Skill · pattern · plan-and-act
Plan / act
A deny-writes planning session that produces a revisable document; act is a permission change against that plan. Use for long-horizon coding or research that a human must agree to before side effects, or when a 40-step plan is generated once and never rewritten.
GET /api/canon/skills/plan-and-act?format=md
Codex/Claude Code plan mode is a permission change. Updating the plan is a tool call. A static todo list while the agent improvises is a named failure.
When
Side effects need a human-agreed shape. Plan is a file, not a thought.
Do
- 01
Plan with writes denied
The artifact is a markdown plan the human can edit. No patches yet.
- 02
Act is a gate change
Allow writes against the plan. Verify in the loop.
- 03
Revise the plan as a tool
When the world disagrees, update the document. Do not improvise around a stale 40-step list.
Don't
- Generate a plan once at t=0 and never touch it.
- Treat plan mode as a system-prompt personality.
Hard rules
- Plan is a permission set plus a file.
- Human edit is HITL on that file.
Refuse
- Plan generated once — A 40-step plan written at t=0 and never rewritten when the world disagrees.
- No verifier in the loop — The agent writes code and declares victory. Tests, typecheck, and build are optional.
Load with this
Load next
Trigger tests
Should fire
- “Add a plan mode before edits”
- “The agent never updates its todo list”
Should not
- “A one-shot classify-and-reply”