token-efficiency

1. Subagent routing

Route by task type, not convenience:

• Haiku — read-only search, fan-out grep, listing, counting. Never analysis.
• Sonnet — edits, tests, refactors, generation with context.
• Opus — architecture decisions, complex debugging, multi-file synthesis.

Never let subagents silently inherit the top-thread model. Pin each agent explicitly.

Detail: reference/subagent-routing.md

---

2. Cache hygiene

Prompt caching is automatic; your job is not to break it:

• Don't /clear mid-task — it voids the cache.
• Keep turns <5 min apart; idle gaps expire TTL.
• Append context; don't rewrite early turns.
• Keep tool-call order stable across turns.

Detail: reference/cache-hygiene.md

---

3. Narrow reads

Grep before reading. Read only the hit range with offset/limit. Never re-read a file after an Edit — the harness tracks state.

---

4. /clear discipline

Clear only between unrelated tasks. Every token in context is re-billed on every turn. Old task context = dead weight.

---

5. Big-skill signal

SKILL.md body loads into context on every invocation. Keep it thin.

• Rule of thumb: >20 KB → split into thin core + reference files.
• Reference files are read on demand, not on invocation.

Detail: reference/big-skill-refactor.md

---

6. Skill-listing trim

Reduce the per-turn skill-listing cost:

• skillOverrides: set unused skills to name-only or off.
• skillListingBudgetFraction: cap total listing size.