Stop throwing “do everything” prompts at AI. Break work into tiny, clear blocks with one role per task. Define context, constraints, and acceptance criteria first—AI executes, you architect.
The story
What I just learned about getting scary-good results from AI:
AI isn’t one super-brain. It’s more like a bunch of hyper-literal junior specialists waiting for clear tickets.
When I stop throwing “do everything” prompts at it and instead give it tiny, well-shaped blocks of work, everything changes.
The moves that are working:
– One role, one outcome.
Not “help with my product.”
More like: “You are my task decomposer. Output: 7–10 tasks, each <30 minutes, each with acceptance criteria.”
- Start human, then delegate.
I frame the spec first: context, goal, constraints, definition of done.
AI is terrible at guessing context; fantastic at operating inside it.
- Think in phases, not blobs.
I separate:
- clarify assumptions
- decompose into tasks
- produce the stuff
- critique against criteria
Each phase is a different “agent hat,” not one chaotic soup.
- Treat acceptance criteria like an API contract.
“5 bullets, ≤20 words, non-technical founder, no jargon.”
or
“JSON only, these 3 fields, must parse.”
The clearer the contract, the less hallucination.
- Split explore vs commit.
First: “Give me 3 different approaches with pros/cons.”
Then: “Refine #2.”
Only then: “Implement step 1.”
Underneath all of this is one meta-skill:
Not “prompting,” but designing the shape of the work so AI can behave like a well-run team instead of a confused oracle.
If you applied this to your next project, what would you delegate differently?
Recent Comments