AI Fluency Checklist

Chat was the doorway. The rooms beyond it — images, voice, audio — are where a lot of the value hides.

• Generate an image you'll actually use — a slide background, a header, a concept sketch. Seeing AI make something usable breaks the "it's just a chatbot" frame and opens a whole category of work you were skipping or paying for.

  Try this: Swap in your own concept, generate it, then drop it straight into your next deck or doc.

  Prompt A clean, minimal hero image of [your concept], muted colors, lots of negative space for text. Copy

  Image generation tools ↗ now and then
• Talk instead of type for a day with a voice/dictation tool. Speaking runs 3–4× faster than typing and kills the friction that stops people using AI for longer thinking.

  Try this: Draft your next long message or note by speaking it — try Wispr Flow, Superwhisper, or even your phone's built-in dictation.

  Voice & dictation tools ↗ now and then
• Turn a dense document into something you'll absorb — an audio overview or a Q&A. The knowledge in a 40-page PDF is useless if you never read it. A different modality gets it into your head on a commute.

  Try this: Upload a report to NotebookLM, generate an audio overview, then ask it questions like a tutor.

  NotebookLM ↗ now and then

You've got the context habit from Level 1. Now a few techniques that reliably lift the output.

• Give the AI a role and an audience, not just a task. "Act as a skeptical CFO" frames the whole response. The same question answered "as an expert X, for audience Y" comes back sharper.

  Try this: Swap in any role and audience that fit your task. A worked example:

  Prompt You're a skeptical CFO. Here's my budget proposal — poke holes in it before I send it to the real one. Copy

  Getting Started → use roles ↗ keep doing it
• Show, don't tell: paste an example of what "good" looks like. One strong example teaches your standard faster than three paragraphs describing it (this is "few-shot" prompting, minus the jargon).

  Try this: Paste real examples of your "good", then ask it to match them:

  Prompt Here are two subject lines I love: [paste them]. Write five more in exactly that style. Copy

  Guide coming soon: A Practical Prompting Guide keep doing it
• Break a big ask into steps instead of one giant prompt. "Outline first, then draft" beats "write the whole thing" — you catch the wrong turn at the outline, not after 800 words.

  Prompt First give me an outline. I'll approve it, then you write it section by section. Copy

  Working With AI, Not Through It ↗ keep doing it
• Ask for three versions, not one. Re-running is nearly free. Comparing options beats polishing the first thing it handed you.

  Try this: Ask for a spread of tones, then steal the best bits of each.

  Prompt Give me three versions: one formal, one warm, one punchy. Copy

  keep doing it

Level 1 was a few bullet points. This is the version that compounds — built once, paying off in every chat.

• Let the AI interview you to build your profile, then save it as custom instructions. You don't have to know what to include — let it ask. The result is a dense brief that upgrades every future answer.

  Try this: Answer honestly; the profile is only as good as what you tell it.

  Prompt Interview me with 10 questions to build a profile that helps you give me better answers. Then write it up as custom instructions I can paste into settings. Copy

  Let AI Interview You ↗ one-time setup
• Add the details that quietly kill "generic" — competence by area, audience, currency/units. "Expert in marketing, beginner in finance" stops the AI both over-explaining and over-assuming. Units and language stop the small daily annoyances.

  Try this: Add lines like this to your profile:

  Prompt I'm expert-level in B2B marketing, a beginner in finance and law. Use EUR and the metric system. Answer in English. Copy

  The Setup That Makes AI Stop Being Generic → Tier 2 ↗ one-time setup
• Spin up one Project (or Custom GPT) for a recurring job, with its context pre-loaded. Stop re-pasting the same background. A Project remembers your style guide and product facts, so every task starts ahead.

  Try this: Make a "LinkedIn posts" project with your tone rules and three past posts loaded. Every new post starts in your voice.

  Project Memory in Claude ↗ one-time setup

Using AI more means trusting it more. These keep that trust earned, not assumed.

• Run the Human-AI Sandwich: think first, let AI generate, then you judge and edit. Outsourcing the framing and the judgment is exactly where AI makes you worse. Keep the bread (your thinking); let AI do the filling.

