___                ___  _             _   
  / __\__ ___   _____/ __\| |_ __ _  ___| | __
 / /  / _` \ \ / / _ \__ \| __/ _` |/ __| |/ /
/ /__| (_| |\ V /  __/___) | || (_| | (__|   < 
\____/\__,_| \_/ \___|____/ \__\__,_|\___|_|\_\

AI talk too much. CaveStack fix.

Left: default verbose /review. Right: CaveStack /review, same patch. Same findings. ~250 words vs ~40.

---

What This

AI builder framework. Caveman mode = default. Every response: short, direct, no fluff. Same 40 skills. Same power. 75% fewer words.

Other AI tools: walls of text. Filler words. "I'd be happy to help you with that." Apologies for things that aren't wrong. Summaries of what you just said back to you.

CaveStack: answer. Done.

Install (one line)

Need Claude Code + Git. Bun auto-installs if missing (with SHA256-verified installer).

curl -fsSL https://cavestack.jerkyjesse.com/install | sh

Open new Claude Code session. Caveman mode active. No /caveman needed. Just works.

Hero Six (featured in v1.0)

Six judgment-layer skills embody the "think before code" moat:

Skill	What it does
/office-hours	Premise challenge + design doc before building. Six forcing questions.
/investigate	Systematic root-cause debugging. Reproduce → isolate → diagnose → fix.
/plan-eng-review	Architecture + tests + coverage lock before implementation.
/plan-design-review	UI/UX audit with mockups before implementation.
/plan-devex-review	Developer experience critique with personas + benchmarks.
/cso	Security audit: OWASP Top 10, STRIDE, supply chain.

34 more skills ship alongside — /ship, /review, /qa, /checkpoint, /health, /retro, /browse, /autoplan, /codex, and more. Invoke any by full name. Run cavestack-skills list in the terminal or /help inside Claude Code to see all 40.

What You Get

Thing	What It Do
40 skills	All installed, all invokable. Hero six featured, others by full name.
Caveman default	No command needed. First response = terse. Automatic.
Intensity dial	/caveman lite (gentle), full (default), ultra (maximum grunt)
Skill discovery	cavestack-skills list or /help — no website round-trip
Short aliases	cs-* for every cavestack-* CLI. Speed over typing.
Tier-2 errors	Every error: what broke + why + exact fix + docs link
Local DX metrics	cavestack-dx show — your TTHW and skill events, zero network
Productivity wrapper	cavestack run <task> with --record (redaction-gated share)
Windows-first	Bun compiles, Git Bash symlinks, PowerShell statusline. All work.
Reversible	cavestack-uninstall. Clean removal. Your files untouched.
Headless browser	/browse for QA, screenshots, page testing. Built in.
Design tools	/design-consultation, /design-review, /design-shotgun
Security audit	/cso — OWASP Top 10 + STRIDE threat modeling
Zero telemetry	No remote data ever. /methodology for how savings are measured.
Chars not tokens	Benchmark measures characters (model-agnostic). Same unit every terminal can count, no API key needed.

Try It Live

See the interactive demo — watch caveman mode compress verbose AI output in real-time.

Before / After

Verbose Claude	CaveStack Claude
"I'd be happy to help you with that! Let me take a look at your code..."	[reads code]
"The issue appears to be related to the authentication middleware where the token expiry check is using a less-than operator instead of less-than-or-equal-to..."	"Bug in auth middleware. Token expiry check use < not <=. Fix:"
"Sure! I'll analyze the changes and provide a comprehensive review..."	[reviews diff, reports findings]
47 lines explaining what it's about to do	5 lines doing it

Why Fork, Not Plugin

Default behavior win. Every opt-in terse workaround — /caveman plugin, be terse in CLAUDE.md, hand-written prompts — require user to remember. They don't. Tool stay verbose. Senior eng close window, go back to grepping.

Only way change default = fork framework, ship terse as baseline.

Also: caveman need fire before first prompt. Need SessionStart hook owned by install. Fork own that. Plugin can't.

Caveman Control

# During session:
stop caveman          # back to verbose (why though)
normal mode           # same thing
/caveman              # re-enable (default: full)
/caveman lite         # gentle compression
/caveman full         # classic caveman
/caveman ultra        # maximum grunt. fragments only.

# Permanent:
cavestack-settings-hook remove-caveman   # disable hooks entirely
cavestack-settings-hook install-caveman  # re-enable

All Skills

/review          code review your diff
/ship            test, review, version, push, PR
/qa              QA test site + fix bugs
/investigate     systematic root-cause debugging
/office-hours    brainstorm ideas, startup diagnostic
/cso             security audit (OWASP + STRIDE)
/design-review   visual QA + fix loop
/browse          headless browser commands
/retro           weekly engineering retrospective
/codex           second opinion from OpenAI Codex
/plan-ceo-review     CEO-mode plan review
/plan-eng-review     architecture + test review
/plan-design-review  UI/UX design review
/autoplan        run all reviews in sequence
/checkpoint      save/resume work state
/health          code quality dashboard
...and 25 more. Full list: /caveman-help

Trouble?

Skill not show up? cd ~/.claude/skills/cavestack && ./setup

/browse fail? cd ~/.claude/skills/cavestack && bun install && bun run build

Stale install? /cavestack-upgrade — or auto_upgrade: true in ~/.cavestack/config.yaml.

