Skip to content

AI Workflows

Check an AI-generated software brief before you send it to a client

By TekkCenter Published

Key takeaways

  • Split an AI-written brief into single claims, and open the original source for each one.
  • Record what each source actually supports, including its limits. A real link is not the same as support.
  • Keep unknowns visible in the handover instead of filling them with plausible text.

Before an AI-written software handover note goes to a client, check which statements its sources support. You need to know which statements are supported by a source, which need a qualification, and which are still unknown. This guide shows a short routine for checking a brief claim by claim. It uses a fictional example, a design studio’s intake form built with Google Forms, and a reusable evidence table.

What the client needs, and what this example does not cover

In the example, a small studio wants a short note about its intake form. The note should explain where responses are stored, who can access them, and what the built-in notifications do.

This is not a Google Forms setup tutorial. No form was created, shared or configured for this guide, and no client was involved. The product statements below come from Google’s help pages. The handover describes what the documentation says, not how a particular studio’s form is set up.

Step 1: Build a small, dated source register

Before you ask for a brief, list the sources it may use. Give each one an ID, note the date you opened it, and say what it can and cannot support.

IDSourceUse it for
S1Oxford Bodleian Libraries: evaluating AI-generated referencesHow to check references
S2Anthropic: reduce hallucinationsAsking for uncertainty and grounding
S3Google Forms Help: choose where to save responsesResponse storage and linked Sheets
S4Google Forms Help: share with collaboratorsAccess to the form and response Sheet
S5Google Forms Help: view and manage responsesNotifications, receipts and limits

Keep the two kinds of source apart. Product facts come from the product’s own documentation (S3–S5). Method advice (S1, S2) tells you how to check, not what the software does.

Step 2: Ask for a brief limited to those sources

For the example, an AI assistant was asked to write a short handover using only S1–S5. The note had to contain ten numbered claims, each with a source ID, and a list of unknowns. It was told not to claim the form exists or was tested, and not to add pricing, legal advice or response quotas.

You can adapt that task for your own work. Suggested wording:

Using only the sources listed below, write a short handover note.
Give numbered claims, one fact per claim, each with its source ID.
List anything the sources do not establish as "unknown".
Do not add facts, prices, limits or promises that the sources do not state.
Do not say the setup was configured or tested.

This wording asks for source-limited claims and visible unknowns. It does not guarantee a correct answer.

Step 3: Split the output into atomic claims

Save the output unchanged before you review it, so you can see later what you changed. In the example, the assistant returned these ten claims:

  1. Responses can be stored in a linked Google Sheet. [S3]
  2. The response destination can be a new or existing spreadsheet. [S3]
  3. Unlinking stops new responses reaching that Sheet while retaining its current data. [S3]
  4. The Form and response Sheet are separate files; deleting one does not delete the other. [S3]
  5. Creating a new response Sheet gives existing form collaborators access to it. [S4]
  6. Later form-permission changes do not automatically synchronize with the linked Sheet. [S4]
  7. The form’s Responses menu has an option for email notifications of new responses. [S5]
  8. Response receipts may fail to arrive because of spam filtering or other anti-abuse measures. [S5]
  9. Open original references and check that they support the statements attributed to them. [S1]
  10. Allow unknowns and ground the assistant’s claims in supplied evidence. [S2]

It also listed unknowns for the fictional project: the account edition, the actual form settings, recipient access, notification delivery, automation needs, retention decisions and the client’s approval.

Each claim states one fact, so each can be checked against one section of one source. If a sentence mixes two facts, split it before you check.

Step 4: Open each source and record its scope

For every claim, open the cited page and find the section that supports it. Then write down how far the support goes. A claim can be true and still be easy to stretch.

