How I use Claude.

Second head

Claude Code is my daily driver, and the contract is written down: it acts as a second head, not a service provider working through tickets. The rule that carries the whole thing: if the answer is in the code, git, logs, or docs, the agent finds it itself — it never asks me something it can verify. If the answer is in my head — intent, scope, the "why is it like this" — it asks early, with its own reading attached. Decisions it makes; missing information it asks for.

The safety layer is inverted

I run with harness confirmations off. No "allow this command?" prompts — which means the agent itself is the only safety layer. So the policy doesn't live in the agent's goodwill: destructive actions require fresh, explicit confirmation every time, scoped to exactly what was asked. Writes to external systems — Slack, issue trackers, anything other people see — are treated as destructive even when nothing is deleted. And whatever can be enforced by config is enforced by config: a declarative deny list in settings, so the harness blocks what memory might forget.

Aliases

Model presets I actually type:

alias sonnet='claude --model "claude-sonnet-5[1m]" --effort xhigh --dangerously-skip-permissions'
alias opus='claude --model "claude-opus-5[1m]" --effort xhigh --dangerously-skip-permissions'
alias fable='claude --model fable --dangerously-skip-permissions --effort xhigh'

# other models through the same harness
alias clamma='OLLAMA_HOST="<homelab>" ollama launch claude'
alias oclaud='ollama launch claude --model glm-5.2:cloud -- --dangerously-skip-permissions'

One config, three machines

The whole setup — global rules, settings, skills, code-style rules — lives in a dotfiles repo; ~/.claude is a symlink into it. Personal Mac, work Mac, and a cloud devbox all run the same brain. Transcripts and history never get committed: they may contain secrets.

Beyond Claude

The same workflow runs other models for comparison: OMP (a Pi-based harness) with Kimi K3, and cloud models through ollama. That habit turned into a real benchmark — the same spec executed independently by four models, and a store now runs the winner. That story is on Arte à Mesa.