| name | use-claude |
| description | This skill should be used when a non-Claude agent (e.g. Codex) is considering delegating a coding task to Claude Code's headless mode (claude -p) — hard long-horizon bug fixing, implementation where first-pass completeness matters more than token cost, test authoring, or a cross-model-family second opinion on a plan or diff. Claude Code spends an Anthropic subscription/API budget instead of the caller's tokens. Invoked for "use claude", "ask claude", "delegate to claude", "claude -p", or "second opinion from claude". |
Use Claude
If you are Claude, you do not need this skill — stop reading and proceed however you otherwise would. Delegating to yourself gains nothing; this skill exists for other agents delegating scoped work to Claude Code's headless mode (claude -p).
Should this go to Claude?
Delegate when the task is scoped, hard, and correctness matters more than token cost:
- Implementation where first-pass completeness counts — Claude spends more tokens but blind reviewers rate its output more complete
- Long-horizon debugging in real repos (multi-file root causes, tangled state)
- Test authoring and verification-heavy work
- Cross-family second opinion / adversarial review of a plan or diff (a different model family catches different bugs)
- Repo work that benefits from project context — Claude Code auto-loads the repo's CLAUDE.md, memory, and skills
A headless one-shot cannot interview the user, so ambiguous or underspecified work, open architecture decisions, and anything needing a long conversation are not delegation targets — they belong to whoever holds the user conversation. Reduce unknowns first, then delegate the well-specified remainder. Also skip delegation for cheap mechanical sweeps if the caller is the more token-efficient model — Claude's edge is quality, not thrift.
Full rationale, evidence, and cost notes: references/task-fit.md. If the decision is "don't delegate", stop here — no need to read further.
Quick start
claude -p --output-format json --model sonnet \
--tools="Read,Grep,Glob" \
"<self-contained prompt with all context Claude needs>" > /tmp/claude-out.json
Headless mode emits nothing until done, then a single JSON object: .result is the final message (intermediate activity is never printed, so tokens aren't double-spent), .session_id enables follow-ups, .total_cost_usd reports spend.
- Models:
sonnet (workhorse), haiku (quick lookups), opus / fable (hardest work); --effort low|medium|high|xhigh|max.
- Write access: default denies edits; add
--permission-mode acceptEdits and/or --allowedTools when Claude should change files (then review the diff).
- Resume:
claude -p --resume <session-id> "follow-up" keeps full context.
- Gotcha: variadic flags like
--tools swallow a following prompt — use = syntax (--tools="Read,Grep").
Full flag reference, permission ladder, session resume, piping, and failure modes: references/invocation.md.
After it returns
Treat the output as a peer's work, not an authority's: verify claims against the code, run tests on any diff, and attribute Claude's contribution when reporting.
