AI Best Practices — Do's, Don'ts & Mental Models

A scannable reference card. The “why” matters as much as the “what”.

• 01

  Give AI a role before the task

  ✅ Do

  Start every substantive prompt with “You are a [specific role] at a [specific context]…”

  ❌ Don't

  Jump straight to the task without framing who is answering.

  💡 Why Models adjust tone, depth, vocabulary and assumptions based on role framing. “You are a senior growth PM” produces fundamentally different output than no framing — it calibrates the model's prior on what matters.
• 02

  Specify the output format explicitly

  ✅ Do

  End every prompt with “Return as: markdown table / JSON array / numbered list / one paragraph.”

  ❌ Don't

  Let the model decide how to structure the output.

  💡 Why Unspecified format means unpredictable format — which breaks downstream automation and makes copy-pasting into docs painful. Format is a contract between you and the model.
• 03

  Use examples, not just instructions

  ✅ Do

  Include 1–2 examples of the output you want (few-shot prompting).

  ❌ Don't

  Describe what you want only in abstract terms.

  💡 Why One concrete example communicates more precisely than 100 words of description. It shows, not tells — and dramatically reduces misinterpretation on complex or nuanced tasks.
• 04

  Decompose complex tasks into steps

  ✅ Do

  Break multi-part tasks into sequential prompts, or use “First do X, then do Y, finally do Z.”

  ❌ Don't

  Ask AI to do 5 things in a single prompt and expect all 5 to be equally good.

  💡 Why AI quality degrades with task complexity in a single prompt. Each sub-task competes for attention. Sequential prompting — where each step builds on the last — produces dramatically better results than one mega-prompt.
• 05

  Verify facts, numbers, and attributions

  ✅ Do

  Treat all specific claims, statistics, and quotes as unverified until you check the source.

  ❌ Don't

  Copy AI-generated numbers or quotes into a brief or presentation without verification.

  💡 Why AI models can confidently generate plausible-sounding but incorrect data — called hallucination. For growth work especially, always verify. Use AI for synthesis and structure, not as a source of truth.
• 06

  Give context about your constraints

  ✅ Do

  Tell AI what you can't do: “No eng bandwidth this quarter”, “Budget under €5K”, “Live in 48h.”

  ❌ Don't

  Ask for ideal-world solutions without real-world constraints.

  💡 Why Without constraints, AI optimises for completeness, not feasibility. A brief that assumes 3 months of engineering time is useless if you have 3 days. Constraints force more creative, actionable outputs.
• 07

  Iterate, don't restart

  ✅ Do

  Build on the previous response: “Good — now make the tone more direct” or “Keep the structure but change the angle.”

  ❌ Don't

  Copy-paste a fresh prompt every time the output isn't perfect.

  💡 Why AI has the full conversation in context. Iterative refinement is faster and more precise than restarting — you're converging on the right answer, not starting over. Think of it like editing a document, not re-writing from scratch.
• 08

  Don't outsource your judgment

  ✅ Do

  Use AI to generate options, drafts, and analysis — then apply your own judgment to decide.

  ❌ Don't

  Accept AI recommendations as decisions, especially for strategy, hiring, or user-facing copy.

  💡 Why AI optimises for plausibility, not correctness. It doesn't know your users, your culture, your brand voice, or your politics as well as you do. Your experience is the filter. AI is the input generator, you are the decision-maker.
• 09

  Use system prompts for recurring contexts

  ✅ Do

  Create a saved system prompt with your role, company, ICP, and current priorities — so every conversation starts informed.

  ❌ Don't

  Re-explain your context at the start of every new chat.

  💡 Why Context switching is expensive. A saved context means every conversation starts informed — like talking to a colleague who already knows your situation, not a stranger you have to brief every time.
• 10

  Treat your prompt library as a product

  ✅ Do

  Save prompts that work, version them, and iterate when they stop performing.

  ❌ Don't

  Throw away prompts that worked well or rely on memory to recreate them.

  💡 Why A prompt that reliably produces a great experiment brief is an asset. It compounds — a team using the same high-quality prompt template produces more consistent outputs than a team improvising every time. Your prompt library is infrastructure.
• 11

  Use AI for the 80%, not the 20% that matters most

  ✅ Do

  Use AI to draft, structure, synthesise, and generate options — then spend your time on the critical decisions and the human context.

  ❌ Don't

  Use AI for the parts of your job that require your unique context, relationships, or judgment.

  💡 Why The 80% (formatting, structuring, first drafts, summarising) is where AI saves the most time with the least risk. The 20% (what to prioritise, how to frame a strategic bet, how to read a stakeholder) is where your human judgment is irreplaceable — and where AI “help” can actively mislead.
• 12

  Tell AI when it's wrong

  ✅ Do

  Push back directly: “That's not right — here's why: [reason]. Try again.”

  ❌ Don't

  Assume the first output is the best the model can do.

  💡 Why AI is not fragile. Correction is the fastest path to better output. Models that are corrected with context — not just told “try again” — produce dramatically better second attempts. Dialogue beats monologue.

How to frame AI in your head. Frames eat tactics for breakfast.

• 1

  AI as a brilliant intern, not an oracle

  Think of AI as an extremely capable person who just joined your team — fast, knowledgeable, eager to help, but with no context about your company, your users, or your history. It needs briefing, not just questioning. The better you onboard it to a task, the better it performs.

  Before asking AI to write a hypothesis, give it your company context, your current experiment backlog, and what you've already tested. The output will be in a different league.

• 2

  The quality of your output = the quality of your input

  AI is a compression function: it produces structured output from structured input. Vague in, vague out. The effort you invest in a clear, constrained, contextual prompt is never wasted — it returns 10x in output quality. Prompting is not a shortcut. It is the work.

  “Write a landing page” takes 30 seconds and returns generic output. A briefed prompt with ICP, JTBD, primary objection and section structure takes 3 minutes and returns something publishable.

• 3

  Iteration beats perfection on the first try

  The best AI users treat conversation as a drafting process, not a single transaction. Plan for 3–5 turns: generate → evaluate → refine → evaluate → finalize. Each turn is cheap. Expecting perfection on the first attempt leads to frustration and worse output than iterating systematically.

  Generate 8 hypotheses → rank by expected impact → write briefs for the top 3 → refine the brief for the one you'll run. 4 turns, 10 minutes, brief-ready output.

• 4

  AI has no memory between sessions

  Every new conversation starts blank. AI doesn't remember what you discussed last week, your team's priorities, or your brand voice — unless you tell it again. A saved system prompt and a prompt library aren't nice-to-haves; they're the infrastructure that turns AI from a novelty into a workflow.

  Build a Notion doc “AI Context — Arnaud” with role, company, ICP, OKRs, and brand voice. Paste it at the start of any important conversation. 30 seconds of setup, much better output.

• 5

  Automation amplifies both quality and errors

  When you automate an AI workflow, its quality — good or bad — runs at scale. A prompt that produces 90% quality is fine one-off. Automated at 1,000 items per week, the 10% errors compound. Design automations with quality checkpoints, not just speed.

  Your translation agent should flag low-confidence translations for human review — especially for marketing copy where brand voice matters.