is in the facts, and both describe people. Delete them, and tighten the prompt so the next draft cannot produce them. For the equivalent problem on the agent's side, neighbourhood pages that avoid steering shows the same discipline applied to marketing copy. Your professional standards will say more about what a neighbourhood description may and may not contain, so treat the rules above as a floor.

Prompt 3: reconciliation comments that carry only your reasoning

Reconciliation is the section where an appraiser explains how the approaches were weighed, so it is also the section where an assistant is most tempted to supply reasons of its own. The fix is to hand over your reasons as a numbered list and instruct the tool to use those reasons and no others. Its job is to turn five bullet points into a paragraph a reviewer can follow. For the wider picture on report narratives, see AI for real estate appraisals: report narratives.

Draft the reconciliation paragraph from my reasoning below. You are
organising my reasoning, not adding to it.

INDICATIONS OF VALUE:
- Sales comparison: Comp 1 $409,900 (net adjustment +2.1%), Comp 2 $418,500
  (+0.8%), Comp 3 $425,200 (-2.6%)
- Cost approach: $431,500
- Income approach: not developed

MY REASONING (use these reasons and no others):
1. The sales comparison approach reflects how buyers in this market compare
   similar homes: most weight.
2. Within it, Comp 2 gets the most weight: same street, closest in gross
   living area (1,440 vs 1,450 sq ft) and condition, smallest net adjustment.
3. Comp 3 needed the largest adjustment: it is larger and has a garage the
   subject lacks.
4. The cost approach supports the upper part of the range but gets limited
   weight: accrued depreciation is an estimate, and land value was extracted
   from three sales.
5. The income approach was not developed: the area is mostly owner-occupied
   and rental data are insufficient.

Rules:
- No reason, comparison or word of weight (strongly, clearly, heavily) that
  is not in my reasoning.
- Leave the final figure as [APPRAISER TO ENTER: opinion of value and
  effective date].
- After the paragraph, list every number you used and its source line.

Illustrative output:

The sales comparison approach is given the most weight because it reflects how buyers in this market compare similar homes. Three sales needed net adjustments of +2.1%, +0.8% and -2.6%, and indicate $409,900, $418,500 and $425,200. Comp 2 carries the most weight within the approach: it is on the same street as the subject, is closest in gross living area (1,440 against 1,450 square feet) and condition, and needed the smallest net adjustment. Comp 3 needed the largest adjustment because it is larger and has a garage the subject lacks. The cost approach indicates $431,500. It supports the upper part of the range but is given limited weight because accrued depreciation is an estimate and land value was extracted from three sales. The income approach was not developed because the area is mostly owner-occupied and rental data are insufficient. The opinion of value is [APPRAISER TO ENTER: opinion of value and effective date].

The judgement checks:

• Count the reasons. The paragraph should contain your five and nothing else. A sixth reason, however sensible, is a conclusion you did not make.
• Check the figures against your grid. Each figure should match your adjustment grid and cost approach exactly. The number list the prompt asks for makes this a one-minute job.
• Read the verbs of weight. "Most weight" and "limited weight" came from you. If the draft says the cost approach "strongly supports" the range, it has upgraded your reasoning, so replace the word.
• Type the opinion of value yourself. The final figure and its effective date should never be produced by the tool, even when it is obvious from the numbers above.

Say the three adjusted comparables average $417,867, and you conclude $418,000. That is your call: the average is arithmetic, the rounding towards Comp 2 is judgement, and only the second belongs in the opinion. Put the reasoning for the rounding into your reasoning list, because if it is not in the list the paragraph will not mention it, and a reviewer will ask.

Prompt 4: a reply to a reviewer's query

A reviewer's query usually arrives at a bad moment and asks for an explanation you already gave in the report, just not in the form they wanted. The risk with an assistant is tone and invention: it will apologise for things you did not do, or add a justification that sounds plausible. So the prompt pastes only the question (with names and file numbers removed), gives your reasons as a list, and bans promises. Drafting a reconsideration-of-value response covers the closely related case where a client challenges the value itself.

Draft a reply to the reviewer's query below. Use only my reasoning.
Tone: courteous, factual, not defensive. Do not promise a revision and
do not apologise.

REVIEWER'S QUERY (names and file numbers removed):
"Comp 2 closed nine months before the effective date yet carries the most
weight. Please explain why it is more reliable than the more recent sales,
and confirm that no time adjustment was needed."

