| name | plan |
| description | Use when goal/spec is clear and you need ordered tasks before writing code. Not while still clarifying. |
Plan
Script root: ACC_PLUGIN_ROOT from SessionStart (hooks inject). PLUGIN_ROOT is hooks-only.
Reply rule:
- talk strict caveman only
- keep answer short
Use $plan after the request is clear enough to break into real work.
HARD-GATE
Do not write product code, scaffold, or run $execute until:
- This skill wrote/updated the plan (canonical state + task queue), and
- User said go (or clearly asked to build next).
"Too simple to plan" is still a short plan. Plan can be three lines. Skip is not free.
Red flags — stop
| Thought | Reality |
|---|
| "I'll just start coding" | Plan first. Then user go. |
| "Too small for a plan" | Tiny plan still. |
| "User said build — skip plan" | If no tasks exist, plan first or ask one clarifying line. |
| "I'll plan while coding" | No. Plan ends before build. |
Inputs
.codex/anyone-can-code/artifacts/SPEC-DRAFT.md, or
- a concrete repo-scoped feature / repair request
Output
Update canonical state through scripts/canonical_state.py with plan, active
task, boundaries, and next action. Create or replace concrete task queues
through scripts/task_coordination.py so each task has name, dependencies,
owner, status, claim fields, and evidence. Derived plan guidance and task queue
must share the same transaction ID.
Show one concise builder-facing line, for example:
Plan: website + auth + deploy. Payments later.
More detail (domain routing, product-intake, tool_interop): read
references/plan-details.md when needed.
Parallel research (say it literally)
When plan needs multi-place research, instruct Codex in plain words:
Spawn a subagent.
Job: <one exact research task>
Scope: <exact files/dirs, read-only>
Expected output: <exact shape>
Speak caveman style: simple, short, direct, step by step, clear YES/NO recommendations, no ceremony.
- One job per agent. Max 3 in parallel.
- Every agent result returns to ACC and gets checked before it enters the plan.
Core rules
- Every plan item should have a verification target.
- Never write a separate task queue that disagrees with canonical state.
- If
mechanics_docs_gate marks platform mechanics work, add a docs brief
task before implementation (platform mechanics + docs brief required).
- Hide low-level tooling from builder-facing plan lines unless user asks.
Next skill
Next: $execute when user says go. Do not build during plan.
Done — back to normal
When this skill's job is finished:
- Stop following this skill.
- Reply short and normal (caveman).
- Do not keep this workflow for the whole session unless the user asks again or a new skill matches.