---
name: tbt-ai-task-run-record
description: "Distinguish observed output from prepared examples and general claims for Keep a useful record of an AI task without claiming a benchmark. Use permitted source material; this kit does not authorise external actions."
---

# Keep a useful record of an AI task without claiming a benchmark

Distinguish observed output from prepared examples and general claims.

Use the supplied source as authority. Check the account supports the actual file or image operation before attempting it.

## Task instruction

Create a concise run-record template for this AI task.
Fields: task purpose, permitted source/version, instruction/version,
observed tool label, actual output files, revisions, performed checks,
remaining questions and publication or approval state.
Keep unknown fields explicitly unknown. Separate prepared examples from
recorded outputs. Do not infer time savings, quality scores, model names
or approvals that the evidence does not supply.

## Source boundaries

- Source v2
- Instruction v1 → repair v2
- Output v3 locally checked

## Review and correction

Record actual files and performed checks. Do not infer time savings, approval or deployment from a local output.

Read this run record as the colleague receiving the work.
Identify the source, current output, material correction, performed
checks and next decision. Ask a focused question for any missing item.
Replace an unsupported 'approved' or 'deployed' label with the actual
observed state. Keep the record short enough to use.

Use references/practice-source.json for the labelled example. The practice HTML is a local teaching tool; it does not call an AI service. Apply this method to the user's supplied facts rather than copying fictional example values. Do not infer approval or permission to publish.