cavekit
A Claude Code plugin: spec-driven development over a single SPEC.md file, a loop of seven commands with no sub-agents
Medium risk
We rate an entry medium when the tool runs code, makes network calls or reads project files. Check what exactly it does before installing.
Why this level
- The build command plans and executes changes in the project code
- Reads repository files and runs tests
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/cavekitDetects the agents on your machine, checks the risk and pins the version.
Other ways to install
Run one by one in the Claude Code chat
/plugin marketplace add juliusbrussee/cavekit
/plugin install ck@cavekit-marketplaceInstall the plugin: /plugin marketplace add juliusbrussee/cavekit, then /plugin install ck@cavekit. Or install just the skills with npx skills add JuliusBrussee/cavekit.
Other ways from the author
/plugin marketplace add juliusbrussee/cavekit
/plugin install ck@cavekitInstalls the plugin with the ck skills and slash commands.
This is third-party code. Review the repository files before installing.
What it does
Cavekit folds spec-driven development into one loop over a SPEC.md file at the repository root. Three core commands run every time: spec creates and amends the spec, build plans and executes changes against it, and check gives a read-only report on drift between code and spec. Four more commands are for when the change earns them: grill sharpens a fuzzy idea with questions, research gathers external facts with sources, review runs an adversarial check of the spec before build, and deepen makes one shallow module deep. The spec is written in a compact caveman encoding to save tokens, and every test failure becomes a bug entry and an invariant the spec remembers. Work stays in a single thread, without sub-agents or orchestration.
Who it is for. For developers and tech leads who work from a spec and want to keep context in one file.
Good fit when
- You need to run non-obvious work from a spec that survives a context reset
- You need to check drift between code and stated requirements
- You need to lock every bug you find as an invariant so it does not return
Not a fit when
- A one-line fix where the whole spec loop is overkill
- You want a swarm of sub-agents and parallel workers, cavekit runs single-threaded
Example request
Write a spec for this task, then build the implementation against it and show which test covers each invariantLimitations
The project is frozen since August 2026: it installs and works, but the author promises no new features or fixes, and development of the family moved to the caveman and caveman-browse repositories. The build and check commands target Claude Code and its slash commands. The previous generation stays a working plugin at tag v3.1.0 with an autonomous loop and sub-agents if that is what you need.
How to disable. Remove the plugin via /plugin, drop the cavekit marketplace, or delete the cavekit skills folder from the skills and plugins directory.
Security check
- The build command plans and executes changes in the project code
- Reads repository files and runs tests
README in short
The README presents cavekit as compressed spec-driven development for Claude Code: one SPEC.md file, one loop, no sub-agents. It describes three core commands and four reach-for ones, plus the caveman encoding that lowers token use and the protocol that turns bugs into invariants. Install works via the skills CLI, the Claude Code marketplace, or cloning the repository. It documents the sectioned spec format and the rule against bloating the process on small edits. The README openly marks the frozen status and keeps the earlier v3.1.0 generation as a working option.
FAQ
It is frozen, should I use it?
The plugin installs and works; the author simply adds no new features. Active development lives in the caveman repository, and cavekit fits if you want a ready compact loop.
How do the three core commands differ from the rest?
spec, build and check are the every-time loop. grill, research, review and deepen are only for when the change earns the extra step.
Related
A skills library that gives coding agents a development process: brainstorming, planning, TDD, subagents and code review
Skills for real engineers by Matt Pocock
Skills For Real Engineers
Small composable skills for engineering with agents: plan grilling, TDD, bug diagnosis, code review and architecture
GitHub toolkit for spec-driven development: the specify CLI adds agent commands and skills to a project, from principles to implementation
Reference MCP servers
Model Context Protocol servers
Official reference MCP servers: Filesystem, Fetch, Git, Memory, Sequential Thinking, Time and Everything