A prompt is a specification, not a question. This lesson reframes the whole course around that one idea.
A prompt is a specification. That sentence sounds small, but it changes how you write every instruction from here on. A question asks a person for information they already hold. A specification states what a result must contain, who it is for, and what counts as finished. Models do not hold your intent; they hold the text you sent. Everything they cannot infer from that text is invented, guessed, or omitted. Writing a prompt well is therefore the work of moving intent out of your head and into the instruction.
Questions hide requirements
Ask "write me a product description" and you have stated a topic, not a requirement. Nothing in that sentence says how long the description should be, whether it should mention price, whether the tone is playful or clinical, or who reads it. The model will produce something plausible because plausible is what it was trained to produce. Plausible is not the same as correct. Three runs will give three different shapes, and none of them is wrong in a way you can point at — which is exactly why vague prompts are so hard to debug.
Example: "Write a 60-word product description for a stainless steel water bottle, aimed at commuters, in a calm and factual tone. Mention insulation and the leak-proof lid. Do not mention price." Every clause removes a decision the model would otherwise make for you.
A specification has four parts
Most working prompts contain the same four elements, in some order. Context says what situation the model is in and what it may assume. Task says what to produce, with a verb you can act on. Constraints say the boundaries: length, format, tone, things to avoid. Output shape says what the result should look like so a machine or a busy reader can use it. When a prompt fails, one of those four is usually missing, and the failure is not random — it is the model filling the gap you left.
Why this matters more than model choice
People often change models when a prompt underperforms. Occasionally that helps, because different models have different defaults. Far more often the prompt had no specification to follow, so every model improvises differently. A clearly specified prompt produces comparable results across models, and a vague one produces noise on all of them. Fix the specification before you change the tool.
Checklist
- Does the instruction name what to produce, not just the topic?
- Does it state who the result is for?
- Does it say what to avoid, not only what to include?
- Could a colleague produce the same result from this text alone?