Context
A harness lets the learner stop trusting vibes. It checks whether listing outputs meet a quality bar and catches repeated mistakes.
Do This Now
- Read the listing rubric.
- Create a gold set with good, bad, and borderline listing examples.
- Ask Codex to evaluate 10 listings against the rubric.
- Turn any repeated failure into a regression case.
Where to paste this:
Paste this into Codex inside the starter workspace.
Paste this into Codex inside the starter workspace.
Prompt To Use
Create a simple listing quality harness.
Inspect:
- rubrics/listing-quality-rubric.md
- examples/good-listing.md
- examples/bad-listing.md
- reports/batch-10-listings.md if it exists
Create:
- reports/listing-quality-harness.md
- examples/borderline-listing.md
- reports/listing-regression-cases.md
The harness should check:
1. no invented facts,
2. target buyer is clear,
3. strongest feature is concrete,
4. generic adjectives are reduced,
5. verification warnings are present,
6. rewrite is different enough from the original,
7. human approval is required before publishing.
Evaluate 10 sample listings and name the repeated failure patterns.What You Should See
Expected result
Codex should create a harness document and identify repeated failure types. It should not claim the workflow is perfect.
If It Goes Wrong
- If it only gives scores, ask for specific evidence and failure types.
- If it approves everything, ask it to apply an adversarial reviewer lens.
- If the rubric is too vague, ask for measurable checks and examples.
Make It Yours
For your market, add checks for legal restrictions, protected-class wording, investment claims, school claims, fees, taxes, and energy ratings.
Practical Output
A listing quality harness, a borderline example, and first regression cases.
Your Working Notes
What I tried
What I need to change
Codex Task Card
| Files to inspect | rubrics, examples, reports. |
|---|---|
| Expected files created | reports/listing-quality-harness.md, examples/borderline-listing.md, reports/listing-regression-cases.md. |
| Verification checklist | Repeated failures become named cases, not vague advice. |
| Follow-up if weak | Update the listing improver skill so it explicitly avoids the top two regression failures. |