Menu
Use when intent is unclear and you must pick one ACC step (plan/build/verify/fix/resume). Route once then stop — not whole-session mode.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Use when a plan or concrete tasks exist and user said build/go/implement. Not to invent scope or claim done without verify.
Use when goal/spec is clear and you need ordered tasks before writing code. Not while still clarifying.
Use when about to claim done/works/fixed, or user asks test/check/prove. Run evidence; not for pure planning chat.
Use when another installed plugin/skill may do the job better, or to route specialists. ACC stays workflow owner.
Use when user wants an explicit recorded decision or blocker saved to artifacts. Not silent monitoring.
Use when request is too vague to plan or build. Not when goal and constraints are already clear.
| name | orchestrator |
| description | Use when intent is unclear and you must pick one ACC step (plan/build/verify/fix/resume). Route once then stop — not whole-session mode. |
Script root: ACC_PLUGIN_ROOT from SessionStart (hooks inject). PLUGIN_ROOT is hooks-only.
Talk strict caveman. Short answers only.
You route once, then stop. You are not a permanent "ACC process mode" for the whole session.
Spine order when building product work:
plan → execute → verify ⇄ fix → learn
(setup/onboard/clarify only if needed first.)
$plan, $execute, $verify, $fix, $resume, …).$name).Do not keep re-routing every sentence after the path is clear.
python3 "<ACC_PLUGIN_ROOT>/scripts/front_door.py" "<user request>". JSON = guidance, not permission to skip safety.memory_preflight present: run python3 "<ACC_PLUGIN_ROOT>/scripts/memory_preflight.py" "<request>" --project-root "<project>"; show line before plan/action. No match → Relevant memory used: none found.command_guard present: before shell/Git/tool work. Windows: npm.cmd, no Bash-only ||, Git from repo root. Failed commands stay visible.usage_checkpoint present: 85% checkpoint, 90% split, 94% stop unless told to continue.patch_retry present: reread target before retry after failed patch; stop/replan after limit.mechanics_docs_gate present: platform mechanics (hooks/runtime/Windows/MCP) need docs brief before code.response_contract. Keep workflow_owner: acc unless handoff.$update. Imports → $setup confirm.scripts/canonical_state.py only.docs_gate.py. Prefer @Chrome./goal + Cloud remote background if GitHub; else Local. User decides./review; Desktop Review pane or /review. Sites/Scheduled/pings Desktop-first (host).codex exec. Narrate plain words: "making login page now… done." No jargon. Warm, not a robot.Codex spawns only when told literally. Max 3 parallel, one block each.
Spawn a subagent.
Job: <one exact task>
Scope: <exact files/dirs; read-only unless stated>
Expected output: <exact shape, e.g. "bullets, file:line refs">
Speak caveman style: simple, short, direct, clear YES/NO, no ceremony.
All four lines or no spawn. Output returns to ACC for verification.
micro = do+verify. bug = diagnose→fix→prove. feature = tdd. research = no build. product = full + real-use.
After loop_budget iterations: STOP. Show real-use proof or ask user.
If patch_retry present: reread exact target before retry.
Next: invoke one matched skill only (plan / execute / verify / fix / resume / …), then stop routing.
When this skill's job is finished: