everything-claude-code

everything-claude-code (ECC) is a repository on GitHub that describes itself as an “agent harness performance optimization system”. It was published as open source by Affaan Mustafa, whose project zenith.chat — built, according to a February 2026 Korean guide, entirely with Claude Code — won an Anthropic x Forum Ventures hackathon in September 2025; the guide says the repository collects settings refined over more than ten months of real use, and likens it to an operating system for Claude Code rather than a collection of configuration files. A Qiita article on Harness Engineering, ハーネスエンジニアリング入門 ── CLAUDE.mdの次に来るAIエージェント制御パラダイム, calls it the culmination of harness design and reads it as more than a collection of CLAUDE.md templates: an executable system that covers all five components the article assigns to a harness — rules, skills, hooks, memory and a feedback loop.

Capabilities

The article maps the repository’s layout onto those components. Rules live in CLAUDE.md, an AGENTS.md defining the roles of 28 agents, and a rules/ directory with language-specific constraints (TypeScript, Python, Rust and a common set). Skills are a skills/ directory of more than 125 workflow definitions as of 23 March 2026 — among them TDD, deep research, security review and continuous learning — alongside an agents/ directory of 28 specialist agent definitions such as a planner, a code reviewer and a TDD guide. Hooks are defined in hooks/hooks.json (more than 25 event triggers), memory in a contexts/ directory of session-persistent contexts for development, research and review, and the feedback loop in a tests/ verification suite.

The Korean guide, from February 2026, describes the same repository from its plugin install and its two companion guides. Its organising idea is to use Claude not as a coding tool but as a virtual development team: the main session acts as project manager and delegates detailed work to role-specific sub-agents. At the time of that post the agents/ directory held 12 agents (planner, architect, TDD guide, code reviewer, security reviewer, build-error resolver, an E2E runner using Playwright, refactor cleaner, documentation updater, and a Go code reviewer and Go build resolver). Each agent is defined with YAML frontmatter naming its tools and model, and the stated design principle is to keep each agent’s tools to a minimum to sharpen its focus and avoid role conflicts between agents — the planner, for example, gets only Read, Grep and Glob. Slash commands such as /plan, /tdd, /code-review, /build-fix, /e2e and /learn (which extracts patterns from a session and saves them as a skill) drive the development routine; everything under rules/ is inserted into Claude Code’s system prompt and always applies; and the hooks the guide shows block a console.log left in an edited TS/JS file, block the creation of unnecessary .md/.txt files other than README and CLAUDE.md, and run Prettier after edits. The author’s Shortform Guide covers setup, basics and philosophy; the Longform Guide covers token optimisation, memory persistence across sessions, continuous learning, verification loops and parallelisation with Git worktrees.

Three design patterns are singled out by the Qiita article:

• Layered hooks. Before a tool runs, hooks block bypassing git hooks with --no-verify, block changes to linter configuration (preventing rules from being weakened), run security monitoring and check MCP server health; after a tool runs, they format, type-check, warn about leftover console.log calls and run quality-gate checks; when a task stops, they persist session state, extract learned patterns and record token cost. What the article finds distinctive is hooks that block an agent’s attempts to get around the rules, rather than only formatting after edits.
• Self-optimization. A harness-optimizer agent analyses the harness’s own configuration and proposes improvements, and a continuous-learning skill extracts patterns from each session to keep updating the harness’s knowledge base.
• Graduated strength. Hooks run in one of three modes — minimal, standard and strict — so enforcement can be matched to a project’s maturity rather than every hook being active all the time.

Adoption & Ecosystem

ECC is built around Claude Code, and the Qiita article treats it as an external plugin. The Korean guide installs it through Claude Code’s plugin marketplace, and notes that the rules are not deployed by the plugin and have to be copied into the user’s rules directory by hand. That guide also passes on a warning against turning on every MCP server at once — a 200K-token context window can shrink to 70K — with the guideline of 20–30 configured MCP servers, no more than 10 enabled per project and no more than 80 active tools, and the advice to wrap CLIs as skills in place of MCP servers. It recommends ECC where production-level quality matters, for its enforced TDD, code-review guardrails and security rules, names a steep learning curve as its limitation, and compares it with Oh My ClaudeCode, which orchestrates agents automatically where ECC leaves the judgment to the user. The Qiita article cautions against adopting a complete setup of this kind on the first day — a flood of hooks slows the agent and managing the configuration crowds out development — and recommends adding components as problems appear. It expects functions that external plugins such as ECC currently provide to be absorbed into agent tools themselves, and points to the repository’s popularity as a sign that good harness design is becoming shared community knowledge.