Blueprint (JuliusBrussee) / cavekit — Summary

Important disambiguation: The GitHub repository JuliusBrussee/blueprint is NOT a planning tool named "Blueprint." The repository was renamed to cavekit (v4.0.0) as of current analysis. The README header reads "cavekit" and the plugin.json name field is "ck". This is a completely different product from imbue-ai/blueprint despite sharing the same repository name slug.

cavekit 4 is a compressed spec-driven development framework for Claude Code that compresses the entire SDLC into three slash commands and one file (SPEC.md) using "caveman encoding" — a ~75% token-reduction notation that replaces prose with symbols, fragments, and pipe tables. Its three commands are: /ck:spec (sole mutator of SPEC.md — creates, amends, or backprops bugs), /ck:build (single-thread plan→execute against spec with auto-backprop on failure), and /ck:check (read-only drift report listing §V/§I/§T violations). The central innovation is backpropagation: every test failure becomes a §B bug entry and a §V invariant in the spec, building institutional memory that prevents recurrence. With 948 stars and a previous v3.1.0 "Hunt lifecycle" version (12+ commands, 12 subagents), cavekit 4 is a deliberate minimization to the essential core. Compared to seeds, it is closest to spec-driver (spec-first, test-enforced, single-file SPEC.md) but differs in caveman encoding, explicit backprop protocol, and refusing all orchestration.

cavekit (JuliusBrussee/blueprint) — Overview

Origin

Created by JuliusBrussee. MIT license. Repository: JuliusBrussee/blueprint (renamed cavekit v4). 948 stars. Pushed 2026-05-17. Previous version v3.1.0 is frozen at git tag.

Philosophy

From README:

"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 4 is a rewrite from the ground up. It keeps only what earns its place."

cavekit's philosophy is aggressive minimalism: remove everything that doesn't pay for itself in token efficiency. The three things it keeps:

1. Durable spec — SPEC.md at repo root survives context resets
2. Caveman encoding — ~75% fewer tokens than prose
3. Backprop reflex — every test failure becomes a §B entry; classes of bug become §V invariants the spec never forgets

Caveman encoding

From README:

"~75% fewer tokens than prose. Symbols, fragments, pipe tables for repeating records."

Section markers: §G (goal), §C (constraints), §I (interfaces), §V (invariants), §T (tasks, pipe table), §B (bugs, pipe table).

Example task table:

§T
| id | status | task | cites |
|----|--------|------|-------|
| T1 | .      | impl auth mw | V2,I.api |
| T2 | ~      | add rate limit | V3 |

Status: . (pending), ~ (in progress), x (done)

Previous version (v3.1.0)

The prior "Hunt lifecycle" had: 4-command cycle (/ck:sketch → /ck:map → /ck:make → /ck:check) + 8 more commands (ship, review, revise, status, design, research, init, config, resume, help = 12 total), 12 named subagents, per-task token budgets, stop-hook state machine, model-tier routing, auto-backpropagation, tool-result caching, Codex peer review, knowledge-graph integration, caveman compression, parallel wave execution, team mode.

v4 deliberately removes all of this. v3.1.0 is preserved at the tag and still fully functional.

Non-goals

From README (explicit):

"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."

cavekit — Architecture

Distribution

• Claude Code plugin: JuliusBrussee/cavekit
• plugin.json name: ck
• Version: 4.0.0
• Install options:
  1. npx skills add JuliusBrussee/cavekit (agentskills.io skills CLI)
  2. /plugin marketplace add juliusbrussee/cavekit && /plugin install ck@cavekit (Claude Code marketplace)
  3. git clone https://github.com/juliusbrussee/cavekit.git ~/.claude/plugins/cavekit

Required runtime

None beyond Claude Code (or compatible AI host)

