How I use AI

A caveat on freshness: I wrote this in early 2026. Models keep getting better, and each generation needs less hand-holding — you can often skip the more prescriptive techniques below and just say what you want. If you're reading this much later, treat it as a snapshot: the mindset (context, conversation, iteration) should age well; the specific tactics will matter less and less.

I've used AI tools almost every day for several years — thousands of conversations across Claude, ChatGPT, and Gemini. In that time I've landed on a small number of techniques that I use over and over again. None of them are complicated. But the gap between someone who uses AI well and someone who uses it badly is enormous, and most of that gap is about how you ask, not about which model you're on or what clever hack you've read about on Twitter.

Most people use AI like Google: type a question, get an answer, close the tab. That works, but it leaves about 80% of the value on the floor. This post is my attempt to show what the other 80% looks like.

The examples below are real prompts I sent, or lightly-edited versions of them. I haven't invented anything for illustration.

The single most important thing

AI is not a search engine. It is a colleague you can talk to.

Everything else flows from that idea. A search engine wants a keyword. A colleague wants to understand your situation, your constraints, and what you're actually trying to accomplish. If you brief AI the way you'd brief a thoughtful new hire — context, constraints, what "good" looks like, what to avoid — you'll get dramatically better results.

Give it context (the 5-minute upgrade)

This is the single change that will most improve your results. Before you ask your question, spend 30 seconds saying who you are and what the situation is.

Google-style:

"Microsoft Dynamics licensing explained"

How I'd ask it:

Please explain to me how Microsoft Dynamics licensing works based on end users.

Assume I have:

• 10 full time customer service agents, each with their own device
• 10 people managing field service
• 100 technicians who work part time and share 20 mobile devices between them

The second version gets you an answer that's actually about your situation. The first gets you a generic summary you'd have to translate yourself.

When I ask about a system migration, I describe the current state and the target state, include the messy real details, mark what I'm unsure about ("an ECS cluster(?)"), and say what format I want the answer in ("a table"). AI won't judge you for uncertainty — flagging it tells the model where to push back.

A few related habits:

You don't need polish. You need information. My prompts are full of typos and I don't fix them. AI isn't marking you on grammar. Don't spend 10 minutes polishing a prompt — spend 2 minutes getting the facts down.

Tell it who you are. AI will answer a cybersecurity question differently for a CTO than for a new hire. "I am a very experienced linux sysadmin" and "I am not a former software engineer" produce completely different — and correctly calibrated — answers to the same question.

Tell it what NOT to do. Most people only say what they want. Explicit exclusions prevent the "cover all bases" mode that produces generic, bloated output:

"I don't want comparisons. I just want to know what the companies do and why Elixir was a good choice."

"I don't care about controversies. Only scientific findings."

Iterate instead of starting over

Do not treat each message as a standalone query. The AI remembers everything earlier in the conversation. Use that.

When the first answer isn't right, don't rewrite your original prompt. Just point at what's wrong:

• "Shorten again"
• "No, they would be more explicit. This is too general."
• "This doesn't flow: [paste the paragraph]"
• "Make it such that no one can pick holes in what I'm saying."

You don't need sentences. AI just needs information. Type faster, think less about phrasing.

This is also how I learn. Here's a real sequence of mine on Bitcoin perpetual futures (I knew nothing going in):

1. "How does a Bitcoin perpetual future work?"
2. "I don't understand why there's an interest rate in the funding, if it's being settled every few hours"
3. "So if the price goes up then those that are long have their contract price increase by the amount the price has gone up?"
4. "Why wouldn't I just buy spot?"
5. "So if you buy spot and short perp do you take any risk?"
6. "What mechanism keeps the perp price in line with spot, and why might that mechanism not work?"

By the end I understood perpetual futures better than I would have from any blog post, because each follow-up targeted my specific confusion. You cannot do this with Google. This is the killer feature of AI, and most people never use it.

When I want to actually think rather than be told the answer, I ask for graduated help: "Please give me the smallest hint possible." Then: "Another tiny hint." And eventually, if needed: "Ok forget it just tell me the answer."

