🤖max.bota practical guide to working with an AI assistant
6 min read

How to prompt an AI assistant

Patterns that work, patterns that don't, and the small tweaks that turn a mediocre answer into a great one.

A person and a robot having a conversation

#The 3-line rule

For most asks, three lines is enough to get a great answer:

1. What you want (the goal)
2. Constraints (what must / must not be true)
3. Context (who's reading, why, what you've tried)
✅ Specific
"Draft an email to my landlord about the broken heater. Polite but firm, in Portuguese. Mention I sent the maintenance request 2 weeks ago and got no reply. Want a response by Friday."
❌ Vague
"Write me an email to my landlord."

#Show, don't describe

For specific styles, give an example. Two short examples beat five paragraphs of description.

✅ With example
"Write a tweet in this style: 'Just shipped X. Felt like Y. Lessons: A, B, C.' Topic = launching my new project."
❌ Without
"Write me a tweet about launching my new project."

#Tell it what to skip

Most failure modes are over-doing it. "Don't pad. Don't list caveats. Don't explain what you're doing — just do it" works.

Try ending prompts with: "Be direct. Skip the preamble. Just the answer."

#Iterate, don't restart

Don't close the chat and try again from scratch when the first answer isn't quite right. Refine in the same thread:

  • "Shorter."
  • "More confident, less hedging."
  • "Cut the third bullet, expand the second with a specific example."
  • "Same content, but in PT-BR, more casual."

#Ask it to think out loud (when stakes are high)

For tough decisions or non-trivial code: ask for reasoning before the commit. Laying out the thinking surfaces mistakes that would otherwise pass.

  • "Before you answer, list the 3 angles I might be missing."
  • "Walk through your reasoning step by step."
  • "What would have to be true for this to be wrong?"

#Give it access to verify

If the assistant can run code, fetch URLs, query APIs — it will. Tell it to:

  • "Run this and show me the actual output."
  • "Curl the page and confirm the email is there."
  • "Check today's date and the day-of-week before drafting."

#Common anti-patterns

✅ One concrete ask
"Refactor pipeline/generate-brief.mjs so the system prompt is read from a YAML config. Smoke-test it after."
❌ Multi-ask soup
"Refactor the generator, also add a fact-checker, and make the dashboard show the result, and write tests for everything."
✅ State the constraints
"Don't add new dependencies. Don't change public function signatures. Keep it under 100 lines added."
❌ Open-ended permission
"Do whatever you think is best."

#Tone patterns assistants respond to

  • "Be honest, even if it stings." — drops the people-pleasing.
  • "Push back if you disagree." — will actually argue.
  • "Confidence-calibrate this — tell me what you're sure of vs guessing." — flags uncertainty.
  • "Show your work." — includes reasoning chain.
  • "Be ruthless." — cuts hard instead of expanding.

#One last meta-prompt

When stuck, ask: "What would a better prompt look like for what I'm trying to do?"

The assistant will rewrite the prompt better than the user would. Then run that.