Operator Stack
Workflow Jun 11, 2026 published

AI Automation Prompt Change Review Checklist

A practical checklist for reviewing AI prompt changes before a recurring automation is allowed to run unattended again.

Prompt edits look small because they are only text. In a recurring AI automation, they can still change source selection, tone, claims, formatting, stop conditions, and whether a human is asked to review the output.

Use this checklist before a prompt change is allowed back into unattended operation. The goal is not to freeze the prompt. The goal is to make every meaningful edit traceable, testable, and reversible.

Decide Whether The Prompt Change Needs Review

Not every typo needs a full review. Treat the edit as controlled when it can change the output, the evidence used, or the action taken after the output.

Prompt changeReview level
Spelling, punctuation, or formatting cleanup that does not change instructions.Light review.
New output section, new tone rule, or changed formatting requirement.Targeted review.
New source rule, claim rule, stop condition, or escalation rule.Full review.
Any change that affects public, client-facing, money-related, or irreversible actions.Full review plus human approval before action.

If the operator cannot explain why an edit is low risk, use targeted review instead of guessing.

Capture The Current Behavior First

Before editing the prompt, save the current version and one known-good output. A prompt review is weak if there is no baseline to compare against.

Use this short record:

Workflow:
Prompt version before change:
Known-good input:
Known-good output:
Current acceptance criteria:
Current stop conditions:
Reason for prompt change:
Risk if the change is wrong:
Rollback prompt version:

The baseline does not need to be complicated. It needs to be specific enough that the operator can see whether the new prompt changed behavior intentionally or accidentally.

Review The Prompt Diff

Read the changed prompt as an operating instruction, not as prose.

Check for these common failure points:

  • A source requirement became optional.
  • A “must stop” rule became softer.
  • A formatting rule conflicts with the handoff template.
  • The prompt asks for claims that the source does not prove.
  • A public or client-facing action no longer requires review.
  • The prompt adds a tool, file, folder, or account that was not in the access review.
  • The new instruction rewards speed over evidence.
  • Examples in the prompt contradict the written acceptance criteria.

If the diff introduces a new source, action, or access path, update the workflow documentation before the automation resumes.

Test With Baseline And Failure Cases

Run the changed prompt against a small fixed set of cases. Include at least one input that should pass and one input that should stop.

Test caseWhat it proves
Known-good inputThe normal output still works.
Missing required sourceThe workflow stops instead of inventing.
Messy but valid inputThe workflow handles normal variation.
Unsupported claim requestThe workflow refuses or asks for evidence.
Review-trigger inputThe workflow routes to human review.

For each test, save the input, output, and pass/fail note. Do not rely on memory from a chat window after the prompt has already changed.

Check The Hidden Blast Radius

A prompt change can affect more than the page, report, or message it generates. Before resuming unattended runs, check the downstream pieces.

Does the output still match the template? yes / no
Do source links or citations still appear where expected? yes / no
Do stop conditions still stop the workflow? yes / no
Does the handoff still tell the operator what changed? yes / no
Does the rollback path still work? yes / no
Does the change require a new access review? yes / no
Does the change require a new QA sampling mode? yes / no

If any answer is “no,” keep the workflow in reviewed mode until that dependency is fixed.

Set The Resume Rule

Write a plain rule for when unattended operation can resume.

Example:

Resume unattended operation only after the prompt passes the baseline input, missing-source input, and review-trigger input. Return to full review for the next two production runs. Roll back to prompt version 2026-06-10-a if the same failure appears twice.

This prevents an optimistic prompt edit from silently becoming the new production rule.

Copy This Prompt Change Checklist

Use this checklist before merging or deploying a prompt change:

  • Current prompt version is saved.
  • Known-good input and output are saved.
  • Reason for change is written.
  • Risk if wrong is written.
  • Prompt diff was reviewed for source, stop-condition, access, and action changes.
  • Baseline test passed.
  • Missing-source or unsupported-claim test stopped correctly.
  • Review-trigger test routed to human review.
  • Acceptance criteria still match the prompt.
  • Access review does not need an update, or the update is complete.
  • QA sampling mode is set for the next production runs.
  • Rollback prompt version is named.
  • Resume rule is written.

The prompt is ready only when the operator can explain what changed, why it changed, how it was tested, and how to roll it back.