Directory tree

JuliusBrussee/blueprint/  (repo name, product is cavekit)
├── README.md
├── FORMAT.md             # SPEC.md schema + caveman encoding rules
├── plugin.json           # {"name": "ck", "version": "4.0.0"}
├── CHANGELOG.md
├── FORMAT.md
├── LAUNCH-POST.md
├── UPGRADE.md
├── commands/
│   ├── spec.md           # /ck:spec slash command
│   ├── build.md          # /ck:build slash command
│   └── check.md          # /ck:check slash command
└── skills/
    ├── spec/
    │   └── SKILL.md      # spec skill (mirrors commands/spec.md)
    ├── build/
    │   └── SKILL.md      # build skill (mirrors commands/build.md)
    ├── check/
    │   └── SKILL.md      # check skill (mirrors commands/check.md)
    ├── caveman/
    │   └── SKILL.md      # caveman encoding utility
    └── backprop/
        └── SKILL.md      # bug → spec backpropagation utility

SPEC.md sections (from FORMAT.md)

§G  goal (1 line, caveman)
§C  constraints
§I  interfaces (external surfaces)
§V  invariants (numbered V1, V2...)
§T  tasks (pipe table: id | status | task | cites)
§B  bugs (pipe table: id | date | cause | fix)

Output file

SPEC.md at repo root (single file, sole state artifact)

Target AI tools

Claude Code primary. agentskills.io-compatible tools via skills CLI.

cavekit — Components

Commands (3)

Skills (5)

Dispatch Modes in /ck:spec

Backprop Protocol (6 steps)

1. Parse bug description
2. Find root cause (read relevant code)
3. Decide: would a new invariant catch recurrence?
4. Append §B row: B<N>|<date>|<cause>|V<N>
5. Append §V invariant (if applicable)
6. If behavioral fix → add/update §T rows

Single State File

SPEC.md — the only persistent artifact. Tasks tracked as pipe table rows with . / ~ / x status.

cavekit — Uniqueness

Differs from Seeds

Closest seed is spec-driver (spec-first, test-enforced, single CLAUDE.md/SPEC.md file workflow). cavekit differs from spec-driver in three ways: (1) caveman encoding — ~75% token reduction vs prose specs, with a defined notation system (§G/§C/§I/§V/§T/§B) rather than free markdown; (2) formal backprop protocol — every test failure triggers a structured 6-step root-cause-to-invariant extraction that updates the spec, whereas spec-driver does not mandate spec learning from failures; (3) explicit non-goals — no subagents, no dashboards, no parallel workers, no YAML/JSON bodies — these are stated design choices, not gaps. Compared to kiro (spec storage in .kiro/specs/requirements.md/design.md/tasks.md), cavekit uses a single flat SPEC.md and custom section syntax rather than multiple files. cavekit is the only framework in this batch or the seeds with a designed token compression encoding for spec files.

Two-Blueprint Disambiguation

JuliusBrussee/blueprint (this repo, slug: blueprint-juliusbrussee) is cavekit 4 — a compressed spec framework. imbue-ai/blueprint (slug: blueprint-imbue) is a planning Q&A tool. They share only the name "blueprint" as a previous identity. See batch notes.

Version history note

v3.1.0 (frozen at tag) is a completely different product — 12+ commands, 12 subagents, parallel waves, peer review. v4.0.0 is a deliberate minimization. Both are valid and functional; v4 is the "distilled core," v3.1.0 is the "autonomous loop." The README recommends choosing based on use case, not defaulting to v4.

Positioning

cavekit occupies the "maximum token efficiency, minimum ceremony" niche in SDD frameworks. It is the only framework that makes token budget a first-class design constraint — the encoding notation exists specifically to keep SPEC.md loadable in full on every turn.

Observable Failure Modes

1. Caveman encoding learning curve: Teams unfamiliar with §V/§B notation must read FORMAT.md before contributing to SPEC.md.
2. Single file as bottleneck: Large projects with many modules may outgrow the single-SPEC.md model (tasks.md + modules could be hundreds of rows).
3. Repository name confusion: JuliusBrussee/blueprint is the GitHub repo name but the product is cavekit — discoverability and naming are inconsistent.
4. Manual plan approval: /ck:build shows the plan and waits for user OK before executing, breaking autonomous pipelines.
5. No merge/PR automation: cavekit tracks task status but does not create branches, commits, or PRs automatically.