Claude Code Config

Production-ready Claude Code configuration for rapid app development.

12 agents · 16 commands · 18 skills · curated rules · token-saving hooks.

Deliberately lean: skill descriptions, agent definitions, and MCP schemas all consume context in every session. This config stays well under the ~10% startup-overhead budget so auto-invocation keeps working and subagents don't blow their context at spawn.

Built on Everything Claude Code (ECC) + obra/superpowers.

---

Install

One-liner (clones to ~/.local/share/claude-code-config, installs the default zero-dep set):

curl -fsSL https://raw.githubusercontent.com/Cyvid7-Darus10/claude-code-config/main/setup.sh | bash

Or pass flags through:

curl -fsSL https://raw.githubusercontent.com/Cyvid7-Darus10/claude-code-config/main/setup.sh | bash -s -- --minimal
curl -fsSL https://raw.githubusercontent.com/Cyvid7-Darus10/claude-code-config/main/setup.sh | bash -s -- --full
curl -fsSL https://raw.githubusercontent.com/Cyvid7-Darus10/claude-code-config/main/setup.sh | bash -s -- --dry-run

Manual clone (if you'd rather inspect before running):

git clone https://github.com/Cyvid7-Darus10/claude-code-config.git
cd claude-code-config
./install.sh

Restart Claude Code. Try /plan, /tdd, /verify, or /code-review.

Minimal install (agents, commands, skills, rules only — no hooks):

./install.sh --minimal

Full install (adds Node.js-based quality-gate hooks — requires Node 18+):

./install.sh --full

Selective install / uninstall:

./install.sh agents skills              # Install only these two components
./install.sh --dry-run                  # Preview without touching disk
./install.sh --uninstall                # Remove everything the installer placed
./install.sh --uninstall skills         # Remove just one component

Available components: agents, commands, skills, rules, mcp, sounds, security (docs), hooks (token-saving Node hooks, opt-in).

Updating

Re-run the one-liner at any time — setup.sh fetches the latest commit, hard-resets to origin/main, and re-runs install.sh:

curl -fsSL https://raw.githubusercontent.com/Cyvid7-Darus10/claude-code-config/main/setup.sh | bash

What happens to each file in ~/.claude/:

Path	Behaviour on update
settings.json	Never overwritten — your customisations are safe. Diff against the repo's settings.json to see new defaults worth merging.
mcp.json	Never overwritten — protects your tokens.
agents/, commands/, skills/, rules/	Replaced with latest. Previous copy backed up to ~/.claude/backups/pre-install-<timestamp>/ first.
~/.local/share/claude-code-config (the clone)	Fast-forwarded to origin/main.

See what changed before updating:

cd ~/.local/share/claude-code-config && git fetch && git log --oneline HEAD..origin/main

Pin to a specific version (useful for reproducible team setups):

CLAUDE_CODE_CONFIG_REF=v1.0.0 curl -fsSL https://raw.githubusercontent.com/Cyvid7-Darus10/claude-code-config/main/setup.sh | bash

Roll back to a previous install:

ls ~/.claude/backups/                    # find the timestamp you want
cp -r ~/.claude/backups/pre-install-<timestamp>/* ~/.claude/

Merge new defaults into your existing settings.json — since the installer won't overwrite it:

diff ~/.claude/settings.json ~/.local/share/claude-code-config/settings.json

Copy over the hook entries or permission rules you want; leave your custom fields alone.

Requirements

	Default install	--full install
bash 3.2+	Required	Required
git	Only for git clone / updates	Only for git clone / updates
Node.js 18+	Not needed	Required (for hooks/ quality gates)
jq / Python	Never	Never

The default install is zero runtime dependencies. The Node-based token-saving hooks in hooks/ (duplicate-read blocking, output compression) are opt-in.

---

What You Get

Default install (zero runtime dependencies):

Component	Count	Examples
Agents	12	planner, architect, code-reviewer, security-reviewer, tdd-guide, build-error-resolver, typescript-reviewer, python-reviewer, database-reviewer, docs-lookup, e2e-runner, refactor-cleaner
Slash commands	16	/plan, /tdd, /verify, /code-review, /build-fix, /e2e, /test-coverage, /refactor-clean, /python-review, /docs, /ponytail + 5 ponytail subcommands
Skills	18	api-design, backend/frontend-patterns, postgres-patterns, database-migrations, security-review, e2e-testing, market-research, ponytail, python + swift/SwiftUI/Foundation Models packs
Coding rules	curated	Common (coding style, git workflow, testing, security, multi-repo-consistency) + TypeScript, Python, Swift
MCP servers	2	context7 (docs, used by docs-lookup), playwright (used by e2e-runner) — GitHub via gh CLI, not MCP

Opt-in with --full (requires Node.js 18+):

Component	Count	Examples
Token-saving hooks	2	pre-read.mjs blocks duplicate file reads · compress.mjs compresses noisy tool output and detects command loops

---

Workflows

• Idea → plan → TDD → ship — /plan uses the brainstorming skill, /tdd runs RED-GREEN-REFACTOR.
• Intelligent routing — the task-router hook analyses every prompt and surfaces the right agent or skill as a hint. Routes include: plan-first, architecture, debug, tdd, security, review, refactor, build-fix, docs, performance, backend (endpoint/migration/queue work → backend-judgement checklist), reasoning (ambiguous/cross-cutting prompts → systematic-reasoning).
• Session continuity — session-restore on start, pre-compact-checkpoint before compaction, session-persist on stop. Pick up where you left off across restarts.
• Deliberate thinking — the systematic-reasoning skill forces a four-pass model-building process before non-trivial changes. The backend-judgement skill applies a 9-check production-safety list to every endpoint, migration, queue consumer, and job.
• Multi-repo consistency — rules/common/multi-repo-consistency.md codifies cross-stack invariants (naming, error handling, logging, boundaries, testing) so Claude behaves the same across your Go / TS / Python / Rust / Java / Kotlin / Swift / C++ / PHP / C# / Perl repos.

