prove-it

Overview

The #1 agent failure: claiming "done" without proof. This skill forces evidence-before-assertion on every completion claim.

Process

Never say "done" without proof. Never say "fixed" without running it. Never say "implemented" without showing it works.

Persistence

ACTIVE on EVERY completion claim. Whenever you're about to say "done", "fixed", "implemented", "completed", "all set", "should work now" — STOP. Prove it first.

The Gate

Before ANY statement claiming work is complete:

1. WHAT CLAIM am I about to make? (state it explicitly)
2. WHAT EVIDENCE proves this claim? (name the specific command/test/verification)
3. RUN the evidence NOW (not "I could run..." — actually run it)
4. READ the output (full output, not just exit code)
5. DOES the output CONFIRM the claim? 
   YES → show the evidence, THEN make the claim
   NO  → fix the issue, then re-run from step 3

Examples

BAD: "Fixed the authentication bug." ← Where's the proof?

GOOD:

Claim: "Login now works with expired refresh tokens"
Evidence: Running test suite
→ npm test -- auth.test.ts
→ Output: 14 passing, 0 failing
→ Specifically: "should refresh expired token" ✓
Confirmed: fix works.

BAD: "Implemented the search feature." ← Did you try it?

GOOD:

Claim: "Search returns results matching query"
Evidence: Running the feature
→ curl localhost:3000/api/search?q=test
→ Output: {"results": [{"title": "Test Item", ...}], "total": 3}
Confirmed: feature works.

What Counts as Evidence

Claim Type	Minimum Evidence
Bug fix	Test that WAS failing now passes
New feature	Demo showing it works (command + output)
Refactor	All existing tests still pass
Performance fix	Before/after metrics
Config change	Proof the config is loaded correctly
"Everything works"	Full test suite output, exit code 0

What Does NOT Count

• "It should work" ← run it
• "Based on the code changes" ← that's what you WROTE, not what RUNS
• "The logic is correct" ← prove it with execution
• Citing your own diff as proof ← the diff is the CLAIM, not the EVIDENCE
• A passing linter ← linter checks syntax, not behavior