10 AI prompting do's and don'ts.

Twenty small habits separate a prompt that gets you a usable answer from one that gets you a shrug in text form. None of them are secret tricks. They all close the same gap: the one between what you actually meant and what you typed.

A good prompt gives the AI three things a topic alone never does: what you want done, what matters most to you and what to avoid. The list below splits that into ten habits worth building and ten worth dropping, each with a one-line example so you can spot the pattern in your own prompts.

The 10 do's

1. Say what you want done, not just the topic. “Smartphone trends” gets you a Wikipedia summary. “500-word exec summary of smartphone trends for a board deck, three sections, one recommendation each” gets you something you can paste into a slide (Kanerika).
2. Front-load the context that changes the answer. A workout plan for someone with a bad knee and stairs at home is a different plan than a generic one, and the AI only knows the difference if you say so.
3. Break big asks into steps. “First explain X, then show Y, then give an example” produces a cleaner answer than one paragraph asking for all three at once (Kanerika).
4. Ask for three options, not one. Forcing alternatives pushes the model past its first, safest guess, useful for anything from a workout plan to a headline (Panaversity, AI Prompting in 2026).
5. Name a role only when it changes the answer. “Explain this like I'm new to spreadsheets” earns simpler language. Skip the role when the task is a fact, not a style.
6. Show one example of the output you want. For anything with a specific structure or tone, a single sample does more work than a paragraph describing it (Kanerika).
7. Give feedback and build on the answer instead of restarting. “Keep the structure, drop the second example, make the tone less formal” gets you further than a new prompt from scratch. Treat it as a dialogue, not a one-shot order, and prompting again beats giving up after one try (Ethan Mollick).
8. Ask it to flag what it couldn't verify. For anything you'll act on, “mark anything you're not sure about” catches more than hoping the answer is right (Panaversity).
9. Retest important prompts in a second chat or model. The same request can land differently in ChatGPT, Claude or Gemini, not by chance but by design, each is built to weigh instructions differently, worth knowing before you rely on one answer for something that matters (ByteByteGo).
10. Save the prompts that worked. A short list of prompts that produced good results is faster to adapt than writing a new one every time (Kanerika).

The 10 don'ts

1. Don't stay vague and expect a sharp answer. “Help me write something” leaves the AI guessing at the one thing that actually mattered to you (Kanerika).
2. Don't assume it remembers an earlier chat. Each conversation works from what's typed into it; nothing carries over from a different tab or yesterday's session.
3. Don't cram five tasks into one prompt. Asking it to draft, format, fact-check and shorten all at once tends to produce a worse version of all four (Treyworks).
4. Don't ask it to judge your own idea without a yardstick. “Is this a good idea?” invites agreement. “Score this against whether there's a real problem, a market and a competitive edge” invites an actual answer (Panaversity).
5. Don't leave out what to avoid. A prompt with no boundaries on length, tone or scope leaves more room to miss than one with two lines of constraints.
6. Don't treat the first draft as final. Good prompts, like good drafts, get there through a round or two of adjustment, not a single try (Kanerika).
7. Don't only test the clean version of your request. A prompt that works with a tidy example can fall apart on the messy, real version of the same task (Kanerika).
8. Don't take confident answers at face value on obscure topics. Frequency in training data tracks roughly with reliability: well-documented topics are safe to trust, a niche regional or technical detail is worth checking against a primary source (Panaversity).
9. Don't dress up a simple request in complicated language. A short prompt in plain language usually beats a long one trying to sound precise (Kanerika).
10. Don't file a bad answer as a dead end. A response that missed the mark is information about what the prompt left out, not a reason to give up on the tool for that task.

Worth remembering

None of this is about the model falling short. Why you get bad AI answers., the previous post in this series, already covered the core issue: a search-engine habit meets a tool built for conversation, and the gap in between is where the bad answers live. These twenty habits close that gap one prompt at a time, by hand.

Most of this list stops being something you have to remember once the input is structured before it ever reaches the AI. That's where CORDON comes in, it provides the structure for you.