Three workflows sit between “task created” and “agent starts coding”: architect review, research, and the prep-task wrapper that runs them in order with a gate in between.

What they are

• architect-task — reviews the task spec for soundness: dependencies, risks, ADR candidates, missing context. Stops the line when the spec itself is broken.
• research-task — gathers patterns, files, symbols, and notes the implementer will need. No code, just findings folded back into the task details.
• prep-task — runs architect first, then research, with a gate. If architect flags blockers, research does not run until the spec is fixed.

When to use them

• The task is non-trivial and you want to catch spec problems before burning time.
• The task touches code you do not know well — research surfaces the patterns to follow.
• You want a single command that does both gated end-to-end: use prep-task.

When to skip them

• The task is small and obvious.
• You already know the patterns and the spec is tight.
• You are mid-implementation — these run before code, not during.

How to invoke

Trigger phrases (per the project’s CLAUDE.md):

• Architecture review: or Architect task: → architect-task
• Research task: → research-task
• Prep task: → prep-task

The agent reads the corresponding workflow file and follows it. None of these write implementation code.

Reference

• architect-task.md
• research-task.md
• prep-task.md