token-efficient-agent

Token-efficient agent: apply all patterns

This is the orchestrator skill. It does not add new rules — it activates the three underlying skills together so you don't have to remember to invoke them separately.

Apply on every non-trivial task

At task start:

1. [[scratch-context]] — check .agent-scratch/notes.md, decisions.md, dead-ends.md before reading any code.
2. [[tier-model]] — decide up front which model handles each piece of work. Default: Haiku for lookup/iteration, Sonnet for judgment, Opus only after Sonnet stalls twice.
3. [[delegate-work]] — spawn subagents for search, grep, file reads >2, test/build runs, and fix loops. Enforce the return contract on every spawn.

During work:

• Append non-obvious findings to .agent-scratch/notes.md as you go.
• Record every ruled-out hypothesis in .agent-scratch/dead-ends.md immediately.

At task end:

• Write one-line decisions with rationale to .agent-scratch/decisions.md.
• If the user starts an unrelated new task, suggest /compact or a fresh session.

When to invoke this skill

• Any task that will touch more than two files
• Any debugging session where the root cause is unknown
• Any test/build/lint fix loop
• Any codebase search or "find where X is defined/used" question
• Any session you expect to run longer than ~10 turns

When NOT to invoke

• One-liner edits to a file already in context
• Direct questions answerable from the current conversation
• Tasks where context growth is clearly not a risk

See also

• [[delegate-work]] — subagent patterns and return contract
• [[tier-model]] — model selection decision table
• [[scratch-context]] — externalize state to .agent-scratch/