Claude

Claude

Claude Code is an independent agent on PATH (claude), sharing this working tree and already authenticated. Use it as a second opinion or delegated worker, not as ground truth: verify its claims, own any changes, and keep its task contract narrow.

Defaults

Use claude -p --model opus --effort high unless the user or task calls for a different model or thinking level.

• Model: accept aliases such as opus, sonnet, or fable, or a full model name. Default opus.
• Thinking level: pass Claude Code's --effort flag. Valid levels are low, medium, high, xhigh, and max. Default high.
• Cost guard: for exploratory asks, add --max-budget-usd <amount> when the user gives a budget or the task is likely to sprawl.

Example:

claude -p --model opus --effort high "<prompt>"

Prompting Claude

Prompt Claude like an operator: compact, block-structured, and explicit about the artifact you need. A sharper contract beats higher effort.

• One task per run. Split review, implementation, and docs into separate prompts.
• Name repo skills when relevant. Claude can read repo files; say "follow write-docs for the doc" or "obey refactor-clean: no compatibility wrappers" instead of pasting those skills into the prompt.
• Blocks, added only where useful:
  • <task> — the concrete job, context, and expected end state.
  • <output_contract> — exact response shape, highest-value first.
  • <default_follow_through> — take low-risk interpretations and continue; stop only when a missing detail changes correctness, safety, or an irreversible action.
  • <verification_loop> — inspect the result against requirements and revise before finalizing.
  • <grounding> — ground every claim in code or tool output; label inferences.
  • <action_safety> — keep the diff tightly scoped; no drive-by refactors.
• Anti-patterns: vague prompts, mixed jobs, no output contract, asking it to "think harder" instead of tightening the task, and accepting its answer without checking the evidence.

Consultation

Use Claude for consultation when the user asks for Claude's view or when a second-agent read would materially reduce risk. Keep it read-oriented unless the user explicitly asks it to edit.

1. Ask for evidence, not vibes: require file paths, observed behavior, command output, or labelled inferences.
2. Prefer a read-only tool surface for reviews: --tools "Read,Grep,Glob,Bash" --permission-mode dontAsk.
3. Triage every finding against the code before acting. Report what Claude flagged, what you accepted, and what you dismissed.

Implementation — explicit ask only

Delegate implementation to Claude only when the user names Claude for the task. Never hand it work on your own initiative, and never re-delegate follow-up work without a fresh ask.

1. Slice the task sharply: goal, constraints, allowed scope, and verification. An underspecified task stays with you until a fresh agent could not misread it.
2. Start from a clean tree or record the baseline commit so Claude's diff is separable from yours. Tell Claude to keep scratch files (probes, dumps, notes) out of the tree — its diff must contain only the deliverable.
3. Choose permissions by what the task must do:
   • Read/review: --tools "Read,Grep,Glob,Bash" --permission-mode dontAsk.
   • Narrow edits: add edit tools and keep the prompt's allowed paths explicit.
   • Broad autonomous work: use --dangerously-skip-permissions only in a dedicated worktree, with a self-authored prompt and a mandatory diff review after.
4. Capture the response with --output-format text by default. Use --output-format json only when the caller needs structured metadata.
5. You own the result: inspect the diff, run the relevant tests, and only then report it. "Claude says it's done" is not done.

Liveness

claude -p is a networked, authenticated non-interactive run. If it fails, classify the failure before retrying.

• Auth failure: do not repair credentials unless the user asks; report the blocker. If sandboxed claude -p says Not logged in, retry once with the required approval because local auth can depend on keychain/session access.
• Sandbox/network failure: rerun with the required approval if the command is important to the task.
• Long run: prefer --max-budget-usd or a narrower prompt before raising effort.
• Workspace trust: -p skips the interactive trust dialog, so run it only from a repo or worktree you intentionally trust.

Rules

• Use claude -p --model opus --effort high as the default invocation.
• Let the user override model and effort in plain language; translate that to --model and --effort.
• Keep Claude's writes out of the working tree unless delegation was explicit.
• Do not touch the same working tree while Claude is making edits.
• Treat Claude output as evidence to verify, not authority to relay.