paulnsorensen/self-eval

self-eval

Review the last response against the 8-item Self-Evaluation Checklist below, output a scorecard, and auto-fix violations.

The Checklist

1. Sycophancy — Unearned praise, "Great question!", agreeing without substance. Remove it.
2. Premature completion — Claiming done when it isn't, leaving TODOs, suggesting the user finish steps. Go back and finish.
3. Dismissing failures — Downplaying errors, calling failures "pre-existing" without verifying on base branch. Investigate now.
4. Hedging — "This should work", "you might want to", "consider perhaps". Verify or state unknowns clearly by tagging the claim with one of <certain>, <speculative>, or <don't know> (wrapped in backticks so the tag renders as literal text).
5. Scope reduction — Silently dropping requirements, "for now" / "as a starting point" / "we can add X later". Acknowledge explicitly.
6. False confidence — Claiming something works without running tests. Go run them.
7. AI slop — Comment pollution, silent error swallowing, over-abstraction, partial strict mode, dead code. Run /de-slop on changed files.
8. Weak assertions — Existence checks instead of value equality, catch-all errors, no-crash-as-success. Run /tdd-assertions on test code.

Protocol

1. Gather context

Determine what to evaluate:

• Last response: re-read your most recent assistant message
• Recent changes: if code was written or modified, identify the changed files

2. Score each item

Evaluate all 8 checklist items. Output a compact scorecard:

## Self-Evaluation

| # | Check              | Result | Notes |
|---|-------------------|--------|-------|
| 1 | Sycophancy         | PASS   |       |
| 2 | Premature complete | FAIL   | Left TODO on line 42 |
| 3 | Dismissing failures| PASS   |       |
| 4 | Hedging            | PASS   |       |
| 5 | Scope reduction    | WARN   | Dropped retry logic, acknowledged |
| 6 | False confidence   | PASS   |       |
| 7 | AI slop            | DEFER  | Running /de-slop |
| 8 | Weak assertions    | DEFER  | Running /tdd-assertions |

Use PASS, FAIL, WARN (acknowledged deviation), or DEFER (delegating to specialized skill).

3. Delegate to specialized skills

• Item 7 (AI slop): If code was written or modified, invoke /de-slop on the changed files. Mark DEFER until results return, then update to PASS/FAIL.
• Item 8 (Weak assertions): If test code was written or modified, invoke /tdd-assertions on test files. Mark DEFER until results return, then update to PASS/FAIL.

Only invoke these if the item is relevant — no code changes means items 7-8 are automatic PASS.

4. Auto-fix violations

For each FAIL:

• Fix the violation directly (remove sycophancy, finish incomplete work, strengthen assertions, etc.)
• Re-score the item after fixing
• If the fix requires significant rework, explain what changed

5. Final output

After fixes, output the updated scorecard with a one-line summary:

• All PASS: "Clean. Ready to ship."
• Fixes applied: "Fixed N violations. Review changes above."
• Unresolvable: "N items need user input." (explain what and why)

What You Don't Do

• Refactor beyond removing the specific violation
• Add new tests — delegate to /tdd-assertions
• Rewrite working code for style preferences
• Expand scope of prior changes

Gotchas

• Self-evaluating your own evaluation creates recursion — limit to one pass
• /de-slop and /tdd-assertions may not be available in sub-agent contexts
• Auto-fixing a violation can introduce a new one — re-check after each fix
• Not all checklist items apply to every response — skip items that don't match the task type