A Spec Is a Hypothesis
A spec is a hypothesis, not a contract. With AI parts especially, you will learn things while building that change the plan — and that is normal, not failure.
The Spec Is Living
You do not write the plan once and march. You run a short loop, and the plan updates as you learn.
- Build a slice — the smallest piece that produces a real result
- Inspect the real result — look at what actually came out
- Update the spec and tasks — fold in what you just learned
- Continue — take the next slice from the revised plan
You Can't Know Until You Try
With ordinary code you can mostly reason about behaviour up front. With a model you cannot — you only find out how it behaves by running it on real inputs.
Let the Evidence Steer
Certain moments are worth more than the roadmap. Notice them, and let them decide the next step.
- Where the output is wrong — the model missed, or the task was ill-posed
- Where you were surprised — the result you did not predict is the most informative
- Where the plan met reality — an assumption that held, or quietly broke
These steer the next step — not a fixed roadmap drawn before you knew any of it.
Between Two Failures
Iterating well means avoiding both extremes at once.
- Thrashing — changing direction every minute, so nothing ever gets far enough to teach you anything
- Rigidity — following a stale plan off a cliff because it was the plan
The corrective for both is the same: steer with evidence, not with mood and not with momentum.
Build It
How to implement: build the riskiest slice of your capstone first — the part you are least sure will work — and let what you learn reshape the rest.
- Weekly AI Tasks tracker — prototype the message→task parse early; its real behaviour will reshape your schema and prompts.
- Personal brand site — draft one section end-to-end (generate → verify → refine) before scaling to all sections.
What you learned
A spec is a hypothesis you revise as you build. Run the loop — slice, inspect, update, continue — and with AI parts, prototype early because you cannot know behaviour until you try it. Steer with evidence, avoiding both thrashing and rigidity.