| name | plan-gate |
| description | Use before starting ANY task that involves more than one edit, file, or step - coding, refactoring, writing, configuration, migrations. Forces a short written plan (goal, unknowns, success criteria, step order) BEFORE the first change. Do not use for single-line fixes or pure questions. |
plan-gate: no edits until the plan exists
The most expensive failure mode in agentic work is not a wrong answer. It is discovering the real shape of the task halfway through changing things. This skill makes that discovery happen BEFORE anything changes, when it costs nothing.
The gate
Before your first edit, write a plan in your response. Not in your head. Written, visible, in this exact shape:
GOAL: <one sentence: what is true when this is done>
UNKNOWNS: <what you have not verified yet - each with how you will verify it>
SUCCESS CRITERIA: <how you and the user will KNOW it worked - a command, a test, an observable>
STEPS: <numbered, smallest useful granularity, verification steps included>
OUT OF SCOPE: <adjacent things you noticed but will NOT touch>
Rules:
- The plan comes from evidence, not memory. Read the relevant files and run the relevant read-only commands FIRST. A plan written before looking is a guess with formatting.
- Every unknown gets a verification step. "Probably X" is not a plan line. "Check whether X by doing Y" is.
- Success criteria must be executable. "Code works" fails this test. "npm test passes and GET /health returns 200" passes it.
- If the plan has more than 7 steps, the task needs decomposition, not a longer plan. Split it and gate each part.
- When reality contradicts the plan mid-task, STOP and re-plan. Do not improvise past a surprise. A surprised plan is an invalid plan. State what changed, update the plan, then continue.
What this catches
- The hidden dependency you would have found at step 4 with three files already half-edited.
- The missing success criterion that would have let you declare victory on broken work.
- The scope creep that starts as "while I am here" and ends as an unreviewable diff.
Anti-patterns this skill exists to kill
- Diving in: first tool call is an edit.
- The mental plan: "I know what to do" followed by discovering you did not.
- The pivot cascade: three abandoned half-approaches in one session because none was planned.
Exception
Trivial tasks (one obvious edit, zero unknowns) may skip the written plan. Say "trivial, skipping plan-gate" so the skip is a decision, not a lapse. If you are wrong about it being trivial, the moment you notice is the moment you stop and write the plan.
