Building Quiet
The public R&D bench behind @build.inquiet. Real Claude Code hooks, skills and the free products, all in one place. Faceless. Reproducible. Free.
Every file here runs in a real working setup. Nothing is mocked for the camera, and this repo is the only code that ever appears on camera.
| Folder | What |
|---|
hooks/ | 47 production hooks for Claude Code. Discipline at the tool-call layer: economic-burn prevention, routing audit, completion-depth enforcement, anonymity-class checks, claim-source provenance. Reference below. |
skills/ | SKILL.md files: voice-dna-multi-brand (one voice file per brand, no bleed), auto-memory-doctrine. |
products/ | Source for the free Gumroad downloads. Each one is built from a tagged commit by scripts/build_product.py, so what you download is exactly what is here. |
demos/ | Small, self-contained examples used in videos. |
tests/ | Tests for the more complex hooks. |
tools/ | rules_audit.py: what Claude Code loads before you type, and what makes rules slip. |
SKILLS.md | Community-sourced skills that pair with these hooks. |
Free list: 5 Claude Code plugins to install first
Four from Anthropic's official marketplace, one from Addy Osmani, every install line included: lists/5-free-claude-code-plugins.md.
Free list: Claude Code subagents, one page per job
Four example agent pages (hook-finder, script-writer, reviewer, cutter) with the docs behind each field: lists/claude-code-subagent-pipeline.md.
Which model for which job, as a three-row table with a subagent page: lists/which-model-for-which-job.md.
Running several sessions at once without collisions: lists/run-parallel-sessions-without-collisions.md.
Free tool: Rules Audit
python3 tools/rules_audit.py /path/to/your/project
Lists every instruction file Claude Code loads before you type a word, with its size, line count and a token estimate. It flags four things that make rules stop holding: a loaded file over 200 lines, a MEMORY.md past its load limit, an @import that points nowhere, and the same rule written in two loaded files. It also reads your permission settings and flags the gaps that matter when Claude runs unattended: an allow rule that lets every shell command run, an allow list with no deny rules, and no deny rule for .env or secrets files. Read-only, standard library only. --selfcheck runs its 26 tests. The load rules come from the memory docs.
Next step, if you want it: the Proof Kit adds guards that block instead of warn. $39: buildinquiet.gumroad.com/l/proof-kit
Free plugin: the Starter
claude plugin install starter --marketplace buildinquiet-tech/building-quiet
A CLAUDE.md template, three setup skills (/starter:claude-md, /starter:done-when, /starter:check-setup) and two guards that block rm -rf, force pushes and .env reads, and still block if they break. Claude Code 2.1.295+. What each one is for: plugins/starter/WHICH-ONE-WHEN.md.
Free downloads
The Proof Kit ($39) is the paid plugin: diff-review checks Claude's own diff against your spec before it says done, plus fail-closed guards for shared checkouts and scheduled routines: buildinquiet.gumroad.com/l/proof-kit. This repo is the readable starting point; you never need the paid one to use anything here.
What this is not
- Not a course. A reference bench: here is the file, here is what it does.
- Not a "look at my system" diagram. Every file is the real version.
@build.inquiet is operator-anonymous. The system is the star.
Why this exists
Claude Code is incredibly powerful and incredibly easy to misuse. Without guardrails, an agent will:
- Burn the wrong model on bulk-mechanical work that should have routed to a spoke (~50% cost overrun).
- Make claims about state ("the build passed", "the queue is empty", "$X spent") without re-verifying.
- Edit files outside its declared scope.
- Repeat the same failure across sessions because no telemetry survives.
- Skip pre-read requirements for skills, plans, and doctrine files.
- Drift between what's claimed and what's actually on disk.
The 47 hooks in this repo close each of those gaps at the tool-call layer — fail-open, never bricking your work, but loudly telling you when discipline is slipping.
Quickstart
# 1. Clone into your Claude Code working tree
git clone https://github.com/buildinquiet-tech/building-quiet.git
# 2. Copy the hooks (or symlink) into your .claude/hooks/ dir
cp building-quiet/hooks/*.py /path/to/your/project/.claude/hooks/
# 3. Wire them into .claude/settings.json (see "Wiring" below)
# 4. Set the project dir env var (if not auto-set by Claude Code)
export CLAUDE_PROJECT_DIR=/path/to/your/project
All hooks default to advisory behavior — they emit [ADVISORY] messages to stderr but exit 0, never blocking. Promote individual hooks to blocking via per-hook env vars (e.g., ORCH_ADVISOR_TIER=blocking).
This is the beginning of the README. Read the rest at the original. Read the full README ↗
