Lesson objective
Write task constraints that keep a draft inside the business's prices and policies, do every calculation in a tool that gives the same answer every time, and refine a draft without letting it drift from its sources.
Plain-language explanation
Why this matters now
A request with good sources can still produce a bad draft. The tool may add a friendly promise, round a figure, extend a warranty by a month, or calculate a discount slightly wrong. Then the owner asks for changes, and after four rounds of "make it warmer" and "shorter please", nobody is sure the figures are still the ones in the price list.
Constraints and disciplined refinement are how the business keeps a draft inside its own facts from first version to last.
For a small business, the "action" is usually a sent quote, a posted price, a refund promise or an invoice. The constraint and the outside calculation stand between the wrong answer and the action.
A question to sit with
A fictional gift shop asks an assistant to apply "15% off everything over $40" to a list of twelve products and write the sale sign. The sign looks perfect and shows its working. Where should the shop's sale prices actually be calculated, and why?
What goes wrong
Constraints that are too vague. "Be accurate" and "don't make things up" are hopes, not constraints. The tool cannot check itself against them, and neither can the checker.
No constraint on promises. Friendly drafting adds "we guarantee", "no problem at all" and "we'll make it right" — each one a commitment.
The tool as calculator of record. Totals, discounts, tax lines and splits come from the draft. A total with working shown underneath was produced the same way as the total.
Rounding and unit drift. Hours become days, square feet become square yards, $1,250 becomes "about $1,200".
Refinement drift. Each round of edits is checked against the previous draft, not against the sources. By round four a price has changed and nobody saw it happen.
Arguing the tool into agreement. The owner says a figure is wrong; the tool agrees and produces another figure, which is also wrong but now feels corrected.
Long chats. Early material scrolls far back, and the draft begins to rely on what was said in passing rather than on the source blocks.
The practical answer
Write constraints the checker can test. A good constraint is specific enough that a person can look at the draft and say yes or no.
| Vague | Testable constraint |
|---|---|
| Be accurate | Use only prices in Source A; list every price with its line |
| Don't overpromise | Do not use the words guarantee, promise, always, never or free unless they appear in Source B |
| Keep it short | No more than 120 words |
| Be careful with dates | Use only dates in Source C; write dates as day and month |
| Don't give legal advice | Make no statement about what the law requires or what the customer is entitled to |
| No made-up claims | Make no claim about quality, speed or results that is not in Source D |
A tool may still break a constraint. Constraints earn their place because they make breaches easy to spot, not because they prevent them.
Claims are a legal matter, whatever the size of the business. The constraint against unsupported claims is there because of the federal ban on deceptive practices that Module 01 described, a duty with no size gate. Module 07 builds the substantiation file; this lesson just keeps unsupported claims out of the draft.
Do the arithmetic outside the tool. Every figure a customer will rely on is calculated in something that gives the same answer every time: a spreadsheet formula, a calculator, the pricing function of your point-of-sale or invoicing system. The draft only carries the figure in.
| Step | Where |
|---|---|
| Work out prices, discounts, totals, splits, tax lines | Spreadsheet, calculator or business system |
| Paste the calculated figures into a source block | The request |
| Draft the wording around those figures | The tool |
| Check each figure in the draft against the calculated figures | Outside the tool |
If the tool is used to calculate at all, treat its figures as a second opinion to compare, never as the answer.
Refine in a disciplined way.
- Change one thing per round. "Make it warmer" in one round; "shorten to 100 words" in the next. Mixed requests produce mixed changes.
- Check every round against the sources, not the last draft. Or, at minimum, check the final version fully against the sources before approval.
- Restate the constraints when you ask for a change. "Make it warmer, keeping every price and date exactly as in the sources."
- Do not argue about figures. If a figure is wrong, fix it yourself from the spreadsheet, or start a fresh request with the corrected source block.
- Start fresh when a chat gets long. Paste the layout and sources again in a new conversation rather than relying on material many screens back.
- Keep the final request with the final draft. The use record should show what produced the version that was sent.
Why "the tool agreed with me" is not a check. Professional guidance (NIST AI 600-1, July 2024; voluntary): Section 2.2 says models "sometimes provide logical steps for how they arrived at an answer even when the answer itself is incorrect." A corrected figure from the tool is a new output, not a confirmed one.
Where the constraint set stops. Constraints keep the draft inside the business's facts. They do not make a pricing decision, set a refund policy or answer a legal question. The owner decides prices and policies; an attorney answers what the law requires.
Remember this
- Write constraints a person can test by looking at the draft.
