Context
The final mission teaches the loop: when the learner corrects the agent, the correction becomes a better example, rubric, skill, or harness check.
Do This Now
- Choose one bad AI rewrite from the batch workflow.
- Write a correction in plain language.
- Ask Codex to turn the correction into a durable update.
- Rerun the workflow on the old failure case.
Where to paste this:
Paste this into Codex after you have at least one batch report or dashboard decision.
Paste this into Codex after you have at least one batch report or dashboard decision.
Prompt To Use
Turn my correction into a durable workflow improvement.
Here is a listing output I did not like:
[paste listing id, original, AI rewrite, and your correction]
Inspect:
- skills/real-estate-listing-improver/SKILL.md
- rubrics/listing-quality-rubric.md
- examples/good-listing.md
- examples/bad-listing.md
- reports/listing-regression-cases.md if it exists
Return:
1. failure type,
2. why it matters,
3. whether this should become a gold example, rubric update, skill update, or regression case,
4. the exact proposed change,
5. a rerun prompt to test whether the mistake is fixed.
Do not apply changes until I approve.
What You Should See
Expected result
Codex should classify the failure and propose a small durable update. The correction should not stay trapped in the chat.
If It Goes Wrong
- If it proposes a giant skill rewrite, ask for the smallest change that prevents this exact failure.
- If it blames missing data, ask what verification warning should have been shown.
- If it cannot test the fix, ask for a regression case.
Make It Yours
Use the same loop for seller emails, market updates, property brochures, neighborhood guides, social captions, and buyer-match summaries.
Practical Output
A self-improving workflow where mistakes become reusable standards.
Your Working Notes
What I tried
What I need to change
Codex Task Card
| Files to inspect | skill, rubric, examples, reports. |
|---|---|
| Expected files created | A proposed durable update and rerun prompt. |
| Verification checklist | The same failure is caught or avoided on the next run. |
| Follow-up if weak | Apply the approved change and rerun the old failure case. Report whether the mistake is fixed. |