SuperAPI AI Directory

subagent-driven-development-conventions

Coding & development

View repository

Updated 10-22 16:17

About this skill

Subagent-Driven Development Conventions

Overview

REQUIRED SUB-SKILL: Invoke superpowers:subagent-driven-development — it owns the actual delegation loop. This skill is a required layer on top: what context to load before delegating, what goes in a subagent's prompt, model choice, and review cadence.

Rules

  • Native Plan Mode is not required for this workflow — go straight to reading the plan and delegating tasks. Don't enable Plan Mode solely to use this workflow, unless the user or another skill separately requires it.
  • Before delegating, read the project's own durable context — its memory file, a fast codebase index (e.g. via codebase-domain-map), the target subagent's role definition, and the rules relevant to the task's scope. The dispatching agent does this itself; a subagent doesn't read these on its own.
  • Pass a full context packet explicitly in each subagent's prompt: goal, scope, interfaces, applicable constraints (including which stack-convention skills to invoke — a subagent doesn't discover skills from a bare task description), the chosen workflow mode, verification policy, and readiness criteria.
  • Default to a mini/cheap model for implementation and review subagents, unless the user specifies a different one.
  • Implementer subagents never commit — explicitly forbid git commit in the prompt; changes stay uncommitted (see repo-workflow-conventions).
  • Route test-writing to the test-writing role, not the implementer — an implementer subagent does not write tests unless the user explicitly asks it to (see testing-philosophy).
  • One final whole-branch review, not per-task. After all implementation tasks are done, run a single final review. Don't run a task-by-task review unless there's an explicit risk, blocker, or user request.

Common Mistakes

Mistake Fix
Entering Plan Mode "just in case" before this workflow Skip it — go straight to delegation
Dispatching a subagent with only "implement X" and no explicit constraints/skills list Build the full context packet first — the subagent has no other way to see it
Using a large/expensive model for a mechanical implementation task Default to mini/cheap unless the user asked for something else
An implementer subagent runs git commit after finishing its task Forbid it explicitly in the prompt
Reviewing after every single task Batch into one final whole-branch review