A Prompt Is Not an Incantation

It's a requirements document. Wherever you didn't specify, the model fills in its own answer.

The web is full of "universal prompt templates" and "prompt hacks," framed like incantations: say it right and it works. The mechanism is plainer: every sentence you write shrinks the range of what the model might produce.

Symptoms you'll recognize:

• A coworker sends a "life-changing prompt"; you copy it and the results are ordinary.
• You are sure you were clear, and what it returns is nothing like that.
• You add "think carefully" and "this is very important" and feel nothing change.

It narrows a range

Back to the mechanics: at every step the model computes "what is the most likely next token," relative to everything it has seen so far.

So when you say only "write copy," the possibilities are nearly infinite: for whom, what product, how long, what tone? None answered, but it must keep writing, so it picks whatever looks most common in its training corpus. The result is a competent, unmemorable generic paragraph.

Every condition you add narrows the range. Add "for users who have been with us three months" and it rules out all new-user approaches. Add "under 100 characters, no exclamation marks" and what remains is already close to what you want.

That is the picture above. Prompt engineering is not a magic phrase; it is systematically saying everything that matters. OpenAI's and Anthropic's official prompting guides say the same thing: be specific, be testable, spell out the conditions. openai-prompt anthropic-prompt

Treat it as a requirements document

Here is a useful comparison: treat the model as a very capable new colleague who knows nothing about your company.

They are smart, quick to learn, and know a bit of every field. But they do not know your product's name, who your users are, why the last proposal was rejected, or what your tone should be. Every piece of background you leave out of the handoff, they have to guess.

How do you hand off work to a new colleague like that? You say who it is for, what outcome you want, what the constraints are, and what done looks like. That is exactly what a prompt should be.

This also explains why someone's "universal prompt" often flops for you: it is packed with their background and constraints, which are not yours. You can borrow the structure, not the content.

Three common misconceptions

First, a magic phrase can dramatically improve results. "Think carefully" and "this is very important" have limited effect, and the reason is straightforward: they carry no new information and narrow nothing. What works is concrete conditions: "use this format," "follow this example," "do not use these words." The one near-exception is asking it to reason before answering, which helps, but not because it is a spell; because writing more is computing more.

Second, longer is better. No. What matters is information density, not length. Piling on irrelevant background dilutes the key instructions and raises cost. Worse, with enough content, the crucial line gets buried in the middle, and the middle is exactly where attention fades.

Third, it understood me. It does not understand; it does statistical continuation. You cannot see the difference when things go well, but you see it immediately when they go wrong: it will not stop and ask when the task description is ambiguous, the way a person would. It picks one interpretation and confidently runs with it. So the cost of ambiguity is entirely yours, which is why the next step is listing the parts you must spell out.

Where to start editing

If you have a prompt that underperforms, check in this order rather than rewriting from scratch.

First, surface your implicit assumptions. You were thinking "obviously for B2B clients," "obviously in Chinese," "obviously no competitor mentions" — the model knows none of these. Write them in, one by one.

Second, add a "don't" list. Positive descriptions are not enough; negative lists narrow faster. "No exclamation marks," "not in marketing voice," "do not invent data" — one of these outdoes three positive sentences.

Third, give an example. If you can provide a sample of "this is what right looks like," it usually beats three paragraphs of adjectives.

Fourth, define done. "Output three items, each under 20 characters, numbered" beats "summarize concisely."

Then look at the results again. Most "the model is bad" cases resolve right here.