  Try this: Before: jot your own take in two lines. After it drafts, paste your take back and challenge it:

  Prompt Argue against this. What am I missing? Copy

  Working With AI, Not Through It ↗ keep doing it
• Make the AI show its uncertainty. A confident tone hides shaky claims. Asking it to flag doubt surfaces exactly what to check.

  Prompt Mark anything you're not fully sure about, and list what you'd verify before I rely on this. Copy

  keep doing it
• Notice when it's just agreeing with you — and push back. AI tilts toward flattery. If every idea you float is "great," you've got a cheerleader, not a thinking partner.

  Prompt Don't validate me. What's the strongest case that I'm wrong? Copy

  Your AI Agrees With You Too Much ↗ keep doing it

The jump from "I use AI sometimes" to "I'd be slower without it" happens here.

• Move three recurring tasks into AI for good. Occasional use stays a novelty. Wiring it into your weekly rhythm is what turns it into a tool you'd miss.

  Try this: Pick three — weekly update, meeting prep, inbox triage. Write a reusable prompt for each and actually reuse it.

  How to Actually Get Good at AI ↗ keep doing it
• Run the same task through two tools and notice the difference. You only learn which tool fits which job by comparing — ChatGPT, Claude, Gemini, and Copilot have real and different personalities.

  Try this: Give the same brief to two assistants. Whose structure, tone, and accuracy do you prefer? Keep the winner for that kind of job. If work only gives you one tool, run the comparison privately on a personal task — the instinct for what each tool does well transfers.

  AI assistants compared ↗ now and then
• Make the free-vs-paid call on your main tool: if you're getting real use out of it, pay for one — and check what the paid tier changes for your data. Once you lean on AI for real work, the free tier's limits start getting in your way. And if you can't get twenty dollars of value a month out of it, that's worth knowing too.

  ✓ Tick it when you've made a deliberate call — upgraded, or decided the free tier genuinely covers you — instead of drifting along on free by default.

  Try this: The honest test: if you can't get $20 a month of value out of AI, you probably don't need it yet — you might just want a better search engine, and that's free. If you can, the upgrade pays for itself in a week. What usually pushes serious users to pay: the free caps (message limits, the older and slower models) start biting, and paid or business tiers often give you more control over your data, like keeping your chats out of model training. Check your tool's current data settings — they change.

  one-time setup

At Level 2 you set up a profile that knows you. Now go one level up: stop re-explaining the same project, product, or set of rules every time you start something. Build the context once, point the AI at it, and every task starts from a running start.

• Build a one-page briefing document for your most-used recurring project — product facts, audience, tone rules, the constraints that never change — and pre-load it into a Claude Project, a Custom GPT, or whatever your work tool offers for pinned, reusable context. Every task you start without it makes the AI guess at things you already know. Loaded once, every task starts from the same solid brief — your facts, your audience, your rules.

  Try this: It doesn't need to be long. A few hundred words covering what the AI would need to act as your specialist, not a stranger off the street. Update it when the facts change.

  Prompt Here's everything you'd need to know about [project] to work on it like someone who's been here a year: [paste your brief]. Confirm what's still unclear, then we'll start. Copy

  Project Memory in Claude ↗ one-time setup
• Build a personal knowledge document — a dense file of who you are, how you work, where you're strong and weak, and your standing rules — and keep it maintained. A few custom-instruction lines were the starter kit. A real, maintained knowledge document is the compounding version — the gap between an AI that knows your name and one that works like a colleague who's been with you for years.

  ✓ Tick it when you have a file you'd confidently paste into any AI tool and trust it to represent you accurately.

  Try this: Don't write it from a blank page — let the AI interview you to pull it out (the Level 2 move, run longer and harder). Capture your role, an honest expertise map, your recurring projects, and the rules you never want broken.

  The Setup That Makes AI Stop Being Generic → Tier 3 ↗ one-time setup
• Structure a long prompt into labelled sections — context, task, constraints, output format — instead of one long paragraph. Models don't weight a wall of text evenly, and instructions buried in the middle get the least attention. Labelled sections tell the AI what each block is for, and you reuse the same skeleton every time.

  Try this: Plain headers work; so do XML-style tags, which models are trained to read cleanly. The point is that "what to do" and "what to work from" never blur together.