- Constraints make breaches easy to spot; they do not prevent them.
- Every figure is calculated outside the tool and carried into the draft.
- Refine one change at a time, restate the constraints, and check the final version against the sources.
- Never argue a tool into a figure; fix it from the calculation or start fresh.
Small business example
Lantern & Leaf is a fictional gift shop with two staff, owned by Ines Moreau. For an autumn sale she wants "15% off every item priced over $40" on eight products and a sign for the window. Her assistant, Callum, puts the eight prices in a spreadsheet with a formula. Three items qualify: a $48 candle becomes $40.80, a $64 throw becomes $54.40, and a $120 lamp becomes $102.00. The other five are $40 or under and keep their prices; a $40 vase stays at $40 because the offer is for items over $40. Callum checks one line by hand: $64 x 0.85 = $54.40. He pastes the eight final prices as Source A, with the offer wording from Ines's sale note as Source B, and uses the standing constraint set, including "use no price that is not in Source A" and "do not use the word guarantee". The first draft is good but lists the vase at $34.00. Callum does not argue with the tool: he corrects the figure from the spreadsheet. Ines asks for a warmer tone. Callum asks for that one change and restates "keep every price exactly as in Source A". The warmer version adds "satisfaction guaranteed". Callum removes it, because Ines's returns policy promises nothing of the kind. He checks all eight prices on the final version against the spreadsheet, not against the earlier draft. Ines reads the sign against the spreadsheet and the sale note and approves it. Callum saves the final request, both sources and the two corrections in the use record. Ines keeps the spreadsheet as the record of the sale prices, so any till question is answered from it, not from the sign.
Responsible-use warning
Do not let the tool be the calculator of record, and do not check a revised draft only against the draft before it. Figures drift quietly across rounds of editing, and a tool that agrees your figure was wrong will produce a new figure with the same confidence as the old one. Calculate outside the tool, carry the figures in, restate the constraints with every change request, and check the final version against the sources before anyone approves it.
Human-review requirement
The owner sets prices, discounts and policies, and approves every customer-facing draft. The owner checks that each figure in the final version matches the calculation made outside the tool, that no word in the draft promises more than the written policy, that the standing constraints were applied and every breach removed, and that the final version was checked against the sources rather than an earlier draft. The owner confirms the request that produced the sent version is stored with it. Any statement about legal entitlement goes to an attorney, and any tax line to the bookkeeper or signing tax return preparer.
Practical exercise
Use a fictional business and fictional prices only.
- Write a fictional list of eight products or services with prices, and a fictional offer such as a percentage discount above a price threshold.
- Calculate every offer price in a spreadsheet or on a calculator and record the results. Check one line by hand.
- Note any item exactly at the threshold and decide from the offer wording whether it qualifies.
- Write six testable constraints for a sign or email announcing the offer.
- Write a fictional first draft that breaks two of your constraints and contains one wrong price.
- Find each breach and the wrong price by checking against your calculation, and record what you found.
- Write three change requests for refining the draft, one change each, restating the constraints in each.
- Write a fictional revised draft that introduces a new promise, and remove it.
- Check the final version against your calculation and sources, not against the earlier draft.
- Write which parts of the offer are the owner's decision and which, if any, would need a bookkeeper or attorney.
- Decide where the final calculation will be kept as the record of the offer prices, and who can answer a customer's price question from it.
- Record the constraints, calculation, breaches and final check in your workbook.
Knowledge check
-
What makes a constraint testable?
Answer guidance: A person can look at the draft and say yes or no, for example "use only prices in Source A" instead of "be accurate".
-
Where should a business calculate the prices and totals a customer will rely on?
Answer guidance: In something that gives the same answer every time, such as a spreadsheet, calculator or invoicing system, with the tool only carrying the figures in.
-
What should you do when a tool agrees that its figure was wrong and gives a new one?
Answer guidance: Treat the new figure as a new unchecked output; fix it yourself from the calculation or start a fresh request with a corrected source.
-
What did the OWASP 2026 project leads say about Misinformation?
Answer guidance: Voters ranked it low but incidents ranked it high, and when fluent, confident output drives a decision or tool call, a wrong answer becomes a wrong action.
Sources and further reading
- OWASP Foundation, OWASP GenAI Security Project — OWASP Top 10 for LLM Applications 2026 (2026-08-04). Accessed 2026-09-15.
- OWASP Foundation — OWASP Top 10 for LLM Applications 2025 (2024-11-18). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.
- OWASP Top 10 for LLM Applications 2026, OWASP GenAI Security Project, 2026