shenxingy/init-profile

Init-Profile Skill

Capture project-specific facts that the SessionStart hook will inject every time. Pairs with the Environment Fingerprint block (host-level facts) by adding the Project Profile block (project-level facts).

What it does

Writes .claude/PROJECT_PROFILE.md with five sections, filled in interactively. The session-context.sh hook auto-loads this file on every SessionStart, so the agent has the facts upfront and stops guessing.

Why it exists

Without a profile, the agent re-discovers the same facts each session ("where does prod run?", "is this hostname local or remote?", "where are the env secrets?"). With a profile, those facts live in the system prompt from session 1.

Process

1. Detect what already exists. Read these for clues:

   • CLAUDE.md (deployment URLs, commands)
   • .env, .env.example, .env.*.local, brands/*.local.env
   • package.json, Dockerfile, docker-compose.yml, vercel.json, netlify.toml
   • ~/.profile / ~/.zshrc for project-specific env vars
   • Recent git log for clues on dev workflow

2. Ask the user (one question per section, skip if already obvious from step 1):

   a. Hosts & topology

   • Which hostname is "production"? Which is "dev"? Are any of them THIS machine?
   • Examples: "prod = vercel.com (remote)", "Aries = this machine, runs local dev"

   b. Service URLs

   • Live production URL?
   • Local dev URL(s) and ports?

   c. Secret/env file locations

   • Where are the real secrets stored? (gitignored .env paths)
   • Which env vars are critical and where do they come from?

   d. Verification commands

   • One-line "is this thing healthy" commands per environment.
   • Example: gh api repos/x/y/actions/runs --jq '.workflow_runs[0].conclusion'

   e. Aliases the user uses for this project

   • "When I say X, I mean Y." (e.g., "the brand env" = brands/*.local.env)

3. Write .claude/PROJECT_PROFILE.md in this format:

   # Project Profile — <project name>
   
   ## Hosts
   - <hostname>: <role> (this machine | remote)
   - <hostname>: <role> (this machine | remote)
   
   ## URLs
   - Production: <url>
   - Local dev: <url(s)>
   
   ## Env / secrets
   - <path>: <what's in it>
   
   ## Verify commands
   - <env>: `<command>`
   
   ## Aliases
   - "<phrase user uses>" → <what it actually maps to>
   

4. Confirm — show the file content, ask the user to approve before writing.

5. Do NOT git-commit. The user decides what's safe to commit. .claude/PROJECT_PROFILE.md may contain internal hostnames or URLs that shouldn't be public — recommend .gitignore if the project is public.

Output

• File: .claude/PROJECT_PROFILE.md
• One-line confirmation of where it landed and that it'll auto-load next session.

Notes

• Keep the profile under 50 lines. Long profiles bloat every SessionStart prompt.
• If a fact changes (new prod URL, new hostname), edit the file directly — no need to re-run /init-profile.
• This skill is project-scoped. Host-level facts (Tailscale IP, sibling projects, gh/gcloud auth) are captured automatically by the Environment Fingerprint block in session-context.sh — don't duplicate them here.