  Prompt <context>[what the AI needs to know]</context> <task>[exactly what you want done]</task> <constraints>[what to avoid, and the format you want back]</constraints> Copy

  Guide coming soon: A Practical Prompting Guide keep doing it

Not every task wants the same engine. A quick reformat and a board-level analysis need different models — and reaching for the wrong one leaves you either slow and overpaying or fast and quietly unreliable. Knowing which to grab is a Practitioner tell.

• Learn the difference between fast models and reasoning models — then consciously pick the right one for your next ten tasks. A reasoning model on a one-line email is slow and pricey; a fast model on a knotty analysis is confidently shaky. This is the lever most people don't know is there.

  ✓ Tick it when you can name at least two tasks where you'd now choose differently than you did a week ago.

  Try this: Fast (Claude Sonnet, GPT-5, Gemini Flash): email, summaries, reformatting, quick lookups. Thinking modes (Claude's extended thinking, GPT-5's thinking, Gemini's): multi-step analysis, weighing trade-offs, checking your own logic for holes.

  The Anthropic Ecosystem ↗ keep doing it
• Match the model to the budget, not just the task: use your strongest model to think — strategy, concepts, planning — then hand the finished plan to a cheaper, faster model to execute. The expensive judgment is in the plan. Once that's solid, a cheaper model follows precise instructions just fine — and the top tier can cost several times more per task than the small one, so paying premium prices to format a list is money lit on fire.

  ✓ Tick it when you've run one task split this way — planned on a strong model, executed on a cheaper one — and trust the result.

  Try this: The gap is bigger than it looks: on Anthropic's line the frontier model runs about 10× the cheapest one for the same work, and the biggest jump is the very first rung. Spend the premium where it changes the outcome — the thinking — and save it on turning that thinking into output.

  Prompt Turn this into a precise, step-by-step brief that a faster, cheaper model could execute without making judgment calls of its own: [task]. Copy

  Think Expensive, Execute Cheap ↗ keep doing it
• Build a two-tool research loop: have a chat model draft the research brief, run it through a research tool (Perplexity, ChatGPT Search, Gemini with grounding), then bring the sourced results back for analysis. A chat model alone invents citations and doesn't know what happened after its training cutoff. Fed a sharp, model-written brief, a research tool gives you sourced answers you can stand behind — the combination beats either tool alone.

  Try this: The move most people skip: don't type your question straight into Perplexity. Ask Claude or ChatGPT to write a detailed, multi-angle brief first, then paste that in.

  Prompt Help me write a research brief for [topic]. Give me 10–15 specific questions that attack it from different angles, which sources to prioritise, and the time period to focus on — formatted so I can paste it straight into a deep-research tool. Copy

  The Research Pipeline ↗ now and then
• Switch on extended thinking or deep-research mode for one problem that genuinely deserves it — a real decision, a thorny analysis, a document that has to hold up. Most tasks don't need it, and turning it on everywhere just burns time. The skill is judgment: spotting the handful of problems each week where the slow, expensive mode actually earns its keep.

  Try this: It's a toggle in the model, not a magic phrase. Turn on Claude's extended thinking or your tool's deep-research mode — the point isn't to use it more, it's to recognise the problem in front of you that's worth it.

  How to Actually Get Good at AI ↗ now and then

By now you can get good output. The Practitioner leap is using AI to sharpen your judgment instead of outsourcing it — building a partner that makes your thinking better, not just your typing faster.

• Write a standing "how to think with me" instruction for your main tool or project: surface my load-bearing assumptions, name the bet I'm making, and tell me what would have to be true for this to work. The leverage isn't in today's prompt — it's in the rules of engagement you set once and keep refining. A partner told to test your assumptions catches the weak point before you build on it.

  ✓ Tick it when your AI names an assumption you were relying on without realising it.

  Try this: Add it to your custom instructions or project, then pressure-test it on a real plan. You're calibrating, not decorating — expect to rewrite these rules a few times before they feel right.

  Prompt When I bring you an idea or plan, don't just polish it. Name the load-bearing assumptions I'm making, the bet I'm actually placing, and what would have to be true for this to work. Then tell me what you'd test first. Copy

  Working With AI, Not Through It ↗ keep doing it
• Before you commit real time to building or writing something, ask your AI partner to make the case that it shouldn't exist — or that there's a cheaper way to the same result. The highest-value thing an AI partner does isn't producing more — it's catching the bad idea early. It's always cheaper to realise something doesn't need to exist than to ship it and watch nobody use it.

