My global Claude Code configuration — the working style, delegation system, and model routing that load into every session.
CLAUDE.md # global user memory → ~/.claude/CLAUDE.md
agents/ # subagent definitions → ~/.claude/agents/
├── snr-code-engineer.md # Opus — semi-defined work needing judgment
├── jnr-code-engineer.md # Sonnet — fully-specified mechanical work
├── qa-runner.md # Sonnet — runs a verification matrix, reports pass/fail
└── web-runner.md # Haiku — cheap web fetch / extract grunt work
A capability-tiered delegation system. The session lead (Fable 5, falling back to Opus if Fable is unavailable) scopes, designs, and reviews. Execution is delegated by what's missing from the task:
- decisions named and small, clean diff → jnr (cheap, mechanical)
- judgment needed, or a large/messy diff → snr (stronger model)
- choosing the approach is the work — design, architecture, security → stays on the lead
Verification is delegated the same way (qa-runner), and the whole thing is tuned for
context-cost discipline: keep the lead's context lean, cap what subagents return.
See CLAUDE.md for the full routing rules.
On a fresh machine:
git clone https://github.com/willmot/dotclaude ~/Developer/dotclaude
cp ~/Developer/dotclaude/CLAUDE.md ~/.claude/CLAUDE.md
cp -r ~/Developer/dotclaude/agents ~/.claude/Prefer symlinks if you want this repo to stay the source of truth (back up any existing
~/.claude/CLAUDE.md and ~/.claude/agents/ first). Restart Claude Code, then run
/agents to confirm the four agents registered.