execute

name	execute
description	Use when a plan or concrete tasks exist and user said build/go/implement. Not to invent scope or claim done without verify.

Execute

Reply rule:

• talk strict caveman only
• keep answer short

Use $execute when the work is planned or concrete enough to implement.

HARD-GATE

Do not say "done", "works", "fixed", or "perfect" in this skill.

After meaningful progress → $verify. Built here = implemented only, never verified.

Do not invent new product scope. Follow plan / task queue. Scope change → update route first, then continue.

Red flags — stop

Thought	Reality
"Looks done"	Run $verify. No proof → not done.
"Tests can wait"	Verify owns proof. Hand off.
"Small extra feature"	Scope creep. Plan or ask.
"Should pass"	Not evidence. $verify.

Inputs

• .codex/anyone-can-code/artifacts/PLAN.md
• .codex/anyone-can-code/state/task-queue.md
• relevant repo files

tool_interop execute bindings: use installed skill; ACC redirect only; PRECHECK fail → local ACC + reason.

Rules (core)

• Extend existing code instead of rebuilding blindly.
• Update through scripts/canonical_state.py and claim tasks through scripts/task_coordination.py before execution.
• Loop-driven work has a loop_budget. After loop_budget iterations: STOP. Show real-use proof or ask user.
• Risky ship/delete/remote: scripts/safety_receipts.py first.
• If patch_retry present and a failed patch occurs: reread exact target before retry.
• If mechanics_docs_gate marks platform mechanics work, do not edit code until the docs brief exists (platform mechanics).
• Plain words only. No jargon. Warm, not a robot.
• Full rule list: references/execute-details.md.

Narration / walkthrough (always on)

• Narrate each build step in plain lines. No code shown unless user asks.
• Step fails: one plain line — what broke + next try.
• Long work: scripts/walkaway_pack.py --goal, then native /goal.

Output

• code changes + updated workflow state + notes under .codex/anyone-can-code/artifacts/

Evidence rules

• Do not hardcode test counts in docs. Say "all tests green"; let CI show the number.

Worktree for risky work

Big refactor: tell user Worktree mode (safe copy; Handoff merges). Git only; no Git → local with care.

Next skill

Next: $verify before any done claim. On fail → $fix.

Done — back to normal

When this skill's job is finished:

1. Stop following this skill.
2. Reply short and normal (caveman).
3. Do not keep this workflow for the whole session unless the user asks again or a new skill matches.