Caveman not fire on new session? Check ~/.claude/settings.json has hooks.SessionStart pointing at caveman-activate.js and hooks.UserPromptSubmit pointing at caveman-mode-tracker.js. Re-register: cavestack-settings-hook install-caveman. Need Node.js on PATH — hooks run under Node, not Bun.

Windows? Works on Windows 10/11 via Git Bash or WSL. Both bun and node on PATH. Bun has known Playwright pipe issue (bun#4253), browse falls back to Node for server.

./setup build error on Windows? Known upstream glob issue in browse/scripts/build-node-server.sh. Main binaries still build. Non-blocking.

Claude can't see skills? Add cavestack section to project's CLAUDE.md. /office-hours does this automatically on first run.

Bug? github.com/JerkyJesse/cavestack/issues — include OS, bun --version, node --version, exact error.

Uninstall

~/.claude/skills/cavestack/bin/cavestack-uninstall

Remove symlinks, hooks, state. Your files untouched.

Credit

MIT. Caveman hooks: Julius Brussee. Built by JerkyJesse. See LICENSE.

---

Verbose README (for the wordy among us)

What is CaveStack?

CaveStack is an AI builder framework that ships with "caveman mode" enabled by default. Caveman mode compresses AI responses by approximately 75% without losing technical accuracy. Instead of opting into terse responses, CaveStack makes them the default behavior from the very first prompt.

Why does this exist?

Every AI coding tool ships with verbose output as the default. Users who prefer concise responses must remember to configure terse mode each session — adding /caveman commands, writing custom instructions in CLAUDE.md, or manually requesting brevity. Most don't bother. The tool stays verbose. Senior engineers close the window and go back to grepping.

CaveStack solves this by forking the upstream framework and making terse the baseline. Caveman mode fires on SessionStart before the first prompt is even processed, so there's nothing to remember and nothing to configure.

Installation

CaveStack requires Claude Code, Bun v1.0+, Node.js, and Git. To install:

git clone https://github.com/JerkyJesse/cavestack.git ~/.claude/skills/cavestack
cd ~/.claude/skills/cavestack && ./setup

After setup, open a new Claude Code session in any project directory. Caveman mode will be active automatically — no additional commands or configuration needed.

Features

CaveStack includes all 40 skills from the CaveStack framework:

• Code Review (/review) — Analyze diffs for bugs, security issues, and style problems
• Ship (/ship) — Test, review, version bump, push, and create PR in one command
• QA Testing (/qa) — Systematically test web applications and fix bugs found
• Investigation (/investigate) — Structured root-cause debugging with four phases
• Office Hours (/office-hours) — Brainstorm ideas with startup diagnostic or builder mode
• Security Audit (/cso) — OWASP Top 10 and STRIDE threat modeling
• Design Tools (/design-consultation, /design-review, /design-shotgun) — Design system creation, visual QA, and variant exploration
• Headless Browser (/browse) — Navigate pages, take screenshots, test interactions
• Retrospective (/retro) — Weekly engineering retrospective with trend tracking
• Plan Reviews (/plan-ceo-review, /plan-eng-review, /plan-design-review) — Multi-perspective plan review pipeline
• Deploy (/land-and-deploy, /canary, /benchmark) — Merge, deploy, and monitor
• And 25+ more — Run /caveman-help for the full list

Additional CaveStack-specific features:

• Always-on caveman mode with three intensity levels: lite (gentle compression), full (classic caveman, the default), and ultra (maximum grunt, fragments only)
• Windows-first installation — Bun binaries compile, setup symlinks work under Git Bash, and the statusline is PowerShell-aware
• Fully reversible — cavestack-uninstall removes all symlinks, hooks, and state files without touching your project files
• Caveman voice on all skill templates — Not just the chat output, but the skill instructions themselves are compressed, reducing output size across the board. We measure in characters, not tokens, because tokens vary by model (GPT, Claude, Gemini all count differently). Characters are model-agnostic and every terminal can count them.

Controlling Caveman Mode

During any session, you can adjust or disable caveman mode:

• stop caveman or normal mode — Return to standard verbose output
• /caveman — Re-enable caveman mode (defaults to full)
• /caveman lite — Gentle compression, mostly drops filler words
• /caveman full — Classic caveman: fragments, no articles, short synonyms
• /caveman ultra — Maximum compression, fragments only

To permanently disable caveman mode, run:

~/.claude/skills/cavestack/bin/cavestack-settings-hook remove-caveman

To re-enable it later:

~/.claude/skills/cavestack/bin/cavestack-settings-hook install-caveman

Troubleshooting

Problem	Solution
Skills not showing up	cd ~/.claude/skills/cavestack && ./setup
/browse fails	cd ~/.claude/skills/cavestack && bun install && bun run build
Stale install	Run /cavestack-upgrade or set auto_upgrade: true in ~/.cavestack/config.yaml
Caveman not firing	Verify ~/.claude/settings.json has SessionStart and UserPromptSubmit hooks pointing at caveman-activate.js and caveman-mode-tracker.js. Re-register with cavestack-settings-hook install-caveman. Ensure Node.js is on PATH.
Windows build error	Known upstream glob issue in browse/scripts/build-node-server.sh. Primary binaries still build successfully. Non-blocking for most workflows.
Claude can't see skills	Add a cavestack section to your project's CLAUDE.md. The /office-hours skill does this automatically on first run.

Reporting Bugs

Open an issue at github.com/JerkyJesse/cavestack/issues with your OS, bun --version, node --version, and the exact error output.

License

MIT. See LICENSE for full attribution covering caveman (hooks) and CaveStack.