Launch audit

Visitors from a launch decide in seconds. This audit finds what would make them leave, what could go wrong once the project is visible, and ranks the fixes by impact. The rubric lives in rubric.md; read it first.

Steps

Read .megaphone/decisions.md first if it exists, and record any new answers there — or, with no human available, the assumptions you made.

1. Get the score. Run the bundled script (Python 3.8+, no dependencies):

   python3 <this-skill-dir>/scripts/audit.py OWNER/REPO            # what's on GitHub now
   python3 <this-skill-dir>/scripts/audit.py OWNER/REPO --readme README.md   # score a local draft

   It's the same code as the Megaphone web checker, so the numbers match. A token in GITHUB_TOKEN is optional. If the repo isn't on GitHub yet or the network is unavailable, score the README-based checks by hand from rubric.md and mark the repo-based ones unknown. If the repo ships a SKILL.md or .claude-plugin/, the report ends with an unscored Skill & plugin section: SK1–SK3 checked automatically, SK4–SK5 left to you in step 2.

2. Do the human review. Read the README top to bottom as a stranger arriving from the channel you'll launch on first — from .megaphone/decisions.md, or the most likely one if it isn't decided yet (a Hacker News reader on a laptop, a 小红书 user on a phone, a V2EX regular…). Answer:

   • Within the first screen (about the first 25 rendered lines), can I tell what it is, who it's for, and why it's better than what I use now?
   • Is there something to see before I have to read?
   • Can I try it in under a minute? Count the steps to the first result. Does that first result actually show the headline feature?
   • Is every hard requirement (browser, GPU, Docker, OS, account, paid key) stated up front?
   • Does anything look abandoned, broken or risky (stale badges, dead links, "WIP" with no date, no license)?
   • Would I know what to say if I told a friend about it?
   • Does the page I land on speak my language? GitHub opens README.md; a visitor from V2EX, 掘金 or 小红书 who meets an English first screen may leave before finding the small 「简体中文」 link (and the reverse for an English channel; a heuristic from our own launch). One line in the visitor's language under the tagline, linking to the twin, helps.
   • Is the fastest way to try it the easiest thing to click (for example, the hero image links to the hosted demo)?
   • Is there proof that someone else got value (a user's words, a before/after)? Optional at launch, but it's what turns a visit into a star later.

   Don't run the project's code unless the user asked you to; reason from the code and docs, and say so. If the repo has a SKILL.md or a .claude-plugin/ folder, also go through rubric.md § Skill and plugin repos (SK1–SK5): open each SKILL.md and read its frontmatter the way an agent would, deciding from name and description alone whether to use the skill.

3. Decide the blockers from rubric.md § Blockers. Any open blocker caps the level at 🔉 Needs polish regardless of the score. A blocker you can't settle without running code (usually Broken try-path) is unverified: check what you can without running it first, say what's left, and don't let it cap the level — it still counts as not closed for the launch gate (launch-plan).
4. Answer risk & readiness (rubric.md § Risk & readiness) — always, even for harmless-looking projects; "none found" is a valid answer. Quote risky wording from the repo and propose factual replacements. Don't give legal advice.
5. Rank the fixes. Label each Blocker, High or Medium by visitor impact, and show the rubric points it's worth (often +0 for the most important ones — that's fine). Name the Megaphone skill that fixes it when there is one.
6. Write .megaphone/audit.md (template below, headings in the user's language, check IDs kept), then summarise in chat: blockers, level, the top three fixes.

Levels

Set the level from the human-adjusted score when there is one, otherwise the rubric score, out of 100 and before any trade-off rescaling (rubric.md). Any open blocker caps the level at 🔉 Needs polish; an unverified one doesn't.

Template for .megaphone/audit.md

# Launch audit — <repo> — <date>

**Level: <emoji> <level>** · rubric score NN/100 · human-adjusted NN/100 (<why, if different>)
Open blockers: N · unverified: N

## Blockers
| Blocker | Status | Evidence | Fix |
|---|---|---|---|
| Quick start doesn't show the point | open/closed/unverified | … | … |

## Risk & readiness
- Third parties: …
- Claims in the repo: "<quote>" → "<factual replacement>"
- Other people's data: …
- Disclaimer: present / draft: "…"
- Capacity: …
- Reachability: …

## Skill & plugin (only for repos with a SKILL.md or .claude-plugin/)
- [x] SK1 Valid frontmatter — <evidence> (a check that doesn't apply: `- n/a SKn … — <why>`)
- [ ] SK2 Description says when — "<current description>" → "<proposed one with trigger phrases>"
- [ ] SK3 Copy-paste install — …
- [ ] SK4 First thing to say — …
- [ ] SK5 Findable — topics: …; directories/lists and when they apply: …

## Fix first
1. <fix> — Blocker/High/Medium — why it matters, exactly what to change (+N pts)
2. …

## Rubric detail
| Category | Score |
|---|---|
| First impression | NN/35 |
| Time to wow | NN/25 |
| Trust | NN/20 |
| Shareability | NN/20 |

- [x] A1 Description … (8/8) — <evidence>
- [ ] A5 Visual in the first screen (0/8) — <evidence>

## Human review
- First screen: …
- Try-it path: N steps — …
- Requirements: …
- Word of mouth: "<how a user would describe it in one breath>"

## Accepted trade-offs
- D1 Two languages — the author keeps a Chinese-only README on purpose (scored NN/95 → NN/100)

## Manual checks (can't be detected automatically)
- [ ] Social preview image set (Settings → General → Social preview)
- [ ] Repo pinned on the author's profile
- [ ] Demo GIF under ~10 MB and loops cleanly

Notes

• Don't pad. A project nobody can try is not "almost there" because it has a license and badges — that's what blockers are for.
• If the user disagrees with a check (e.g. no second language), record it as an accepted trade-off rather than nagging.