Most people aren't bad at using AI. They're bad at asking.
I've spent the last few months deep in LLMs, RAG pipelines, and agentic workflows - building data pipelines by day, and breaking down prompting patterns for my mentees by night. One pattern shows up again and again: the gap between a mediocre AI output and a genuinely useful one almost never comes down to the model. It comes down to the prompt.
Here's the playbook I've landed on.
The Do's
- Give context before the ask
Don't just say "write me an email." Say who it's to, what you want the reader to do after reading it, and what tone fits.
Weak prompt:
Write an email about the project delay.
Strong prompt:
Write a short email to my manager explaining that the data migration is delayed by 3 days due to a schema mismatch. Keep the tone calm and solution-focused — include what we're doing to fix it and the new ETA.
AI can't read your mind, but it can read your context. The more situational detail you give upfront, the less back-and-forth you need later.
- Show, don't just tell
One good example output is worth three paragraphs of instructions. If you know what "good" looks like, paste it in.
This is especially powerful for tone, formatting, or style-matching tasks — instead of describing your voice, show a sample of your past writing and ask the model to match it.
- Ask for reasoning before the answer
"Think step by step before you respond" sounds like a cliché, but it genuinely improves accuracy on anything involving logic, math, or multi-part decisions. Asking the model to lay out its reasoning first also makes it easier for you to catch errors before they reach the final answer.
- Iterate like it's a conversation, not a search query
Your first prompt is a draft, not a final exam. Treat the second and third prompts as refinements, not failures. The best outputs I get are almost never from the first prompt — they come from 2-3 rounds of "good, now adjust X."
- Specify format explicitly
Bullet points, table, word limit, tone, structure — say it upfront. Don't make the model guess your output shape. If you need a table with specific columns, say so. If you want under 200 words, say so. Ambiguity in, ambiguity out.
The Don'ts
- Don't ask vague, open-ended questions and expect a sharp answer
"Tell me about marketing" gets you a Wikipedia summary. "Give me 3 GTM strategies for a B2B SaaS targeting mid-market fintechs" gets you something you can actually use.
Specificity in the prompt is what buys you specificity in the output.
- Don't skip the "why"
Telling the model your goal — not just your task — changes the quality of the output dramatically. A task tells the model what to produce. A goal tells it why, which shapes the reasoning behind how it gets there.
- Don't treat every answer as final
AI is confidently wrong sometimes, especially on numbers, citations, or niche facts. Verify anything factual, numerical, or high-stakes before you ship it. Confidence in tone is not the same as correctness in content.
- Don't over-engineer simple asks
Not every prompt needs five paragraphs of instructions. Match your effort to the complexity of the task — a quick rephrase doesn't need the same scaffolding as a multi-step analysis.
- Don't forget you're the editor
The model drafts. You still decide what's true, what's on-brand, and what actually ships. Treating AI output as a first draft — not a finished product — is the single biggest mindset shift that improves the quality of what you put out into the world.
The real unlock
It isn't a smarter model. It's getting specific about what you actually want, and treating the AI like a sharp junior collaborator who needs context — not a search engine that needs keywords.
What's one prompting habit that's changed your output quality? I'd love to hear it in the comments.
Top comments (0)