Drive Development with Subagents

superpowers core skill · subagent-driven-development

In one line: Don't grind a long task to the end in one session. Make the main session the "dispatcher" — a fresh subagent per task, discarded when done.

Scenario: build a coupon system

A coupon system has several fairly independent pieces: create coupons (admin), users claim them, redeem at checkout, expire and reclaim. You already have an implementation plan. Write it all in one session, and by the time you hit "redeem," the AI's head is still stuffed with "create coupon" details — things blur and quality drops.

How to ask the AI

Use subagent-driven-development to execute this coupon plan: for each task (create/claim/redeem/expire),
dispatch a fresh subagent, then a spec review (does it implement the requirement exactly), then a code-quality review — both pass before the next.

Answer the subagent's questions seriously. When it asks "what if the same coupon is redeemed concurrently," that pre-work question is far cheaper than discovering over-redemption later.

How it'll walk you through it

The main session reads the plan once and feeds the "create coupon" task's full text to a fresh subagent (it doesn't read the whole plan — cleanest context). The subagent implements -> tests -> self-reviews -> commits, then passes spec review (only created coupons, nothing extra) and quality review (structure/naming). Issues get fixed by the same subagent and re-reviewed; only a pass moves to the next task, "claim" — again a fresh, clean subagent.

Remember one line

subagent-driven-development = main session dispatches, a clean subagent per independent task + spec-then-quality review. For multi-module features like coupons, do one piece and review one piece — no context bleed.