MY REASONING:
1. Comp 2 is on the same street, closest in gross living area (1,440 vs
   1,450 sq ft) and condition, with the smallest net adjustment (+0.8%).
2. The two more recent sales needed larger net adjustments (+2.1% and
   -2.6%), which carry more adjustment risk.
3. Market-area median sale price: $409,500 in the month of Comp 2's sale,
   $412,500 in the effective month, a change of 0.7%.
4. In my judgement a change of that size is not supported as a separate
   time adjustment.
Close with: [APPRAISER TO CONFIRM: any further support to attach]

Illustrative output:

Thank you for the query. Comp 2 closed nine months before the effective date. It is given the most weight because it is on the same street as the subject, is the closest match in gross living area (1,440 and 1,450 square feet) and condition, and needed the smallest net adjustment (+0.8%). The two more recent sales needed larger net adjustments (+2.1% and -2.6%), which carry more adjustment risk. The market-area median sale price moved from $409,500 in the month of Comp 2's sale to $412,500 in the effective month, a change of 0.7%. In my judgement a change of that size is not supported as a separate time adjustment. [APPRAISER TO CONFIRM: any further support to attach]

The judgement checks:

• Answer the question that was asked. The query has two parts, a reliability question and a confirmation about time adjustment. Check that both are answered, in that order.
• Verify 0.7%. (412,500 - 409,500) / 409,500 is 0.73%. Rounding is fine, but you should have recalculated it, because it is the number a reviewer will test.
• Look for concessions you did not make. Phrases such as "on reflection" or "we will review" are commitments. Delete them unless you mean them.
• Attach what the slot asks for. If the reply says the median moved by 0.7%, the monthly medians should be in the file you send.

Treat the reply as a draft of your position, not of the reviewer's relationship with you. If the query contains something you disagree with, or something you now realise the report got wrong, the prompt cannot help: that is a decision about the report, and it is yours.

Prompt 5: the audit that catches what the first four let through

The audit is the prompt I would keep if I could only keep one. It reads the finished narrative against your facts and reasoning and lists every sentence that has no source. It does not rewrite anything, so there is nothing to approve except the list. Run it in a fresh chat, with both the facts and the draft pasted in, so it is not marking its own work in the same conversation.

Below are (A) my facts blocks and reasoning and (B) a narrative draft.
Audit B against A.
1. For every sentence in B, quote the line in A that supports it, or
   write NO SOURCE.
2. List any word in B that classifies, predicts or praises (for example
   strong, stable, popular, likely, desirable) and say whether A contains
   an appraiser conclusion that justifies it.
3. Recalculate every percentage and total in B from the numbers in A and
   say whether each matches.
4. Do not rewrite anything.

Illustrative output for a draft that has picked up two extra sentences:

Three of those five rows need action, and none of them would have been obvious on a quick read. The audit is not infallible either: it can mark a sentence as supported when the support is only approximate. Treat it as a second pair of eyes that is good at tedium, and keep your own read as the final one.

What stays out of the prompt and what to check in the settings

The placeholders only work if they are used every time, so keep a short list of what never gets pasted: client and borrower names, the subject address and any parcel or tax identifiers, loan and file numbers, the client's instructions and internal notes, interior photographs, and anything your engagement letter marks as confidential. If a facts block needs a real address to make sense, the block is probably carrying more than it should.

Then check the tool's side. On an individual plan, look for the model-training switch in the privacy settings and turn it off; labels change often, so read yours today rather than trusting a screenshot. The business plans (ChatGPT Business and Enterprise, Claude Team and Enterprise, Gemini in Workspace, Microsoft 365 Copilot) are stated by their vendors not to train on business content by default. Both ChatGPT Business and Claude Team need at least two seats, so a solo appraiser usually ends up on an individual plan with training switched off. Keep memory features off, or confined to one project, so that facts from one assignment cannot leak into the next. A temporary chat is not erasure either: ChatGPT's Temporary Chat is not used for training, for example, but may be kept for up to 30 days. What to check in an AI tool's privacy policy and terms covers the questions to put to any vendor.

The settings are the tool's side. The other side is your own obligations: the engagement letter, the client's policy on third-party tools and the professional standards you work under. Some clients restrict outside tools and some ask for disclosure of AI assistance. Read them before the first use. What appraisers should check first on AI guidance is a good starting list.

