agentic-coding

My Claude Code configuration: global rules, skills, sub-agents, hooks and settings. Linked into ~/.claude via symlinks so the repo is the single source of truth.

Forked from sammcj/agentic-coding (Apache-2.0). Most skills and the hook code are Sam McLeod's work, trimmed and adapted. Upstream stays configured as upstream; changes are cherry-picked file by file, not merged.

Layout

• Rules/CLAUDE.md - global instructions, loaded in every session. Kept under 60 lines on purpose.
• Rules/CLAUDE-WEB.md - optional variant for claude.ai project instructions.
• Skills/ - agent skills, one directory per skill with a SKILL.md. Loaded on demand.
• Claude/agents/ - sub-agents (critical-reviewer, step-back, quick-researcher, software-research-assistant, compression-editor).
• Claude/commands/ - slash commands (/compact-prep).
• Claude/hooks/ - approve-compound-commands (Go, auto-approves piped and chained commands when every part is allow-listed) and allow-skills-edit.sh (lets Claude edit its own skills without prompting).
• Claude/output-styles/terse.md - optional output style, enable with /output-style terse.
• Claude/settings.json - permissions (allow, ask, deny), sandbox, hooks, plugins, status line.
• Claude/statusline-command.sh - context usage and rate-limit status line. Needs jq.

Install

./install.sh

Symlinks everything into ~/.claude, builds the Go hook and installs the pre-commit hook. Existing files are moved to .backup-before-install/. Requires jq, optionally go and pre-commit.

Warp scans ~/.claude/skills/ as well, so the skills are available there without extra setup.

Maintenance

• Regel ändern: Datei hier editieren, fertig. Die Symlinks zeigen direkt hierher.
• Claude Code schreibt selbst in settings.json (Plugin-Installationen, Permission-Freigaben). Das Repo wird dadurch dirty, einfach committen.
• Upstream sichten: git fetch upstream && git log --oneline main..upstream/main -- Skills Claude. Einzelne Skills holen mit git checkout upstream/main -- Skills/<name>.
• Neue Skills nur anlegen, wenn du dieselbe Erklärung zum zweiten Mal gibst. Vorher skill-creator-primer lesen.

Working rules I keep coming back to

Condensed from Sam's tips, kept because they match my experience:

• Only add rules for behaviour that differs from the default. Review them with the aim to reduce.
• Skills over rules: anything longer than ten lines that is needed only sometimes becomes a skill.
• Hooks over prompts for anything that must happen every time.
• Fresh sessions aggressively. When stuck, have the agent write a handoff and start over.
• No MCP server where a CLI does the job.
• After a hard-won fix, ask the agent what led it astray and turn the answer into a rule or skill.

License

Apache-2.0, see LICENSE. Original work by Sam McLeod.