# AI audit instructions and prompts

Free companion to the AI audit guide. Teaching rows are fictional and must be replaced. No row is a JAC or client result.

## Use the workbook

Import audit-workbook.csv into a spreadsheet. Use record_type values fact, question, observation, and repair. Give each record a stable ID. Connect questions, observations, and repairs using question_id. Keep fact references in fact_source and evidence_reference.

1. Enter approved public facts and their sources. Leave an unverified capability unknown.
2. Select a branded and an unbranded question relevant to a buying decision. Keep a fixed panel_version.
3. Record each actual attempt with time, platform, surface, locale, and settings. Save the answer in an approved evidence location.
4. Use attempt_status completed, failed, or not run. A failed attempt has no assessed mention, citation, recommendation, or accuracy value.
5. Classify outcomes only after reading the completed answer. Inspect a cited page before claiming it contains or omits a fact.
6. Define one proposed repair with its permission, scope, owner, and acceptance check. Use repair_status proposed, blocked, implemented, verified, or unresolved.
7. Save a changed_at date only after the change occurred. Repeat comparable checks and store their evidence separately.

## Collection prompt

Copy the prompt and supply permitted material.

```text
Investigate this product-description question using only the permitted sources and the answer observations supplied or actually collected.

Approved product facts and source dates:
[PASTE FACT RECORDS]
Question panel and version:
[PASTE QUESTIONS]
Permitted platforms, surfaces, and settings:
[PASTE]

For every attempted check, record the exact question, time, platform, surface, locale, settings, attempt status, error if any, and evidence reference.
For completed answers, separate brand mention, owned citation, recommendation, and accuracy against the fact records.
For unavailable or failed checks, leave outcome fields unassessed. Do not infer absence.
Do not quote a source you could not fetch. Do not combine surfaces or panels into an unexplained score. Do not change robots, account settings, or public content.
Return the records and unknowns, not a sales forecast.
```

## Repair and review prompt

```text
Review this audit record against the approved facts and original evidence.
Records: [PASTE]
Facts: [PASTE]
Evidence: [PASTE]

Identify one observed issue that a controlled change could address. Distinguish an incorrect capability claim, absent mention, technical obstacle, and insufficient evidence.
Propose a repair with evidence, fact owner, change owner, included work, exclusions, required access, and an observable acceptance check.
Keep external search adoption separate from implementing the repair.
If evidence or access is missing, name the dependency and leave the item unresolved.
Use Oxford commas. Do not invent company facts, results, or quotes.
```

## Teaching record

In the fictional example, the approved fact is CSV import with manual owner assignment. A hypothetical completed answer describing universal automatic CRM sync would be a claim to check against that fact. The example doesn't include an actual collected answer. Collect your own evidence before setting a real accuracy status.

The proposed page repair passes only when the public explanation matches verified facts, the URL and links work, and the agreed owner has checked it. A later assistant answer may still differ; record that observation without treating it as a guarantee or a validated score.

## Before sharing

Remove personal information and private source content. Share only approved facts and permitted evidence references. A completed workbook can support your own next page change or a one-time AI Audit conversation.