Skill v1.0.0
currentLLM-judged scan100/100name: idea-clarify description: Surfaces what the user actually wants when the repo is empty or the request is one paragraph of vibes. Use before scope-knife on empty repos or vague briefs; do not invoke after a plan exists. when_to_use: | Trigger when scope-knife refuses to run because the repo is empty, the idea is too vague, or the user gave a one-paragraph brief. Run as a pre-step to scope-knife. Do not invoke once scope-knife has produced a plan; switch to scope-knife directly.
version: 1.0 category: scoping tags: ['empty-repo', 'vague-brief', 'pre-scope'] dependencies: [] triggers: ['we have no idea', 'what are we building', 'empty repo', 'vague brief', 'scope is unclear', '需求不明确', '明确项目目标', '想法落地', '要做什么', '到底要做']
idea-clarify
When a team shows up with an empty repo and a sentence, the problem is not "we have no features". It's "we have no shared picture of what we're building." This skill forces that picture into the open.
Input contract
Required:
user_brief: the raw one-paragraph description from the useraudience: who the demo is for (judges, customers, internal team)
Optional:
time_remaining_minutes: integer >= 0 (default 36 * 60)prior_artifacts: list of any pre-existing notes, sketches, links
Execution
1. Surface the problem
Ask the user the four canonical questions:
| # | Question | Why it matters | |
|---|---|---|---|
| 1 | Who is the user and what are they trying to do? | defines the value prop | |
| 2 | What does success look like at the demo? | locks the demo_path later | |
| 3 | What's the smallest thing we can show that proves the idea works? | prevents over-building | |
| 4 | What's the one thing we cannot ship without? | forces a hard priority |
Refuse to proceed until each answer is at least one concrete sentence.
2. Detect contradictions
If two answers contradict (e.g. "we want ML" + "no ML allowed"), surface the contradiction. Do not silently pick one. The agent must resolve it before scope-knife runs.
3. Emit a one-page brief
Write .hackathon/artifacts/idea-brief.md with:
- the four answers verbatim
- any contradictions + how the agent resolved them
- a one-line
demo_goalcandidate - a one-line
mvp_axis(the smallest axis along which the MVP can grow)
Output contract
Files written:
.hackathon/artifacts/idea-brief.md(human-readable, one page)
This skill does NOT write to .hackathon/state/*.json directly. It is a pre-scope-knife step; scope-knife reads the brief and writes plan.json from it.
Acceptance criteria
- [ ] The four canonical questions are asked exactly once.
- [ ] Contradictions are surfaced, not silently resolved.
- [ ]
demo_goalis a single sentence (max 20 words). - [ ]
mvp_axisis a single noun phrase. - [ ]
.hackathon/artifacts/idea-brief.mdexists and is <= 1 page.
Failure modes
| Mode | Behavior | |
|---|---|---|
| User gives one-word answer | Ask again, refuse to guess | |
| User refuses to answer | Surface as contradiction; refuse to proceed | |
| Contradiction detected | Block until resolved | |
| Time < 1h | Skip step 1 questions; ask 3 + 4 only |
Trigger phrases (for agent intent matching)
- "we have an empty repo"
- "what should we build"
- "the brief is one paragraph"
- "we have an idea but no plan"
- "idea is too vague"