claude-scaffold

claude-scaffold

**[English](README.md)** | **[Русский](README.ru.md)**

**Claude Code infrastructure for ML and AI engineers — deploy once, sync everywhere, and cut input token costs by 71%.**

Three things in one repo: **scaffolding** (22 skills, profiles, hooks deployed to any project), **token optimization** (bash output filters + context defaults measured at 71.4% savings on Sonnet 4.6), and **multi-repo management** (`update --all` keeps every project in sync with one command).

[](https://github.com/pyramidheadshark/claude-scaffold/actions/workflows/ci.yml) [](https://www.npmjs.com/package/claude-scaffold) [](https://www.npmjs.com/package/claude-scaffold)

---

Three Pillars

1. Scaffolding — deploy discipline to every project

claude-scaffold is a **central infrastructure layer** you configure once and deploy everywhere. 22 domain skills load automatically based on what you're working on. Profiles adapt the CLAUDE.md for different roles. Hooks enforce quality checks without any manual wiring.

claude-scaffold  ← you configure this once
      │
      ├── deploy → project-a/.claude/
      ├── deploy → project-b/.claude/
      └── deploy → project-c/.claude/

2. Token Optimization — 71.4% input token savings (measured)

v2.1.0 ships a bash output filter hook that wraps verbose commands before they run. Instead of Claude reading 2,000 lines of `pytest` output, it reads the 20 lines that matter.

Benchmark: 25 tasks × Sonnet 4.6 via OpenRouter
  Baseline:  25,084 input tokens
  Optimized:  7,178 input tokens
  Savings:      71.4%

Also included: `CLAUDE_CODE_DISABLE_1M_CONTEXT=1` context cap as deploy default, `/compact` reminder at call 25, and opt-in `claude-haiku-4-5` routing for administrative agent tasks.

3. Multi-Repo Management — one command keeps 29+ projects in sync

npx claude-scaffold update --all   # sync all registered projects
npx claude-scaffold update --all --dry-run  # preview what changes

...