verification-before-completionlisted

# Verification before completion Claims of success cost nothing. Evidence costs a single command. Always pay it. ## When to use - About to say "done", "fixed", "passing", "ready to merge". - About to mark a task `DONE` in a `Status:` block. - About to close a `Plan:` checkbox. ## Delegation Two things this skill does **not** delegate, and one it must not do. Native `/verify` builds and runs the app — *"without falling back to tests or type checks"* — and since v2.1.215 Claude cannot invoke it on its own. It is a good manual complement, not a substrate this flow can call. `/goal` is the closer relative: a session-scoped prompt-based Stop hook whose evaluator re-checks a completion condition after every turn. It enforces *"keep going until X holds"*. This skill enforces the different thing — *"do not claim X without pasting the command and its output"*. Use both; they compose. **Do not instruct the model to verify.** Current models already do. Anthropic's Opus 5 guidance is explicit that prompts containing *"include a final verification step"* or *"use a subagent to verify"* cause over-verification, and that *"the same applies to legacy harness scaffolding that adds separate verification steps."* What survives is the **evidence format** below — a completion claim carries the command and its output, because that is what a human reader needs, not because the model needs reminding. ## The gate Before claiming, run through these in order: 1. **Identify** — what command p