TASK — Direct Agents Through Scoped, Reviewable Work
Phase 3 of 4 in the AI Driven Development Methodology. The PLAN says what happens in what order. The TASK is how you actually hand one piece of that plan to an agent — with the context, constraints, and verification rules it needs.
What a TASK is in this methodology
A TASK is a single, scoped instruction to an AI agent — one responsibility per agent, per task. Not "build the feature." Not "fix the module." One reviewable unit of work, with enough context that the agent doesn't have to guess, and enough constraint that it can't quietly expand its own scope.
A working TASK prompt typically includes:
- Exactly which step of the PLAN this task covers — nothing more
- The specific files, modules, or interfaces the agent is allowed to touch
- Constraints carried over from the PRD (security, conventions, performance)
- How the agent — and you — will verify the task is actually done
Why scope discipline matters here. Agents are fast enough to generate large amounts of code in one pass. Without a tightly scoped TASK, "fast" turns into "hard to review" — and review is exactly where the human stays the architect.
What goes wrong without one
Broad, unscoped instructions are where multi-file, multi-agent work quietly breaks down: one agent changes a module's contract while another still assumes the old one, and nobody wrote down what the contract was supposed to be. Tight, single-responsibility tasks are what keep parallel or sequential agent work coherent.
How this looks in practice
In AAM's methodology, a TASK prompt is written per PLAN checkpoint — never for the whole PRD at once. Each task ends with an explicit review step before the next task begins, so architectural drift gets caught one small step at a time instead of at the end of a long agent session.
Start with the free TASK template
The Workflow Starter Kit includes a ready-to-use TASK-prompt-template.md — part of a
four-part kit (PRD, PLAN, TASK, IMPLEMENTATION) that works in any language or framework.