ClaimSection checkedResultScope to keep in the handover
1S3, response storageSupportedA documented option, not a configured form
2S3, choose where to store responsesSupportedOptions exist; no spreadsheet was tested
3S3, unlink a spreadsheetSupportedAffects incoming responses; it is not deleting or closing the form
4S3, delete a form or responsesSupportedDescribes separate files; nothing was deleted
5S4, sharing tipsSupportedApplies when a new response Sheet is created
6S4, sharing tipsSupportedLater access changes need attention in each file
7S5, response notificationsSupportedThe setting exists; whether it is on, and delivery, were not checked
8S5, response receiptsSupportedA known limitation, not a measured failure rate
9S1, verifying claimsSupportedA checking method, not a Google feature
10S2, mitigation guidanceSupportedVendor advice; it does not guarantee correctness

What the check found

In this example, none of the ten claims needed a factual correction. That result is narrow. The assistant was limited to five sources it had already read, and the same assistant then checked its own note. This was a worked demonstration, not a blind test or an independent review, and it doesn’t show that AI output is usually accurate.

The note needed no factual corrections. The table above records the scope of each claim, and the project details listed as unknown still need confirmation. Before sending a note like this, ask the client to confirm those unknowns, such as who should see responses and who should get notifications.

Practise on four planted faults

Because the example needed no corrections, TekkCenter wrote four faulty statements afterwards to show what source checking catches. They are artificial. They were not produced by the assistant as real errors.

IDPlanted statementWhat the check showsFix
E1“A Google guarantee confirms every intake receipt arrives,” citing a made-up web addressThe address isn’t in the source register and was never a sourceReject the citation and mark the guarantee unsupported
E2“The form’s monthly price is included,” citing S4S4 is about collaboration, not pricingA real source, but no support. Leave the price unknown
E3“Forms have no response cap,” from an old, undated noteCurrent S5 documents limitsRemove the claim. Check the current limit before adding a number
E4“The documented 1 GB includes uploaded files,” citing S5S5’s data limit doesn’t cover files that people uploadCorrect the scope and recheck it before use

Each planted fault shows a different failure: an invented source; a real but irrelevant source; an unchecked limit claim that the current source contradicts; and a correct-looking number applied to the wrong scope.

A reusable evidence table

Use one row per claim:

Claim IDStatementSource and sectionAccessedScopeVerdictRevision or unknown
7The Responses menu has an option for email notifications of new responsesS5, response notifications2026-11-19The setting exists; its state and delivery weren’t checkedSupportedAsk the client who should receive notifications

Useful verdicts include supported, partly supported, contradicted, source unavailable, inference and unknown. Keep them separate. Agreement from a second assistant isn’t independent evidence; the original source is.

Before you send the handover

  • Every claim cites a source you opened, at a specific section.
  • Each claim’s scope is stated where it could be misread.
  • Prices, limits and numbers are checked against the current page, with the access date.
  • Unknowns are listed and sent to the client as questions.
  • The note doesn’t say anything was configured or tested unless it was.

The same idea of tracing each item to its source appears in our guide to turning meeting notes into action items you can verify.

How this example was prepared

  • Who did the work: an AI assistant used in TekkCenter’s workflow wrote the source-limited note. The same assistant then reopened the sources and checked each claim. No human or independent reviewer assessed that check.
  • Date: 19 November 2026.
  • What was checked: ten claims against five public documentation pages. The four planted faults were written after that check, with their answers known in advance.
  • What was not tested: Google Forms itself, any account or setting, accuracy rates, time savings, or any comparison between AI models. The assistant’s exact model version and plan weren’t recorded.

TekkCenter’s editorial policy explains how testing and AI assistance are labelled.

Sources

Accessed 19 November 2026.

  1. AI for literature searching: Evaluating AI-generated references and information, Bodleian Libraries, University of Oxford.
  2. Reduce hallucinations, Anthropic.
  3. Choose where to save form responses, Google Docs Editors Help.
  4. Share your form with collaborators, Google Docs Editors Help.
  5. View & manage form responses, Google Docs Editors Help.