A practical checklist for writing better AI prompts

Most disappointing AI answers trace back to an underspecified request. The assistant is not reading your mind; it is continuing the text you gave it. A short checklist, applied before you press enter, fixes most of it.

The checklist

• Task. Say what you want done, in a verb: summarize, compare, rewrite, list, explain, find the bug.
• Purpose. Say why. "I am choosing between two quotes for my landlord" leads to a different answer than "compare these quotes."
• Audience. Say who it is for. An explanation for a ten-year-old and one for an engineer are different texts.
• Context. Include the facts the assistant cannot know: your constraints, what you have already tried, definitions of your terms.
• Source material. Paste the text to be used instead of asking the assistant to recall it.
• Constraints. Length, tone, what to include, what to leave out.
• Format. A table, five bullet points, a two-paragraph email, steps in order.
• Success test. Say how you will judge the result, for example "it should fit on one screen" or "every claim should have a quote."

You rarely need all of these. Pick the ones whose absence would change the answer.

A before and after

Before: "Write about our new feature."

After: "Write a 150-word announcement for existing customers about our new export-to-spreadsheet feature. The tone is friendly and plain, with no marketing superlatives. Mention that it works on the Pro plan only, and end with a link to the help page. Here are the release notes: ..."

The second version tells the assistant the task, audience, length, tone, a constraint, and the facts, and gives it the source text.

Techniques that help

• Break big tasks into steps. Ask for an outline first, agree on it, then ask for each section. Mistakes are cheaper to fix early.
• Show an example. One sample of the format or tone you want is often clearer than a paragraph describing it.
• Ask it to ask you. "Before you start, ask me any questions you need answered" surfaces what you forgot to say.
• Ask for assumptions. "List the assumptions you made" shows where it filled a gap by guessing.
• Request alternatives. "Give me three different approaches and the trade-offs" beats one confident answer.
• Say what not to do. "Do not invent statistics" or "do not change the headings" prevents common problems.

Iterate instead of starting over

If an answer is close, reply to it: "Shorter, and drop the second paragraph," or "Keep that structure but use plain words." Replying keeps the useful parts. If a conversation has wandered, starting a new one with a better first message often works better than correcting the old one, because the old text still counts against the context window. See what a context window is for why.

Common mistakes

• Asking for certainty. If the assistant does not know, a prompt that demands an answer invites a guess. Why AI makes things up covers this.
• Leading questions. "Isn't it true that..." gets agreement more often than accuracy.
• Too many tasks in one message. Each added task dilutes the others.
• Hidden requirements. If something matters, say it. "Obviously" is not shared context.
• Pasting secrets. Passwords, access keys, and other people's private details do not belong in a prompt. See what to check before sharing data with an assistant.

Quick answers

How long should a prompt be? As long as it takes to carry the facts the assistant needs. A simple request can be one line. A task that depends on context needs more. Length is not quality.

Do phrases like "think step by step" help? Asking for the working can help with reasoning tasks on some models, and the effect varies. Being specific about the task and giving the facts matters more than any phrase.

Does politeness matter? A courteous tone does no harm, but clarity is what changes the answer.

Should I repeat my instructions in a long chat? Yes, restate the key constraints now and then. In a long conversation, early text may no longer be in view.

Where NAMELESS stands

NAMELESS is an AI model that is still in training. It is built to remember what you have already told it, so you do not repeat yourself, and it does not reread the conversation to follow it. Clear requests help with any assistant.

You can join the waitlist to follow progress. To see how tokens count what you write, read what a token is.