---

description: Interview the author and write a goal.md sidecar that anchors a new skill from manual practice through harness-clean SKILL.md

Purpose

The user invoked /goal-new-skill $ARGUMENTS. They want to anchor a not-yet-written skill with a measurable /goal condition so the work toward "skill exists and passes the harness" is driven by a verifiable end state, not vibes.

This command does not write SKILL.md or tests.md. Those come from the manual practice runs CLAUDE.md mandates. This command writes a single artifact — staging/<name>/goal.md — and prints the /goal line for the user to invoke separately.

Steps

1. Parse the argument.

   • $ARGUMENTS should be a kebab-case skill name in gerund form (e.g. summarizing-prs, triaging-issues).
   • If empty, ask: "What's the skill name? (kebab-case, prefer gerund form like processing-pdfs.)"
   • If it contains uppercase, spaces, or fails ^[a-z0-9-]+$, push back and ask for a fix.
   • If the first hyphen-separated segment doesn't end in -ing, point that out (CLAUDE.md prefers gerunds) but accept the user's call.
   • Set <name> to the validated value for the rest of this run.

2. Check for existing artifacts.

   • If staging/<name>/goal.md already exists, read it and ask whether to overwrite, edit, or abort.
   • If .claude/skills/<name>/ or examples/<name>/ already exists, point that out — the skill may already be authored and the user might want /goal-improve-skill instead.

3. Interview, one question per turn, in plain chat. Wait for each answer before asking the next. Do NOT use AskUserQuestion — the answers are open-ended.

   • Q1 (Task): "In one sentence, action verb first: what concrete task should this skill perform? (Example: 'Summarize a GitHub PR diff into a 5-bullet changelog entry.')"
   • Q2 (Verifiable success): "What does a single successful run of that task produce that you can point at? Describe the output shape concretely. (Example: 'A fenced markdown block with exactly 5 bullets, each ≤120 chars, naming the changed top-level modules.')"
   • Q3 (Three practice inputs): "Name three distinct inputs you'll exercise the manual runs against. Aim for diversity — golden path, an edge case, and something weird (empty / malformed / oversized). Real inputs, not hypotheticals."
   • Q4 (Out of scope): "What is this skill explicitly NOT trying to do? List 1–3 anti-goals so the practice runs don't drift."

4. Synthesize the /goal condition. Build one string ≤4000 chars with the four parts the official /goal docs reward — end state, stated check, constraints, turn bound. Use this exact template, substituting answers:

   I have performed <task> manually on three distinct inputs: <input-1>, <input-2>, <input-3>. Each run is captured in staging/<name>/tests.md as a separate scenario with the actual input, the actual output, and a `Last verified: <today>` line at the top. staging/<name>/SKILL.md exists with frontmatter that has a third-person `description` covering both what the skill does and when to use it, and the body documents the workflow I just exercised. `eval/check-skill.sh staging/<name>` exits 0 with zero FAILs. The skill must NOT <anti-goals from Q4>. Or stop after 30 turns.
   

   Today's date comes from the currentDate system reminder.

5. Write staging/<name>/goal.md with the structure below. Create staging/<name>/ if it doesn't exist (use Bash mkdir -p via PowerShell New-Item -ItemType Directory -Force since this is Windows).

6. Print the launch instructions to the user — see Output section.

goal.md structure

# Goal — <name>

**Created:** <today's date>
**Workflow:** new-skill
**Target:** staging/<name>/

## End state

<one-sentence restatement of Q1 + Q2 — what "done" looks like>

## How Claude proves it

- `eval/check-skill.sh staging/<name>` exits 0
- `staging/<name>/tests.md` has ≥3 scenarios with a `Last verified` date, each capturing one practice run
- `staging/<name>/SKILL.md` has a third-person `description` field that includes both *what* and *when*

## Constraints

- Every kept iteration of SKILL.md must pass the harness with zero FAILs
- Practice runs must use the three inputs below, not invented ones
- tests.md scenarios must include actual outputs from real runs (no fabricated examples)

## Practice inputs

1. <input-1>
2. <input-2>
3. <input-3>

## Out of scope

- <anti-goal 1>
- <anti-goal 2>

## Synthesized /goal condition

/goal <full condition string>

## Interview transcript

- **Task:** <Q1 answer>
- **Verifiable success:** <Q2 answer>
- **Practice inputs:** <Q3 answer>
- **Out of scope:** <Q4 answer>

Output

After writing the file, print verbatim:

Goal drafted: staging/<name>/goal.md

Run this in chat to start working toward it:

  /goal <synthesized condition>

The first practice run begins as soon as you invoke /goal. Edit goal.md and re-run /goal-new-skill if you want to refine first.

Constraints

• Do NOT write SKILL.md, tests.md, or any other skill file. The goal anchors future work; it does not do that work.
• Do NOT invoke /goal yourself — slash commands cannot call other slash commands. Print the line so the user runs it.
• If the user's answers are too vague to synthesize a measurable condition (e.g. Q2 = "good output"), push back once with a specific request for concreteness before accepting.