Drop-in starter templates that turn the agent from an eager intern into a disciplined collaborator. Plan. Confirm. Implement. Verify. Every single time.
60 seconds · macOS, Linux, Windows · Node 18+
Every kit-managed session walks the same four-step loop. The agent can't skip steps; the rules are read at session boot, the hooks enforce the non-negotiable parts.
For any task touching three or more files, the agent writes a plan to tasks/todo.md before touching code. No silent assumptions, no leap-before-look.
agent_docs/workflow.md
Plans pause on protected changes — new dependencies, schema migrations, auth logic, build system edits. Two approaches with tradeoffs, then your call.
CLAUDE.md → Protected Changes
Touch only files directly required. No opportunistic refactors. Unrelated issues logged under tasks/todo.md → Not Now. Discipline as a deletion test.
agent_docs/conventions.md
Mandatory order: typecheck, lint, tests, smoke test. The agent asks itself — would a staff engineer approve this? — and answers with evidence.
CLAUDE.md → Verification
“Would a staff engineer approve this?” — the question every task ends with.
The prompt: “Add a search feature to the users page.”
The agent immediately starts coding…
The agent restates, plans, asks, and verifies…
Same model · same prompt · different rules
Every piece is in plain markdown or shell. Read every line in a sitting. Audit before you trust.
Loaded automatically via Claude Code's semantic matching. Description-only triggering — no manual import.
/code-quality-auditCode smells, error handling gaps, maintainability — table-of-findings report.
/architecture-reviewSOLID, module boundaries, dependency health. The structural lens.
/deepening-reviewFind shallow modules. Grilling loop turns one into a deep one. Interactive.
/interface-designDesign It Twice. Three parallel sub-agents propose, one wins, hybrid optional.
/shipFull deploy pipeline — tests, coverage, CHANGELOG, bisectable commits, PR.
/debugEvidence-before-fix. Reproduce, isolate, fix, verify. Hypothesis ledger if it gets long.
The rules pages explain the system. The recipes show how to use it.
Custom /skill-name for your codebase. Auto-loads via semantic matching.
Port your existing rules. Map Cursor concepts to kit structure.
User-level rules + per-project overlays. Coordinated upgrades.
Each template ships a CLAUDE.md tuned for the framework and a CODEBASE_MAP.md scaffolded with the right sections.
Server Component rules, build verification, Prisma + Tailwind ready.
Layered architecture, API design conventions, Knex.js patterns baked in.
Async patterns, dependency injection, Alembic-aware migration playbook.
That works fine until your rules file passes 200 lines and you can't reason about it anymore. The kit wins on tiered context loading, deterministic hooks, project overlay separation, and multi-tool export. It loses on learning curve and bash dependency. Honest comparison →
“Versus raw Claude Code, .cursorrules, AGENTS.md, and Aider config — the comparison covers twelve dimensions including tiered context, deterministic enforcement, project overlay, and cross-tool export.”
Read the full comparison
One source. Seven outputs.
Edit one CLAUDE.md, run ./scripts/convert.sh all, ship to every tool your team uses.
One command. Sixty seconds. The next session your agent runs is shaped by the kit.
works on macOS, Linux, Windows · node 18+