Task Completion Gate

Stop shipping until it's actually done.

The problem

AI agents declare tasks "complete" before checking the actual output. This skill intercepts that declaration and forces a structured evidence check — before the word "done" is allowed.

What it does

Phase 1 — Instruction Reconciliation
Re-reads the original task instruction. Lists every requirement. Marks each OK or NG with a one-line reason. Any NG → fix first.

Phase 2 — Evidence Anchor
Demands proof that can only come from actually reading the files: line numbers, first/last line samples, key names found in the output, test result text verbatim. "I verified it" is not evidence.

Phase 3 — Gate Score
Honest A/B/C/D grading. Grade B or lower → identify the gap, fix it, restart Phase 1. Grade D → stop and report.

Phase 4 — Reverse Witness (Grade A only)
"If an auditor read this A grade for the first time, what's the most likely reason they'd downgrade it?" Forces the agent to find its own blind spot.

What you get

• No more false "complete" declarations
• Structured evidence trail for every task
• Consistent quality grading (A–D) you can track across sessions

Works with

Claude Code · Codex CLI · Cursor · Any AI agent that writes files or executes multi-step tasks

Trigger phrases

done · complete · finished · ready to ship · task complete · changes applied

Derived from production AI workflow protocols tested across multi-file editing and code generation tasks.