Same model, two answers
Your team launches a new app this week. You open the same AI, twice, and send each of these prompts. The model is identical — only the instruction differs. Pick the prompt you would trust to give you something you could actually ship, then check your reasoning.
The model did almost exactly what each prompt asked. When an AI answer disappoints, the reflex is to blame the model. But far more often the instruction was the problem: a vague prompt leaves the model to fill every gap with a confident guess of its own. A prompt is not luck — it is something you can audit and rebuild. The rest of this lab is that craft, run on one real prompt.
Build the prompt, part by part
A strong prompt is composed, not guessed. The same five checks find what is missing in any instruction — run them in order and you can audit a prompt in thirty seconds. C.R.I.S.P. is that lens.
Context
Background the AI can’t know.
Role
A role only when it changes the answer.
Instruction
One explicit aim, as a directive.
Specifics
Constraints plus a few examples.
Presentation
The exact output shape.
The Forge — assemble it and watch the meter
The task: reply to a customer whose order #4471 arrived broken. Toggle each part into the prompt. The good parts (C.R.I.S.P.) raise a live clarity meter; the over-constraint extras are flagged — they lower it, because contradictory rules and shouting make a modern model worse, not better. The clarity you reach here becomes the bar the Refinement Room scores against.
Audit the weak prompts
1. Two ways a prompt fails
Most prompts fail in one of two opposite directions. Under-specified — a part is missing, so the model invents it. Or over-constrained — so many (often contradictory) rules that the model can’t satisfy them all and drops one silently. Sort each prompt into the failure it commits. Click an item to move it between bins, then check.
2. Spot the dead weight
Here is the over-constrained brief from Scenario 4, laid out clause by clause. Click every clause that adds no value or actively conflicts with another — the dead weight a calm rebuild would cut. Leave the one clause worth keeping. Then check.
Click each clause that should be cut, then press Check.
The Refinement Room
Refinement is a loop, not a one-shot: collect the reply, spot the gap, change one part, re-run. Here is one prompt — “Summarize this report” — carried forward across rounds. Each round, pick a single change. The prompt and the model’s reply rewrite, and the clarity meter moves. It can move down. Add the one part that matters and it climbs; refine many things at once or pile on rules and it falls.
Know when more instruction starts hurting
1. The over-constraint cliff
The prompt is at 80% — clear, scoped, good enough to ship. The hardest judgment in prompting is right here: do you keep adding, or stop and verify? Choose a path, watch the consequence on the reply and the meter, then try the others.
2. Estimate the lift of one good example
Examples (few-shot) are routinely the highest-ROI fix in prompting — especially for locking output shape. Estimate this: switching a format-sensitive task from zero examples to a few worked ones, by roughly how many percentage points does task success typically improve?
Takeaway
1. Write the missing part
Here is a prompt that is strong on Context and Instruction but missing its Presentation — the output shape. Write the Output-Format line you’d add, then self-assess against the checklist and compare to a model answer.
A strong Output-Format line
“Return exactly 5 bullet points, each under 20 words. Lead each bullet with the relevant figure, then the one-line takeaway. End with a single sentence naming the biggest risk.”
It names a concrete shape (5 bullets), a checkable length (under 20 words each), and where the numbers go — so the board can read it in seconds and two runs would be usable in the same slot. That is the P in C.R.I.S.P. doing its job.
2. Your audit, on a card
One page to keep. Run these five checks over any prompt before you send it.
CContext
- Paste the source the AI can’t see
- Name the real audience
- Omitted context is the top cause of generic output
RRole
- Set a role only when it changes the answer
- “Explain to a board” vs “to an engineer”
- “World-class expert” is usually decoration
IInstruction
- One verb, one deliverable
- A topic is not a task
- Decompose four asks into a verified sequence
SSpecifics
- Add a few worked examples (few-shot)
- Keep the constraints that matter, drop the rest
- Positive checkable rules beat vague negatives
PPresentation
- Name the exact shape and length
- Static content first, variable data last
- Use delimiters to separate instruction from data
TThe edge
- More instruction isn’t monotonically better
- The over-constraint cliff: rules can conflict
- Know when to stop refining and start verifying
You’ve audited the prompt.
You built a prompt from its parts, watched clarity climb as you added the right one, saw the meter fall when you piled on contradictory rules, and learned where to stop refining and start verifying. That is the difference between prompting as luck and prompting as a craft you control.