claude-config

Problem

Every project on a machine ends up with the same agent problems: instructions that grow until they crowd out the work, skills that only run when someone remembers to call them, expensive models doing cheap retrieval, and a configuration that drifts quietly until the agent behaves differently from one week to the next. claude-config is the machine-wide answer: one configuration living in ~/.claude, applied to every project, with a verifier that says whether it is still intact.

Structure

• CLAUDE.md — the router, capped at 150 lines. It loads into every session, so it points at detail files instead of containing them.
• context/stack.md — the stack facts the router points to.
• Seven skills — grill-me, scaffold-project, verify-visually, capture-learning, brand-kit, campaign and video-pipeline — each triggered from its own description field.
• Two subagents with pinned models: bulk-researcher on a cheap model, visual-qa on a vision-capable model.
• settings.json — the permission rules, the status line and a stop-hook notification.
• statusline.py — model, directory, branch, context percentage and session cost in the status line.
• verify.py — 68 structural checks across five sections (prereqs, router, skills, agents, hygiene), exit-code gated for CI.

Decisions

• Context budget. A CLAUDE.md router capped at 150 lines routes to detail files instead of containing them.
• Description-driven dispatch. A skill states when it applies and the agent triggers it; nothing is invoked by hand.
• Model routing per task class. Bulk retrieval goes to a cheap model and visual verification to a vision model, which keeps cost predictable.
• Schema-validated frontmatter on every skill and agent, so a malformed file fails verification instead of failing quietly in a session.
• A 100-rule permission policy — 75 allow, 23 deny, 2 confirm — where deny beats allow.
• Deny-by-default .gitignore, so session keys and transcripts are structurally unstageable.
• No project-specific facts in global files. The verifier fails on known project strings.

Result

One command, python3 ~/.claude/verify.py all, returns 0 failing check(s). The same configuration applies to every project on the machine, and structural drift fails the check before it reaches a session.