Self-installable Claude Code skill via HTTP

Lets a teammate install your skill with a single curl-pipe-bash command: curl -fsSL http://your-host/install/skill-name.sh | bash. The installer fetches a SKILL.md over plain HTTP, drops it into ~/.claude/skills/, and prompts the user to set two env vars (API URL + bearer token). New teammates onboard in 30 seconds; the LLM becomes the frontend to your internal platform.

Building a UI for every internal operation is expensive and rarely pays back. Exposing a versioned REST API and shipping a Claude Code skill that talks to it lets any teammate run "show me founders," "send email to @x," "list replies" in natural language without you maintaining a frontend. A founder-discovery CLI uses this pattern and rolls onboarding to a 30-second curl.

1. Build a versioned REST API under /api/v1/* with bearer-token auth.
2. Write SKILL.md documenting each endpoint and the natural-language workflows ("list founders" = GET /api/v1/founders/feed).
3. Serve SKILL.md over HTTP at /install/skill-name.md as raw text — never via the Read tool from a git checkout (otherwise the parent CLAUDE.md gets injected as a <system-reminder> and confuses everything).
4. Ship a bash installer at /install/skill-name.sh that does: create ~/.claude/skills/skill-name/, curl SKILL.md, prompt for env vars, write to shell rc.
5. Generate per-user API keys (a small CLI: manage_keys.py issue <email>).
6. In SKILL.md, use ${VAR:?missing} shell idiom so misconfigured installs fail loudly, not silently.

• Serve raw text, not git-cloned files. If another Claude reads via the Read tool from a local checkout, the parent project's CLAUDE.md is injected as a system-reminder — looks like prompt injection, breaks behavior.