Cut the work into agent-sized tasks
Lesson 3 of 5 — the unit of work that keeps you in control.
How to size a task for review, sequence tasks to manage risk, and recognise when a prompt is too big to trust.
Prerequisites
Lesson 2 — you have a spec to decompose.
The reviewable unit is the real unit of AI-augmented work. If you cannot hold an entire change in your head well enough to spot what is wrong, the change is too big — no matter how good the model is. So you cut.
Task decomposition
Cutting a feature into units small enough to review and verify in a single pass — the unit of work in AI-augmented development.
Want the full picture? Open the reference →
How to size a task
- One concern per task — the route, OR the validation, OR the UI. Not all three.
- Independently verifiable — each task ends at a state you can run or test.
- Reviewable in one sitting — if the diff would not fit on your screen, split it.
- Ordered by risk — the contract and data model first; they are expensive to change once code leans on them.
1. Schema + migration: post_draft table (postId, body, updatedAt). [verify: migration runs]2. API route: POST /api/posts/:id/draft, owner-only, upsert. [verify: 3 tests pass]3. Client hook: useAutosave(postId, body) — debounce + retry queue. [verify: saves after 2s idle]4. Wire into editor + a "saved at HH:MM" indicator. [verify: by hand]DetailSequencing: why contract-and-data go first
Code written against a contract is cheap to keep; a contract changed after code depends on it is a cascade of edits. Lock the shape (schema, signatures) in an early, small, heavily-reviewed task. Everything downstream then leans on something you already trust, and each later task stays a one-pass review.
Decompose your feature into 3–5 tasks. Write the one-line verification for each before you prompt for any of them.
Concepts covered · 2
More like this
Review what comes back like a senior
Models produce confident, plausible, well-formatted code — which is exactly what makes their mistakes dangerous. This lesson is the review rubric: the specific failure classes to hunt for, because "it looks right" is not review.
Orchestrate agents without the codebase rotting
Putting it together: context, guardrails, and a verification loop that lets you run coding agents across a real codebase without it drifting into an incoherent pile. This is the senior loop at full speed.
Structure & style the app
Lay down the HTML skeleton and a calm dark style for a todo app. No JavaScript yet — just the shape, so the behaviour you add next has somewhere to live.
Liked this one?
Pass it on
Discussion
Be the first to ask
Your questions stay private with the team. We can pin answers to share with everyone.
Sign in to ask a question privately on this tutorial.
Sign inNo pinned answers yet for this tutorial.