Backprop

> Status: frozen (August 2026). cavekit is no longer in active development. > Everything below still installs and works, but expect no new features or > fixes. Active development of the family lives in > caveman and > caveman-browse.

cavekit

compressed spec-driven development for claude code one file · one loop · zero sub-agents

what this is

Plan-then-execute forgets. SDD remembers — but most SDD frameworks bury that value under agent swarms, dashboards, and ceremony that costs more tokens than it saves.

Cavekit is the simplest full loop: grill → spec → research → review → build, over one SPEC.md file, no sub-agents. Three commands you run every time; four more you reach for only when the change earns it.

The spine is three properties that earn their tokens:

• durable spec — SPEC.md at repo root survives context resets. It is the agent's long-term memory: lose the window, reload the spec, keep going.
• caveman encoding — ~75% fewer tokens than prose. Symbols, fragments, pipe tables. All nine skill descriptions cost ~1.1k context — 16× lighter than spec-kit's 18.6k. That is the whole point.
• backprop reflex — every test failure becomes a §B entry; classes of bug become §V invariants the spec never forgets.

And one rule that keeps it from bloating into the frameworks it replaces: right-size. A one-line fix is just /build. The full chain is for genuinely uncertain or high-blast-radius work — never for a typo.

commands

the loop — run these every time:

reach for these — only when the change earns the ceremony:

install

One line, via the skills CLI:

npx skills add JuliusBrussee/cavekit

Installs nine skills into ~/.claude/skills/: spec, build, check (the loop), grill, research, review, deepen (reach-for), plus caveman and backprop (the utilities). Claude activates each when its trigger context matches — e.g. "write a spec for…" invokes spec, a fuzzy idea invokes grill, a risky change before build invokes review. Claude Code picks them up on next launch.

Or via the Claude Code marketplace (also adds the /ck:spec, /ck:build, /ck:check, /ck:grill, /ck:research, /ck:review, /ck:deepen slash commands):

/plugin marketplace add juliusbrussee/cavekit
/plugin install ck@cavekit

Or clone directly:

git clone https://github.com/juliusbrussee/cavekit.git ~/.claude/plugins/cavekit

format

See FORMAT.md. Sections: §G goal, §C constraints, §I interfaces, §R research (optional, pipe table), §V invariants, §T tasks (pipe table), §B bugs (pipe table). Each verb owns specific sections — no verb rewrites a section it does not own.

files

FORMAT.md             spec schema + caveman encoding + sectioned ownership
commands/             seven thin slash-command entry points → the skills (loop + reach-for)
skills/spec           spec mutator — sole writer
skills/build          plan-execute, verification contract
skills/check          drift report
skills/grill          sharpen a fuzzy idea → §G/§C before spec
skills/research       external knowledge → §R, every finding sourced
skills/review         adversarial senior review of the spec → hardens §V
skills/deepen         spare-budget design pass — make one module deep
skills/caveman        encoding utility
skills/backprop       bug → spec protocol (six steps)

non-goals

• no sub-agents. Main Claude does the work.
• no dashboards. cat SPEC.md is the dashboard.
• no parallel workers. One thread, one spec, one diff.
• no JSON / YAML spec bodies. Markdown + pipe tables.
• no hooks, no orchestration binaries, no TypeScript helpers.

older cavekit (the Hunt lifecycle, v3.1.0 and earlier)

The previous generation is not deprecated — it is frozen at tag v3.1.0 and remains a fully working plugin.