| name | mushanghai-onboarding |
| description | muShanghai Longevity Month 2026 onboarding bootstrap. Use when a user is new to the program (typically the May 12 AI-Empowered Onboarding session) and needs to: (1) confirm Claude Code is installed, (2) install the obra/superpowers + mattpocock/skills skill packs, (2.5) install Tavily + Firecrawl web-search MCP servers for live evidence research, (3) run brainstorming or grill-me to surface their seed hypothesis and pick a cohort + track, (4) hand off to mushanghai-experiment-leader to draft their proposal. |
| when_to_use | Trigger phrases: 'mushanghai onboarding', 'I am new to muShanghai', 'help me get started with muShanghai', 'install muShanghai skills', 'install muShanghai MCPs', 'May 12 onboarding', 'set up Claude Code for muShanghai', 'I want to join muShanghai', 'first day of muShanghai', 'install Tavily for muShanghai', 'install Firecrawl for muShanghai', 'add web search to my muShanghai Claude'. |
| allowed-tools | Read Bash |
muShanghai Onboarding
Valid for: May 12, 2026 onboarding session. Refresh install instructions before any later reuse.
Prerequisite: Read ../mushanghai-canon/references/safety.md before showing the banner to the user. Read ../mushanghai-canon/references/two-cohorts.md and ../mushanghai-canon/references/three-tracks.md so you can answer cohort/track questions.
Always: show the canon safety banner first
Before any other content, surface the safety banner verbatim from ../mushanghai-canon/references/safety.md. Do not skip this step.
Mode router
Ask one question:
Where are you in your muShanghai onboarding?
(A) I just walked in — set up Claude Code and the skill packs from scratch.
(B) I have Claude Code; install just the muShanghai-supporting skill packs.
(C) Everything is installed; help me figure out my seed hypothesis and which cohort/track to join.
- A → run Section 1 (Claude Code check) → Section 2 (skill packs) → Section 2.5 (web-search MCPs) → Section 3 (hypothesis surface) → Hand off
- B → skip Section 1; run Section 2 → Section 2.5 → Section 3 → Hand off
- C → skip Sections 1–2; ask whether Tavily / Firecrawl MCPs are already installed (
claude mcp list); run Section 2.5 only if missing → Section 3 → Hand off
Section 1 — Anthropic client preflight
Load references/anthropic-client-install.md. Walk through, in order:
- Confirm Node.js ≥ 18 (
node --version) — required by npx skills
- Confirm git (
git --version)
- Install an Anthropic Claude client. Two complementary options (the user can install both):
- Claude Desktop (macOS / Windows / ChromeOS — no Linux): download from https://claude.com/download. Includes a built-in Code tab that runs Claude Code natively.
- VS Code "Claude Code" extension (any OS with VS Code 1.98+): publisher
anthropic.claude-code. Shares ~/.claude/ with the Desktop app, so skills installed once are visible in both.
- Sign in with the user's Claude account in whichever client they installed.
Section 1.5 — (Optional) Install cc-switch
Load references/cc-switch-install.md. Walk through only if the user expects to switch between API providers (official Anthropic vs third-party relays, or multiple accounts):
- Confirm the user actually needs a provider switcher. If they only plan to use the official Anthropic API with one account, skip cc-switch entirely — it is optional.
- macOS install (recommended):
brew tap farion1231/ccswitch && brew install --cask cc-switch. Or download the Universal .dmg from https://github.com/farion1231/cc-switch/releases/latest.
- Open the app, add the providers the user has access to.
Important caveats the skill must surface:
- cc-switch is third-party (MIT, by Jason Young) — not endorsed by Anthropic.
- cc-switch manages Claude Code's config, not Claude Desktop's chat tab and not the VS Code extension's UI — though both surfaces read the same
~/.claude.json that cc-switch edits, so changes propagate to Claude Code regardless of surface.
- Hot-switching providers works for Claude Code without a restart; other managed CLIs (Codex, Gemini CLI, OpenCode) need a terminal restart.
Section 2 — Install the two skill packs
Load references/superpowers-install.md and references/mattpocock-skills-install.md.
Show the install commands. Show the exact commands; do not execute them silently. Wait for user confirmation before suggesting they run anything.
After install, run the macOS symlink workaround if the user is on macOS:
mkdir -p "${HOME}/.claude/skills"
for d in "${HOME}/.agents/skills"/*; do
[ -d "$d" ] || continue
skill="${d##*/}"
[ -e "${HOME}/.claude/skills/$skill" ] || ln -s "$d" "${HOME}/.claude/skills/$skill"
done
Verify by asking the user to run, in a fresh Claude Code prompt: "Tell me about your superpowers." Expected: Claude lists the superpowers skills.
Section 2.5 — Install web-search MCP servers (recommended)
Load references/search-mcp-install.md. Walk through, in order:
- Explain why: the user's seed hypothesis is a longevity claim that needs live evidence (PubMed abstracts, ClinicalTrials.gov pages, TCM literature). Without web-search MCPs, Claude Code can only reason from training data with no live citations.
- Install Tavily first (lowest-friction OAuth flow, 1,000 free credits/month). Sign up at https://app.tavily.com/home, then run the one-liner
claude mcp add command from the reference.
- Install Firecrawl second (500 free credits at signup). Sign up at https://firecrawl.dev/app, then run the one-liner.
- Show the exact commands; do not execute them silently. Wait for user confirmation before they run anything that embeds an API key into a URL.
- Remind the user to start a new Claude Code session — MCP config is read only at session start.
Verify by asking the user to run claude mcp list (should show both servers), then in a new session try: "Use Tavily to search for recent NMN clinical trials, then use Firecrawl to read the top result."
Skip allowed: If the user explicitly wants to work only from PDFs they already have, skip Section 2.5 and note it in the handoff message so the Track Leader knows their evidence base is offline-only.
Section 3 — Surface the seed hypothesis
Ask one question to pick the interview style:
How would you like to think through your seed hypothesis?
(a) Light alignment interview — quick one-question-at-a-time interview to align on cohort + track + a draft hypothesis. (Uses the grill-me skill from mattpocock/skills.)
(b) Full design session — longer brainstorming workflow that ends with a committed design doc. (Uses the brainstorming skill from obra/superpowers.)
Default recommendation: (a) light alignment for May 12 onboarding — most users want to walk away with cohort + track + hypothesis sketch, not a full spec.
Section 4 — Hand off to mushanghai-experiment-leader
After the user has a seed hypothesis + cohort + track choice, output a 1-line message they can paste to invoke the next skill:
"Help me draft my experiment proposal for muShanghai — cohort: [Longevity Month | BioBloom], track: [Decoding TCM | Lifestyle | Proof of Living], seed hypothesis: [one sentence]."
Remind the user that mushanghai-experiment-leader is the next step, and that their Track Leader has final approve/revise/reject authority.
References
Assets
