From Spec to Tasks
A spec describes the whole thing you want — and that is almost always too big to build in one go. The move is to break it into small, ordered tasks that a person or an agent can finish and verify, one at a time.
What Makes a Good Task
Not every to‑do is a task worth handing off. A good one has four properties.
- Small — finishable in one focused sitting, not a week.
- Clear done‑condition — you can tell when it is finished.
- Independent where possible — not blocked on three other tasks.
- Limited surface — it touches a small, known part of the system.
Decomposing the Spec
Slice the spec by feature or by layer — then order the pieces so each task builds on a working previous one.
Agent‑Sized vs Human‑Sized
Once tasks are small, you can decide who takes each one.
- Give the agent well‑scoped, verifiable tasks — a clear done‑condition it can be checked against.
- Keep for yourself the ambiguous work and any high‑stakes judgment call.
- When in doubt, tighten the task until it is verifiable — or hold onto it.
Each Task Is a Loop
A task is not just a line on a list. Each one is a build → inspect → decide loop — the same rhythm from Module 1 — and each carries its own acceptance check, exactly as you defined in Module 4, Lesson 4.
Build It
How to implement: take your capstone spec and write the first 5 tasks in order, each with a one‑line done‑condition.
- Weekly AI Tasks tracker — 1) data model, 2) list‑this‑week endpoint, 3) week view, 4) message webhook, 5) parse‑message → task.
- Personal brand site — 1) page skeleton, 2) about section, 3) projects from real repos, 4) styling, 5) deploy.
What you learned
A spec is too big to build at once, so you break it into small, ordered, verifiable tasks. Slice by feature or layer, build a walking skeleton first, hand the verifiable work to the agent — and treat every task as its own build→inspect→decide loop.