---

MCP Servers

Edit ~/.claude/mcp.json to enable.

Server	Purpose	Setup
context7	Live library/API docs	None — zero config
playwright	Browser automation	None — zero config
magic	Magic UI components	None — zero config
github	Manage repos/PRs	GitHub token + Docker

GitHub setup:

gh auth token    # paste into ~/.claude/mcp.json, replace <YOUR_GITHUB_TOKEN>

---

Security

Deny-list defaults protect credentials and destructive commands — see settings.json. Permissions block:

• Reads of ~/.ssh, ~/.aws, .env, credentials*, secrets*
• rm -rf /, rm -rf ~, rm -rf /*
• curl|bash, wget|bash, curl -d exfiltration patterns
• git remote add, npm publish, nc/netcat

See security/SECURITY.md for the full threat model.

---

Customisation

Add a command

Create ~/.claude/commands/my-command.md:

---
description: What my command does
---
Instructions for Claude when this command is invoked.

Add a skill

Create ~/.claude/skills/my-skill/SKILL.md:

---
name: my-skill
description: When to activate this skill
---
Domain knowledge and instructions.

Add an agent

Create ~/.claude/agents/my-agent.md:

---
name: my-agent
description: What this agent specializes in
tools: ["Read", "Grep", "Glob", "Bash"]
model: sonnet
---
System prompt for the agent.

Add / remove lifecycle hooks

Edit ~/.claude/settings.json under hooks.{SessionStart,UserPromptSubmit,PreToolUse,PostToolUse,PreCompact,Stop,Notification}.

The default config avoids PreToolUse and PostToolUse because they fire on every tool call and can add several seconds of latency per turn. If you add them, keep them small, set timeout low, and prefer async: true.

---

Recommended companion plugins

After installing this config, layer these official marketplaces on top — they're curated, actively maintained, and complement our agents without duplicating them:

# In Claude Code:
/plugin install superpowers@claude-plugins-official
/plugin marketplace add trailofbits/skills
/plugin marketplace add wshobson/agents

Plugin	Why add it	Source
superpowers	TDD methodology, brainstorming, systematic-debugging — officially accepted into Anthropic marketplace Jan 2026	obra/superpowers
Trail of Bits skills	Pro-grade security auditing: static analysis, variant analysis, differential review, supply-chain risk, constant-time analysis	trailofbits/skills
wshobson/agents	184 agents across 25 categories — cherry-pick conductor (track management), comprehensive-review (multi-perspective analysis), plugin-eval (anti-pattern detection)	wshobson/agents

---

Troubleshooting

• Hooks not running — restart Claude Code (it reloads settings.json only on start).
• settings.json didn't change after re-install — expected: the installer never overwrites an existing settings.json. Delete or rename your copy, then re-run, or hand-merge from this repo.
• Node hooks failing — you installed --full without Node 18+. Either install Node or re-run with ./install.sh (default omits Node hooks).
• Skill descriptions being dropped — run /doctor in Claude Code; if skills are shortened or dropped, you've added too many. Run /context to see startup overhead.

---

Project Layout

claude-code-config/
├── .claude-plugin/
│   └── marketplace.json   # Plugin marketplace entry (Anthropic convention)
├── agents/                # 12 specialized subagents
├── commands/              # 10 slash commands
├── skills/                # 17 workflow skills
├── rules/                 # Curated coding rules (common + python/swift/typescript)
├── hooks/                 # Opt-in Node token-saving hooks
├── scripts/hooks/         # Scripts invoked by the opt-in hooks
├── security/              # Security framework docs
├── mcp-configs/           # Reference MCP server configs
├── sounds/                # Optional macOS notification sounds
├── images/                # README demo GIFs
├── docs/                  # Developer docs (PLUGIN_SCHEMA_NOTES.md, tapes/)
├── install.sh             # Zero-dep installer (--minimal / --full / --uninstall / --dry-run)
├── settings.json          # Starter settings (permissions + token-saving hooks)
├── mcp.json               # MCP server config (add your tokens)
├── plugin.json            # Claude Code plugin manifest
└── AGENTS.md              # Full agent reference

---

Credits

• Everything Claude Code (ECC) — Agents, commands, rules, hooks, scripts, security guide.
• superpowers — Brainstorming, planning, TDD skills.
• claude-code-hooks-multi-agent-observability — Monitoring patterns.
• ruflo — Task-router and session-persistence patterns.

License

MIT — see LICENSE.

---

Made by Cyrus David Pastelero — if this helped you, consider starring.