| name | delegate-coding |
| description | For coding/implementation tasks, use your own judgement to delegate the actual code-writing to a subagent running an appropriately lower-power model, to conserve top-tier tokens. Triggers whenever a task is primarily writing or editing code (features, fixes, refactors, mechanical edits). Skip for judgement-heavy work (design, debugging, auditing, synthesis). |
Delegate coding to a lower-power subagent
When a task is primarily writing or editing code, don't burn top-tier model tokens
on the mechanical work. Use your own judgement to pick an appropriately lower-power model
and run the implementation in a subagent. Keep judgement, review, and synthesis in the
main loop.
Why: Implementation work rarely needs the top-tier model. Cost/efficiency — the
expensive model's judgement is better spent on deciding what to build and verifying the
result than on typing it out.
How to apply
-
Judge the task first. Ask: is this mostly producing/editing code, or is it
judgement-heavy?
- Delegate — feature implementation, bug fixes with a clear cause, refactors,
boilerplate, test writing, mechanical/repetitive edits.
- Keep in the main loop — architecture/design, ambiguous debugging, security or
code review, data synthesis, anything where the hard part is deciding not typing.
-
Pick the model by weight (your judgement, not a fixed rule):
sonnet — substantive implementation that still needs real coding ability.haiku — trivial, mechanical, or highly-specified edits.
-
Spawn an Agent with the model override and a self-contained prompt: include the
files, the exact change, constraints, and how to verify. The subagent doesn't share your
context, so give it everything it needs.
Agent(subagent_type: "general-purpose", model: "sonnet",
description: "...", prompt: "<self-contained task + files + verify steps>")
-
Review in the main loop. Read the subagent's result, sanity-check it, run/verify as
needed, and only then commit. You own correctness, not the subagent.
When to skip
If the task is small enough that delegating (writing a self-contained prompt, spawning,
reviewing) costs more than just doing it, do it inline. Judgement over ceremony.
