← ClaudeAtlas

autolisted

Explore to ship in one run. Calls the built-ins in order and lets the gates hold each step honest.
IsthisLee/did-you-check · ★ 0 · AI & Automation · score 73
Install: claude install-skill IsthisLee/did-you-check
# End to end Take $ARGUMENTS from start to finish. **Reply in whatever language I am writing to you in.** Follow the four stages from the official best practices: Explore → Plan → Implement → Commit. **Do not reinvent a stage — call what already exists.** ## 0. Size it first If you can describe the diff in one sentence, skip planning. The official docs: "If you could describe the diff in one sentence, skip the plan." Wrap ceremony around small work and people stop using the command. For anything bigger, walk the steps below. ## 1. Explore Hand it to the **built-in Explore subagent**. It runs in its own context and returns only the conclusion, so this conversation does not fill up with file contents. The official docs: "Use subagents to keep research out of it." ## 2. Plan Enter **plan mode** (`Shift+Tab`) and write the plan. No code changes before approval. If the plan turns out to be a large feature, suggest settling a spec first with `/check:spec`. ## 3. Implement Follow the `/check:tdd` loop: failing test first, confirm RED, minimum implementation, confirm GREEN. All four gates are live through this stage. Assert without evidence and the evidence gate stops you; try to finish without running the checks and the completion gate does; try to neuter a test and the integrity gate does; try to rewrite history and the project guard does. **If a gate blocks you, assume the gate is right and go measure.** ## 4. Review Call the **built-in `/code-review`**. A fresh sub