Turns a request into a short written plan — problem, risks, phases, tasks — before any code is written.
4 stars
0 votes
0 copies
1 view
Added September 26, 2026
ai-agentsgo
Security analysis
A100/100
Scanned September 26, 2026
$npx -y skills add xiaolongnk/SuperHarness --skill plan --agent claude-code
Installs into .claude/skills of the current project.
Are you the author of Plan?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/xiaolongnk-plan)
---
name: plan
description: Turns a request into a short written plan — problem, risks, phases, tasks — before any code is written.
domains: [planning]
tier: both
---
# Plan skill
## When to use
Invoke for any **multi-step** task: a feature, a non-trivial bug fix, a refactor, anything where
the steps aren't obvious in one line. Plan first, then execute against the plan.
Skip for genuinely trivial one-liners (a typo, a one-line config change) where writing a plan costs
more than the change. For a big or risky feature, get the plan reviewed before implementing.
## Steps
1. **Restate the request in your own words.** One or two sentences. If your restatement and the
request disagree, you've found a misunderstanding — resolve it before planning further.
2. **State the problem.** What are we actually solving, and for whom? Name the user-visible outcome
that means "done."
3. **List the risks and unknowns.** What could go wrong, what's uncertain, what might break that's
adjacent. Flag anything you'd want a second opinion on. An honest risk list now saves a rewrite
later.
4. **Break the work into ordered phases.** Each phase is independently checkable — you can stop
after it and have something coherent. Order them so the riskiest or most uncertain phase comes
early, where a wrong assumption is cheap to correct.
5. **List the concrete tasks inside each phase.** Specific enough to execute without re-deciding:
which file, which function, what behavior. Note where tests drive the work (write the failing
test first).
6. **Define done.** The observable outcome and the check that proves it (a test, a command, a
behavior in the running app) — this becomes the acceptance bar for `/verify`.
7. **Write the plan down** (a short doc or the task board) and, for a big feature, hand it to a
fresh-context reviewer before any code. Record the agreed plan so the implementer executes it
rather than re-improvising.