Most disappointing AI output is not a model problem. It is a missing-information problem. The assistant answered the question you typed, not the question you meant, because the things that made your version obvious — who it is for, what came before, how long it should be — only existed in your head.
The five pieces
- Role — who the assistant should behave like, but only when it changes the answer. "You are a careful financial analyst" matters. "You are a world-class genius" does not.
- Context — the real material. Your notes, last quarter's numbers, the actual client brief. Paste it rather than describing it.
- Task — one thing. If your prompt contains the word "and" three times, split it.
- Format — length, structure, sections, whether you want a table or prose.
- Constraints — what to leave out, what not to invent, and permission to say "I don't know".
What it looks like in practice
Here is a prompt for a task almost everyone has some version of: turning messy notes into something you can send.
You are helping me write an internal update for a non-technical leadership team. Context: the raw notes below are from a 45-minute project review. Task: turn them into an update. Format: a two-sentence summary, then three bullets for decisions made, then three bullets for what is still open with the owner's name. Constraints: use only facts in the notes, do not soften bad news, and list anything ambiguous at the end under "Needs confirmation" rather than guessing.
>
Notes: [paste]
The prompt is longer than what most people type. That is the point — you are writing the brief you would give a competent new hire.
Iterating beats rewriting
The second habit matters as much as the structure. Treat the first answer as a draft. Tell the model what was wrong and show it what right looks like:
Two problems with that draft: the summary buried the delay, and the tone is too formal. Here are two sentences of my own writing for tone: [paste]. Rewrite it, keeping every fact and inventing nothing.
Three rounds of correction on one good prompt will beat thirty new prompts every time.
Try this today
Take a prompt you already use regularly and add the two pieces it is almost certainly missing: real context, and constraints on what not to do.