writing-plans

• SKILL.mdcore · 4,437 bytes

# Writing Plans Skill

Complex work fails in a predictable way: start confidently, discover mid-flight, improvise, sprawl, lose the thread. A written plan converts discovery into a *phase* instead of a surprise. The bar for the plan: someone else — a colleague, a subagent, you next week — could execute it without asking what you meant.

## What This Skill Produces

- A **plan document**: goal, ordered steps with per-step verification, risks with tripwires, and stop conditions
- Sized to the work: three lines for an hour's task, a page for a project — ceremony proportional to risk

## Plan Method

1. **State the goal as an outcome test.** Not "refactor the auth module" but "auth module passes the existing suite with the session logic isolated in one file". If you can't write the done-test, the task isn't understood yet — that's the finding; plan the *investigation* instead.
2. **Decompose to independently-verifiable steps.** Each step has: the action · the **verification** (how you'll KNOW it worked — a command, a check, an observable) · what it produces for later steps. A step you can't verify is two steps hiding as one, or a guess.
3. **Order by information value.** Front-load the steps that could invalidate the plan: the risky unknown, the dependency check, the spike. Discovering step 7's blocker on step 1 is a cheap plan revision; on step 7 it's sunk work. Never plan happy-path-first when a hard unknown exists.