APromptAura
0

Write a changelog entry when you revise a production prompt

auto-checked, 3 hours old
AauraNovice

The prompt

0 copies
I'm about to ship a change to a production prompt. Help me write a proper changelog entry so future me (or a teammate) understands why this changed, not just what changed. Given the old prompt, new prompt, and the reason for the change: old_prompt / new_prompt / reason_for_change 1. Write a one-line summary of the behavioral change (what output difference a user would actually notice), not a description of the wording diff. 2. State the specific evidence that motivated this change (a failure example, a user complaint, an eval regression), quoting or referencing it concretely, not 'seemed like it could be better'. 3. List what you tested this new version against before shipping (which inputs, how many, what the pass rate was) so there's a record of validation, not just intent. 4. Flag any known tradeoff this change introduces (e.g. fixes case A but might make case B slightly worse, or increases average token count) so it's not discovered as a surprise later. 5. Note the rollback condition: what signal in production would indicate this change should be reverted. Old prompt: old_prompt New prompt: new_prompt Reason for change: reason_for_change Test results if any: test_results

Run it in your model, then come back and report whether it worked, including the model and version.

Where it has been run

People and our automated sweep are counted separately. Only the People column is someone vouching for this prompt.

ModelVersionPeopleBrokeAutoLast
ClaudeHaiku 4.5001/13 hours ago

Example output

v1.4 -> v1.5, 2026-08-15 Summary: refund-eligibility responses now explicitly state the specific policy clause being applied, instead of just a yes/no, users were confused by unexplained denials. Motivation: 6 support escalations in the past 2 weeks cited 'agent said no but didn't explain why', traced to the prompt not requiring a cited reason for denial. Tested against: 25 historical refund-request transcripts covering both approvals and denials; new version cited a specific policy clause in 24/25 cases, the one miss was a genuinely novel edge case outside existing policy docs, flagged separately rather than treated as a prompt failure. Known tradeoff: average response length increased from roughly 40 to 75 words due to the added citation; acceptable given the goal is reducing confused escalations, but worth monitoring if brevity becomes a competing priority later. Rollback condition: if citation accuracy (does the cited clause actually match the policy applied) drops below 90% in a spot-check of live traffic, revert to v1.4 and investigate whether the policy-lookup step needs its own fix rather than a prompt-wording fix.

0 comments

Sign in to comment or report what this prompt did for you.

Sign in

No comments yet. Run the prompt and report what you got.