Writing Specs with GenAI
A spec is the bridge from a rough idea to a build you can actually make. And GenAI is a fast, tireless partner for drafting one — and for stress‑testing it before you write a line of code.
What a Good Spec Contains
A spec does not have to be long. It has to be clear. Five parts do most of the work:
- The goal — what problem this solves, in one sentence
- The user — who it is for and what they are trying to do
- The key features — the handful of things it must do
- Out of scope — what you are deliberately not building
- Acceptance criteria — how you will verify it works
Let the Model Draft First
Describe your idea in plain words. Ask the model to turn it into a first spec — then ask it what you left out.
Pressure‑Test the Spec
You are too close to your own idea to see its holes. The model is not — so make it look for them.
- “What would make this fail?” — surfaces the risky assumptions
- “What did I leave undefined?” — catches the vague corners
- “What happens at the edges?” — empty input, huge input, no network
You Stay the Author
The model drafts and critiques; you decide. It is a partner, not the owner. A spec you do not understand is worse than no spec at all — it hands the whole build to a machine and hides that from you.
Build It
How to implement: give an LLM your three‑line idea from Lesson 1 and ask for a one‑page spec with acceptance criteria — then ask it to poke holes in that spec and fix what it finds.
- Weekly AI Tasks tracker — spec the “message → task” feature precisely (inputs, outputs, failure cases) with the LLM’s help before any code.
- Personal brand site — spec the sections and the honesty rule (“every claim cites a source”) so generation has guardrails.
What you learned
A spec names the goal, the user, the features, what’s out of scope, and how you’ll verify it. Use GenAI to draft it and to attack it — but you stay the author who decides.