How a made-up growth figure nearly reached a report

Here is an illustration of the mistake the whole pack exists to prevent. An appraiser, short of time, asks a general chat assistant for "a market conditions paragraph for a three-bedroom detached house in [MARKET AREA]", with no facts block. The draft comes back fluent and confident: "Prices have appreciated about 4% a year over the past five years, with inventory remaining tight." Neither claim is based on the appraiser's data, because the assistant had none. It is describing a market in general, not this one.

It shows up when the audit prompt runs. Both clauses come back NO SOURCE. The appraiser's real figures say the median price rose 3.6% year on year, not 4% a year over five, and months of supply went up from 2.2 to 3.1, which points the opposite way from "tight". Had the paragraph gone out, a reviewer comparing it with the data would have found a contradiction. Fluent text is not evidence, which is why the facts block comes first and the audit comes last.

Minutes saved, and what goes in the file

These timings are illustrative, for one file that draws a reviewer query. Written by hand, the four narrative pieces might take 90 minutes: 25 for market conditions, 20 for the neighbourhood, 20 for reconciliation and 25 for the reply. With the pack, including the checks, the same pieces take about 48: 10, 8, 10 and 12, plus 8 for the audit of the whole narrative. On a file with no query the totals are 65 minutes by hand and 36 with the pack. The saving is real but modest, and the checks account for a good part of what is left, which is the right trade for a document you sign. If you spend the saved time on a longer look at the comparables, the pack has paid for itself in the way that matters.

Keep a record. Save the facts blocks, the filled prompts, the audit output and a note of the tool and date with the workfile, for as long as your standards and engagement terms require records to be kept. Version the pack too: when you add a rule, such as a new banned word, note the date, so you can tell which version of the prompts produced which report. If a client or standards body ever asks how a paragraph was produced, the file answers in a minute.

Appraiser questions about prompts, tools and disclosure

Can the assistant choose my reconciliation weights?

No. Weights are an opinion you must be able to defend, so they belong in your reasoning, not in the tool's. Give it your weights and your reasons, then ask it only to organise them into clear sentences and to list any number it used. If its wording implies more or less weight than you intended, change the wording.

Do I have to tell a client I used AI on their report?

That depends on your engagement terms, the client's own policy and the professional standards you work under, and I can't answer it for you. Read all three before the first assignment, ask the client in writing if the terms are silent, and keep a note of what you decided and when.

How do I stop the assistant inventing market statistics?

Give it a facts block, tell it to use only that block, and make 'NO SOURCE' the required answer for anything else. Then run the audit prompt on the finished draft. A fabricated figure usually shows up there as a sentence with no matching line in your facts.

Do these prompts work for commercial assignments?

The structure does: facts block, empty conclusion slots, audit. The content changes, because a commercial file leans on lease data and income analysis rather than a handful of sales. Build the facts block from your abstracts and rent schedules, and keep every capitalisation and rent conclusion as a slot that you fill.

Further reads

• USPAP Workfile Checklist: What to Keep and How AI Can File It — What belongs in the workfile when AI touched a draft.
• AI for Commercial Appraisers: Lease Abstracts and Income Prep — The same facts-first method for income-producing assignments.
• Can ChatGPT Write Home Inspection Report Narratives? What to Check — Another report writer's checks on AI-drafted narratives.
• Using AI for Appraisal Work: Comps Prep and Draft Report Checks — Preparing comparables and checking the draft report.
• How to Build a Shared Prompt Library for Your Team — How to keep a pack like this shared and versioned.
• How to Catch Made-Up Figures in AI-Drafted Proposals — The same idea for catching invented figures in proposals.
• AI-Ready Appraisal Software: What UAD 3.6 Means for Your Practice — The UAD 3.6 dates, the listed software with published prices, the AI features worth paying for and a two-report test for a solo appraiser.
• UAD 3.6 Checklist for Solo Appraisers: Templates, Photos and AI Tools — A grouped UAD 3.6 readiness checklist for a one-person appraisal practice, with the AI drafting steps that save time and the ones to leave alone.
• AI for Canadian Appraisers: CUSPAP, Privacy and Practical Uses — CUSPAP confidentiality, competence and work-file rules read against AI, plus privacy points and practical appraisal uses that keep client data out.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI, Anthropic, Google Workspace and Microsoft privacy pages (business-plan training defaults); OpenAI's Temporary Chat FAQ; ChatGPT Business and Claude pricing pages (seat minimums); OpenAI and Claude help pages on Projects.