blueprint (repository name Cavekit, author JuliusBrussee) is a compressed spec-driven development plugin designed for Claude Code, currently in frozen status (August 2026). Cavekit is no longer under active development, but the features described below can still be installed and work normally; there will simply be no new features or fixes. Active development of this tool family has moved to caveman and caveman-browse.
Purpose
This plugin solves the problem that the "plan first, then execute" pattern is easy to forget: spec-driven development (SDD) can remember context, but most SDD frameworks are dragged down by agent swarms, dashboards, and ceremonial processes, consuming more tokens than they save. Cavekit provides the simplest complete loop: grill → spec → research → review → build, centered on a single SPEC.md file, without using subagents. Three commands are used every time, and the other four are used only when a change is worth it.
Core Features
- Persistent spec: SPEC.md at the repository root survives context resets and serves as the agent's long-term memory—when the window is lost, the spec can be reloaded to continue working.
- Caveman coding: reduces tokens by about 75% compared with prose. It uses symbols, fragments, and pipe tables. All nine skill descriptions occupy only about 1.1k context, 16 times lighter than spec-kit's 18.6k.
- Backprop reflection: each test failure becomes a §B entry; a class of bug is escalated to a §V invariant, and the spec will never forget it again.
- Right-size: a one-line fix only needs /build; the full chain is used only for work that is truly uncertain or has broad impact, never for typos.
Main Capabilities
Nine skills: spec, build, check (the core loop), grill, research, review, deepen (used as needed), plus caveman and backprop (utility tools). Claude activates each skill when the trigger context matches, for example "write a spec for…" invokes spec, a vague idea invokes grill, and a risky change before building invokes review. Claude Code will automatically recognize them on the next startup.
Usage
It can be installed through the skill CLI (installing the nine skillsInstall to ~/.claude/skills/); or install via the Claude Code marketplace (which also adds the /ck:spec, /ck:build, /ck:check, /ck:grill, /ck:research, /ck:review, /ck:deepen slash commands); or clone the repository directly. The SPEC.md format is described in FORMAT.md and includes sections: §G Goals, §C Constraints, §I Interfaces, §R Research (optional, pipe table), §V Invariants, §T Tasks (pipe table), §B Bugs (pipe table). Each verb owns specific sections and will not rewrite sections that do not belong to it.
Design principles
No subagents (the main Claude does the work), no dashboard (cat SPEC.md is the dashboard), no parallel workers (single-threaded, single spec, single diff), no JSON/YAML spec body (Markdown + pipe tables), no hooks, no orchestration binary, no TypeScript helper scripts.
Legacy version and choices
The previous generation cavekit (Hunt lifecycle, v3.1.0 and earlier) is not deprecated but frozen at tag v3.1.0 and remains a fully usable plugin. It includes the four-command Hunt lifecycle (/ck:sketch → /ck:map → /ck:make → /ck:check), plus /ck:ship, /ck:review, /ck:revise, /ck:status, /ck:design, /ck:research, /ck:init, /ck:config, /ck:resume, /ck:help, for a total of 16 slash commands, 12 named subagents, as well as per-task token budgets, a stop-hook state machine, model-tier routing, automatic backpropagation on test failure, tool result caching, Codex peer review, Karpathy behavioral guardrails, caveman token compression, knowledge graph integration, design system enforcement, parallel wave execution, and team mode. If you need a fully autonomous loop, parallel agents, peer review, or design system workflows, choose v3.1.0; if you want a streamlined loop (single spec, no orchestration, ceremony only where it fits), choose v4. Migration instructions are in UPGRADE.md, and it is a two-way door—SPEC.md is plain Markdown and will not lock you into either direction.
Applicable scenarios and limitations
Suitable for developers using Claude Code who want to manage development specs in a lightweight, low-token way. Limitations: the project is frozen, with no new features or fixes; no further information is provided on pricing, cost, or platform compatibility; no actual usage results or performance data are provided; the current materials do not mention support for non-Claude Code environments.
