use-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.