Cavekit is a spec-driven development workflow for Claude Code built around a single SPEC.md file at the repository root. The loop is grill, spec, research, review, build: three commands run every time (/ck:spec, /ck:build, /ck:check) and four more are used only when a change is uncertain or has a wide blast radius. It runs in the main Claude thread with no sub-agents, dashboards, hooks, or orchestration binaries.
Three properties underpin the design. The spec is durable: it survives context resets and acts as the agent's long-term memory. The spec is written in caveman encoding, a compressed style of symbols, fragments, and pipe tables that uses about 75% fewer tokens than prose; the nine skill descriptions together occupy roughly 1.1k tokens of context. And a backprop reflex turns every test failure into a §B bug entry and recurring bug classes into §V invariants that the spec retains.
Development is frozen as of August 2026; the package still installs and works, and active work on the family continues in caveman and caveman-browse. The previous generation (v3.1.0, the Hunt lifecycle with 16 slash commands, 12 named sub-agents, parallel wave execution, and team mode) remains available at its tag for projects that want the autonomous, multi-agent loop.
Features
- The loop: /ck:spec (sole spec mutator), /ck:build (plan then execute, naming which test proves each invariant), /ck:check (read-only drift report)
- Reach-for commands: /ck:grill sharpens a fuzzy idea, /ck:research gathers sourced facts into §R, /ck:review runs an adversarial pre-build review with a go/no-go gate, /ck:deepen deepens one module
- Spec format: sections for goal, constraints, interfaces, research, invariants, tasks, and bugs; each verb owns specific sections
- Caveman encoding: compressed spec syntax that lowers the token cost of reading and writing the spec
- Backprop reflex: a six-step protocol that promotes test failures to spec bugs and bug classes to invariants
- Right-sizing rule: a one-line fix is just /build; the full chain is reserved for high-uncertainty work
- Install options: npx skills add, the Claude Code plugin marketplace, or a direct clone into ~/.claude/plugins/
