launch-check Skill
为 AI 生成的应用和网站(Lovable、Bolt、Replit、Cursor、v0、Claude)提供上线前体检报告。一条命令被动检查线上 URL,按红绿灯给 4 个领域打 0-100 分——安全(页面 JavaScript 中可见的密钥、Supabase service_role 密钥、通过 --owner-check 检查暴露的 Supabase 或 Firebase 表、缺失的安全响应头、HTTPS、cookie)、375px 手机端设计、Google 和 AI 搜索可发现性、速度与基础项——然后生成一份可直接粘贴给 AI 的通俗易懂的"优先修复"清单和一页 HTML 报告。当有人问"我的应用上线安全吗"之类的问题时使用。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
Launch Check
scripts/launch_check.py is a deterministic checker (Python standard library; Playwright optional). It loads the
site the way a visitor's browser does, scans what it downloaded, runs two vendored audits, and prints one
report card. Do not judge the site by reading it yourself: run the tool, then explain and fix what it found.
Run it
- Find Python 3.9+:
python3 --version, elsepython --version, elsepy -3 --version(Windows). Call it. = the folder that holds this SKILL.md."/scripts/launch_check.py" https://the-app.com --report launch-report.html
--owner-check: ONLY when the user confirms they own the site. If the page contains a Supabase URL + public
anon key (or a Firebase databaseURL), it asks the API which tables exist and reads at most 1 row per table
(GET only, max 20 tables, never writes, row contents are never printed or saved; only "table X: readable by
anyone (N columns)"). Without the flag it only says that a database was found.
--repo .: also scan the local project for committed.envfiles and secrets in tracked files.--prompt: print only the fix list as a prompt for the AI that built the app.--json: everything.--skip phone,search,speed: faster partial run.--fail-under 80: exit 1 below this score.
- Phone design and page-load metrics need Playwright (` -m pip install playwright && -m playwright
install chromium`). Without it, phone design is skipped with a note; the rest still runs.
Typical time: 6-8 s on a small site, 30-60 s on a large marketing site (the search check reads 5 pages politely).
Read the output
SCORE 30/100 NOT READY: 5 things to fix before you launch
FAIL Security 0 Anyone can see your Stripe secret key in your website's code (+7 more)
WARN Phone design 77 The page is wider than a phone screen (+4 more)
FIX THIS FIRST
1. HIGH Anyone can see your Stripe secret key in your website's code
in /assets/index-4f9c2a.js value sk_l…(50 chars)
- PASS / WARN / FAIL = green (80+) / amber (50-79) / red (under 50) per area. SKIP = the area could not run (reason in a
note:line). - Overall score = weighted mean (security 40 %, the other three 20 % each). Any exposed secret or open table caps it at 39.
- Verdict: NOT READY = a HIGH security finding (or score under 50); ALMOST READY = other HIGH findings or score under 80; READY TO LAUNCH otherwise.
value sk_l…(50 chars): secrets are always cut to their first 4 characters. The full value is never printed or sent anywhere.- A Supabase anon / publishable key or a Google (Firebase) key in the page is normal and only INFO/LOW: safety depends on Row Level Security and key restrictions, which
--owner-checktests for Supabase and Firebase. - Exit code 0 = ready (score >=
--fail-under, nothing HIGH), 1 = findings, 2 = bad URL or the site could not be loaded (HTTP 403 usually means it blocks automated visitors: run it on a preview URL).
What to do next
- Exposed secret key (Stripe, OpenAI, Anthropic, AWS, Supabase service_role ...): it is already public. Move the
call server-side (API route, edge function), put the key in a server environment variable, and ROTATE the key
in the provider dashboard. Deleting it from the code is not enough.
- Open table (
db-table-open) or open Firebase: turn on Row Level Security / database rules with per-user
policies, then re-run with --owner-check until the table is gone from the list.
- Paste the "Paste to your AI" block from the HTML report (or
--prompt) into the AI that built the app, apply,
redeploy, re-run. Repeat until the verdict is READY TO LAUNCH.
- Headers (CSP, HSTS, X-Frame-Options, nosniff): set them in the host config (
vercel.json,netlify.toml,
_headers, Cloudflare rules) or the server middleware.
- "Google and ChatGPT see an almost empty page": the app renders all text with JavaScript; prerender or
server-render the public pages (landing, pricing) so crawlers read them.
How it stays safe
- Passive by default: it downloads only what a browser does (the page, its same-site scripts, links on the home
page, /favicon.ico, plus robots.txt, sitemap.xml, llms.txt for the search check). It never probes hidden
paths such as /.env or /admin.
- Hostile-input rules: http(s) only, redirects and scripts only on the same site, size and count caps, terminal
output stripped of control characters, HTML report fully escaped. --owner-check only talks to
https://.supabase.co, Firebase hosts, or the audited site itself, and never uses a service_role key.
scripts/polish_audit.py(phone design) andscripts/geo_audit.py(Google & AI search) are vendored copies of
the site-polish-audit and geo-seo-audit tools; launch_check.py runs them with --json.
Limits (honest)
- Not a penetration test: no login flows, no API fuzzing, no server-side auth or paywall checks, no write tests.
A good score means none of these common launch mistakes were found, not that the app is secure.
- Scripts served from another host (CDN subdomain) are not scanned; the report says how many were skipped.
- Secret patterns cover the common providers; custom tokens and keys inside source maps are not detected.
- Sites behind bot protection may answer 403 to the checker; rate limits (429) are not counted as broken links.
- Phone design and speed need a Chromium browser through Playwright.