Claude Code Kit

Personal Claude Code setup used by [email protected]. Use at your own risk.

A plugin marketplace with eight skills, plus the permission ruleset and global instructions I actually work with daily.

Install

Add the marketplace once:

/plugin marketplace add padak/claude-code-kit

Then install only the plugins you want:

/plugin install trip-master-plan@claude-code-kit

Each skill is its own plugin, so nothing arrives that you did not ask for. Browse and install interactively with /plugin, and update everything later with /plugin marketplace update claude-code-kit.

Plugins

trip-master-plan is the only plugin with runtime dependencies: python3 (stdlib only) and curl. On first use it downloads ~85 MB of Natural Earth GeoJSON into a geo/ directory in your working folder; the geodata is not vendored here.

Subagents

trip-master-plan ships trip-fact-checker — a read-only adversarial verifier (Read, Grep, WebSearch, WebFetch) spawned one instance per section, in parallel. It reports only problems, each with an exact quote, what is wrong, a drop-in correction and a source URL. Running it as a subagent rather than inline keeps a long audit out of the main session’s context.

Configuration

Plugins cannot install user-level configuration, so .claude/settings.json, CLAUDE.md and .claude/rules/ are copied by hand:

git clone https://github.com/padak/claude-code-kit.git
cd claude-code-kit
cp .claude/settings.json ~/.claude/settings.json
cp CLAUDE.md ~/.claude/CLAUDE.md
mkdir -p ~/.claude/rules && cp .claude/rules/*.md ~/.claude/rules/

Read them before overwriting yours — they are opinionated.

Why this settings.json

Permissions are split by how reversible an action is, not by how dangerous it sounds:

• allow — things that cannot break anything: git status, ls, grep, reading files, running tests, python, build tools. Auto-approving these is where the time saving actually comes from; they are also the calls Claude makes most often.
• ask — things that destroy work or are hard to undo: rm, git push --force, git reset --hard, git clean, package installs, docker, kubectl, and edits to lockfiles and package.json. Not blocked, just deliberate.
• deny — reading or writing secrets: .env files, SSH keys, .aws/credentials, certificates, keystores, anything matching *token*, *password*, *credentials*. This is the rule set worth copying even if you take nothing else. It is not about trust; it removes an entire class of accident where a secret gets pulled into the transcript and from there into a commit, a bug report or a PR description.

Two hooks:

• PostToolUse runs py_compile on edited Python and tsc --noEmit on edited TypeScript. Catching a syntax error at the moment of the edit beats discovering it ten steps later, when the fix means unwinding everything in between.
• Notification fires a macOS notification when Claude needs you — worth it once tasks run long enough that you switch windows.

Why this CLAUDE.md

CLAUDE.md is loaded into every session, so it holds only what applies everywhere and stays under 50 lines. Language-specific rules live in .claude/rules/ with a paths: front matter, so the Python rules load only when Claude touches a .py file and cost nothing otherwise.

What the global file does:

• Sets the audience. I am a product manager. Claude reports outcomes and user impact — what is now, what changes, what it means — not implementation detail, and shows code only when a decision needs it. Output is short because I read results, not process.
• KISS. The simplest thing that solves the problem; no speculative architecture. If complexity looks necessary, Claude has to argue for it first.
• Splits thinking from doing. The main session (Fable) plans, decides and reviews; execution goes to Sonnet/Opus sub-agents, in parallel when tasks are independent, never for work that is faster inline.
• No mocks, no stubs, no TODO: implement later. If something is in the plan it gets built, or Claude asks. Left unstated, a blocked step quietly becomes a fake one that passes tests and fails in production.
• No hardcoded values, no silent defaults. Application config lives in config files; a missing required variable fails loudly at startup instead of falling back to an invented value. A wrong default is much harder to debug than a crash.
• Research before implementing unfamiliar tech, via Perplexity MCP, with explicit cost tiers so the cheap tool is the default and the expensive one is deliberate.
• Czech in conversation, English in files. Separating the language you think in from the language the artifact ships in keeps the codebase readable to everyone else.
• Clean, safe commits — no Co-Authored-By, no generated-with footers, no secrets in any tracked file.

.claude/rules/python.md adds the Python conventions: uv for environments and dependencies, httpx, pathlib, pytest, python-dotenv, specific exceptions.

.zshrc

alias cc="claude --allow-dangerously-skip-permissions --chrome"

I run Claude as cc in full YOLO mode — all permissions bypassed, Chrome MCP included. That is how I work daily; if you want the guardrails above, use claude directly. Also included: auto-activation of .venv if present, and a gtimeout alias for macOS.

Customization

• Personal overrides go in ~/.claude/settings.local.json, which git ignores
• Language — set "language": "Czech" (or yours) in settings.local.json
• Add a skill to a plugin — create a directory with a SKILL.md under plugins//skills/
• Add a subagent — drop a Markdown file with name / description / tools front matter into plugins//agents/
• Add a new plugin — create plugins//.claude-plugin/plugin.json and list it in .claude-plugin/marketplace.json
• Develop against a local checkout — /plugin marketplace add ./claude-code-kit points at your working copy instead of GitHub

License

MIT