The simple explanation
Say what you need done (the task). Supply the information needed to do it (the context). State the limits — length, tone, what to avoid (the constraints). Show the shape you want back — bullets, a table, three sentences (the format). And keep source material visibly separate from instructions, so it’s clear which text is data and which is direction. Instructions buried inside retrieved documents are untrusted content: the model can’t tell who wrote them.
See this idea move.
The Prompt Anatomy experiment walks through this concept step by step — press run and watch it happen. Everything is simulated in your browser; no real AI runs.
A concrete example
Build a prompt to summarize a support ticket: “Summarize this ticket in 3 bullets for a manager. Flag any refund request. If anything is unclear, say so instead of guessing. Ticket: …” Change the format from bullets to a table and only the requested output changes — the underlying facts don’t. That’s the point of the format line: it controls presentation, not truth.
Now compare that with the most common bad prompt: “Tell me about the support ticket.” No audience, no length, no indication of what matters. The model guesses all of it — and every guess is a chance to be wrong about what you wanted. Most prompt failures aren’t model failures; they’re missing specifications. The model answered the question you wrote, not the one you meant.
Where people get misled
The “prompt engineering” mystique. A clear brief beats clever tricks; magic words and incantations don’t reliably change behavior. What works is boring: specific task, sufficient context, explicit constraints. The second trap: over-iterating the prompt instead of fixing the inputs — if the model lacks the facts, no wording supplies them.
One more useful habit: iterate on one variable at a time. Change the format, test. Add a constraint, test. Shotgun-editing the whole prompt teaches you nothing about which change helped. Prompts are specifications, and specifications get better through controlled edits, not vibes.
The honest limits
Better prompts cannot supply missing evidence or guarantee correctness. A perfect prompt on top of no sources is still guessing — politely formatted guessing. Treat instructions found inside retrieved documents as untrusted, and verify what matters.