AI-firstify audit: add .claude/skills/ and expand CLAUDE.md

[tofutim]

Context

An AI-firstify audit was performed across the srclight ecosystem (srclight engine + srclight-app). Model-radar shares the same infrastructure patterns and would benefit from the same improvements.

Recommendations

1. Create .claude/skills/ for repeatable workflows

Model-radar has complex operational workflows that should be captured as prescriptive Claude Code skills:

• release/ — Version bump, PyPI publish, MCP registry update workflow
• provider-setup/ — Adding a new provider (API key config, model list refresh, verification)
• benchmarking/ — Running quality benchmarks across models, interpreting results

2. Review CLAUDE.md for AI-first principles

Check that CLAUDE.md:

• Is under 200 lines and well-structured
• Covers build/test/release procedures prescriptively
• Uses progressive disclosure (references/ for detailed docs)
• Includes a "do/don't" section for agent behavior

3. Safety audit

• Verify no API keys are committed (they should only come from environment/config)
• Confirm .gitignore excludes ~/.model-radar/config.json, .env, credentials
• Ensure MCP tools that modify state (configure_key, restart_server) have appropriate annotations

4. Expand MCP tool instructions

The server instructions are already good but could benefit from:

• A "Setup Workflow" section showing the agent-driven happy path
• Clearer separation of read-only vs write tools in the tool guide

Reference

See srclight/srclight-app commit history for the pattern — skills, MCP tool expansion, and widget extraction were done there as part of this audit.

Acceptance Criteria

• .claude/skills/ directory exists with at least one skill
• CLAUDE.md reviewed and updated if needed
• No credentials in committed code
• MCP server instructions include setup workflow

[tofutim]

📡 Conductor dispatched agent (pid 445232, Claude Code (default)) in /home/tim/Projects/srclight/model-radar on feature/issue-9-ai-firstify-audit-add-claude-skills-and (worktree)

[tofutim]

🤖 Agent picking this up. Working in /home/tim/Projects/srclight/model-radar.

[tofutim]

📍 Audit complete. CLAUDE.md (182 lines, private/gitignored) is well-structured. No credentials in code. .gitignore needs .env coverage. MCP instructions need setup workflow happy path and read-only/write separation. Now creating CLAUDE.md public version, .claude/skills/, and expanding MCP instructions.

[tofutim]

✅ Done

PR #10 opened: #10

Changes:

1. .claude/skills/ — 3 prescriptive skills: release (version bump + PyPI + MCP registry), provider-setup (end-to-end new provider), benchmarking (quality audit + tier validation)
2. CLAUDE.md — Public version (88 lines) with commands, key modules, MCP tool classification (read-only/execution/write/informational), version bumps, release process, do/don't section, docs references
3. MCP instructions — Added setup workflow happy path (7-step agent-driven onboarding), separated read-only tools from write tools in the tool guide
4. Safety — No credentials in code. Added .env/.env.* to .gitignore. Un-gitignored CLAUDE.md so agents can read it.

All 167 tests pass.