Module
Module 4 of 5Lesson 3 of 3~14 min

Review an agent's plan against the spec

A coding agent often lays out a plan before writing a single line, and that is a valuable chance to check that it understood the spec. This lesson teaches you to hold that plan up against the spec to spot the gaps, then check what was actually delivered, without needing to be a developer.

Lesson objective

By the end of this lesson, you will be able to review the plan a coding agent proposes before writing code, spot the gaps with the spec (missing criterion, out-of-scope change, unstated assumption) and then check the evidence of what was delivered.

Topics covered

  • coding agent plan review
  • coding agents
  • plan mode
  • AI code review
  • spec vs implementation

Where it fits

Specs for coding agents

What do you need to write so an AI coding agent builds the right thing, and how do you check before and after?

Lessons in this module

  1. What a coding agent needs
  2. The intent file and repository conventions
  3. Review an agent's plan against the spec (this lesson)

What you will learn in the course

This lesson is part of the course Write specs that developers and AI agents understand

  • Choose the level of spec suited to the decision at hand (PRD, feature spec, ticket) and know what each must contain.
  • State the intent of a feature (problem, target outcome, constraints, non-goals) separately from the solution, to leave the right room to the team and the agent.
  • Write verifiable, declarative and unambiguous Given/When/Then acceptance criteria, including as tables of examples.
  • Identify edge cases and non-functional requirements (performance, security, accessibility, personal data) and make them testable.
  • Write a spec and an intent file that a coding agent can use (context, constraints, files involved, out of scope, verifiable definition of done).
  • Review a coding agent's plan against the spec, spot gaps before it writes code, then check the evidence of what was delivered.
  • Keep a spec alive (versions, decisions, updates when the code reveals a case) and spot the anti-patterns that make it useless.