  ✓ Tick it when you've killed or shrunk at least one idea because the case against it actually held up.

  Try this: Works for features, documents, processes, even hires — anything you're tempted to over-build. The discomfort when it lands a real hit is the value, not a bug.

  Prompt I'm about to put serious time into [idea]. Make the strongest case that I shouldn't — that it's over-engineered, solves a problem I don't really have, or has a cheaper path. Where's the waste? Copy

  Working With AI, Not Through It ↗ keep doing it
• Learn to spot both ways an AI partner fails you — the yes-man that loves everything, and the reflex cynic that tears everything down — and notice which one yours is doing right now. Dialling up the pushback to fix flattery just swaps one useless mode for the other. A partner that shreds every half-formed thought is as worthless as one that praises it; the skill is calibrating to the narrow band in between.

  ✓ Tick it when you can catch your AI in either mode mid-conversation and pull it back.

  Try this: Test it: float a deliberately half-baked idea. A yes-man calls it brilliant; a reflex cynic shreds it; a calibrated partner asks the one question that exposes whether it's worth pursuing.

  Your AI Agrees With You Too Much ↗ keep doing it

One-off prompts are tactics. A playbook is strategy. This is where you stop improvising every time and start building something you can lean on — and hand to someone else.

• Run one real task as a chain of separate passes — outline, draft, then a fresh critique pass that only critiques — instead of one prompt that tries to do everything at once. Ask a model to write and judge in the same breath and you get neither. Separating the passes — ideally letting a clean chat or a second model critique the first one's work — catches what self-editing never will.

  Try this: Outline first; approve it; draft section by section; then open a fresh chat and paste the draft in cold: "You're a skeptical reviewer seeing this for the first time — where does it fall apart?"

  Working With AI, Not Through It ↗ keep doing it
• Write your personal AI playbook: your five most-used prompts, when to reach for each, and the one thing each must always include. Your best prompts don't improve by living in your head; they improve by being written down, reused, and refined. A playbook turns scattered wins into a system you can lean on.

  Try this: Five candidates: weekly update, meeting prep, drafting feedback, explaining something complex, reviewing a document. One template per task, plus the line you must never drop.

  How to Actually Get Good at AI ↗ one-time setup
• Name two or three tasks where you've found AI makes your work worse — and write them into your playbook as AI-free zones. Knowing when not to reach for AI is as much a Practitioner skill as knowing when to. The trap is that AI often feels faster while actually slowing you down, so the honest audit beats the gut feel — and blind delegation quietly erodes the judgment you spent years building.

  ✓ Tick it when you can say, without hesitating, which tasks you'd never hand to AI — and you've told apart the ones you're just doing badly from the ones that are genuinely better by hand.

  Try this: When something feels slower with AI, diagnose it honestly: are you overcomplicating a simple job (a skill issue you can fix), or is the task just faster done yourself (a real AI-free zone)? Only the second belongs on the list. Usual suspects: anything needing your unfiltered personal voice, relationship-sensitive messages, work where checking the output takes longer than doing it.

  Code vs. Automation vs. AI ↗ one-time setup

Teaching is the fastest way to find the holes in your own understanding. It's also pull, not push: one colleague who sees your setup actually working does more for the people around you than any mandate. Bring one along.

• Help one colleague set up their AI tool — custom instructions, memory, and at least one reusable prompt or Project. Walking someone through the setup you built exposes every part you only half-understand, and turns you into the person who makes others better — not just the one who's good alone.

  Try this: Run the "let the AI interview you" prompt together with someone who doesn't know where to start. The whole onboarding takes about half an hour.

  Let the AI Interview You ↗ one-time setup
• Share one prompt or workflow from your playbook with someone who'd benefit — and add one sentence on why it works. Choosing which prompt to share, and saying in one line why it works, is the fastest way to find where your own understanding is thinner than you thought.

  Try this: It doesn't need to be a presentation — drop it in a message. The act of picking one and explaining it forces you to understand what you've been doing on autopilot.

  How to Actually Get Good at AI ↗ now and then