Let AI interview you

This is the technique that most surprises people when I show them. Instead of trying to write a perfect prompt up front, ask AI to extract the information from you:

Ask me one question at a time until I say STOP. When I say STOP, output [the thing I want].

Use it when you're writing a document and want to make sure you haven't left anything out; when you're writing a PRD and don't want to fight through a blank page; when you have a vague sense of a problem but haven't pinned down what you want to do.

For software projects I use a two-part version constantly:

Ask me one question at a time so we can develop a thorough, step-by-step spec for this idea. Each question should build on my previous answers, and our end goal is to have a detailed specification I can hand off to a developer. Let's do this iteratively and dig into every relevant detail. Remember, only one question at a time.

Then, when you've answered enough: "Now compile our findings into a comprehensive, developer-ready specification."

Why does the interview work? You know 80% of what needs to go in a document. AI knows the 20% you forgot. The interview catches the gaps — and it kills blank-page paralysis.

If you'd rather react to options than answer open questions, flip it: "Give me 5 different ideas and then I will pick one."

Writing with AI (most people do this wrong)

If your writing workflow is "ask AI to write something, copy-paste it, done" — you're doing the thing that makes AI output sound like AI output. Everyone can tell. Nobody reads it. Here's what I do instead.

Never ask AI to write from nothing. Give it your draft. Every writing task I send looks like "Improve this email:", "Lightly edit this:", or "Complete this email I started to write:" followed by my own words. Your half-finished draft contains your voice, your priorities, your opinions. AI finishes the thought — it doesn't replace the thought.

"Lightly edit" vs. "improve" vs. "rewrite". These produce very different results. "Lightly edit" means: my words, just cleaner — use it where your voice matters. "Improve" allows rewritten sentences but keeps intent. "Rewrite as…" is a full rework. Most people default to "improve" when they should say "lightly edit".

Tell it who will read it. "Write this as an engaging article for a smart, thoughtful and ambitious 12-year-old" and "I am sending this email to experienced developers who have no knowledge of Wordpress" produce appropriately different output.

Banish AI slop with explicit constraints. The word "dense" is magic — it suppresses the padding:

"I need a highly structured and very dense report."

"This is too wordy. Make it dense by compressing the language, removing fluff."

"Do NOT add flowery language or pleasantries."

Ask for specific changes, not general feedback. "Can you review my article?" gets you vague thoughts you can't act on. Instead:

Review each section in this draft and give me a list of specific changes that would make it better. Please make each change specific so that I can easily go through the list and make the changes atomically.

Do flow and fine-editing in separate passes. Ask for structural suggestions first ("just tell me which paragraphs to move around"), apply them, then do copy-editing. And tell the AI your changes will be applied sequentially, so each suggestion can assume the prior one is done.

AI as a thinking partner

The most underused mode. Most people never ask AI to push back.

I don't ask "is this good?" — it will say yes. I ask it to find what's wrong:

"Which sections might critics pick on?"

"If they were going to quote one part of it to show how pointless my work is, which part would they single out?"

"Roast my thesis in a tweet."

This is red-teaming your own work. It's much harder to do alone than with AI. Two of my most-used phrases:

"Steel man this and then explain the shortcomings."

"What questions will the person ask me after they read this doc?"

The first protects me from my own confirmation bias. The second lets me answer objections in advance of a meeting.

And before forwarding an article or repeating a claim, I paste it in with "Is this true?" or "What are the flaws in this report?" Thirty seconds of checking has saved me from looking foolish more than once.

Power moves

Negotiate the format before the answer. When I ask for something structured, I don't let AI just start:

Please make a table that explains the sequence of steps we should take. Before you make the table, please suggest the columns the table will have, so I can suggest changes.

It's much cheaper to iterate on the columns of an empty table than on the rows of a full one.

Filter with explicit criteria. When researching options: "Each of the items in the list must satisfy ALL of these criteria: …" The capital "ALL" signals a hard filter, not a nice-to-have.

Chain AI tools. The quality of a Deep Research run or a coding-agent session depends entirely on the prompt. So I use Claude to write the prompt, via the interview technique:

Please help me to write a prompt for a deep research agent. Ask me one question at a time until I say STOP. When I say stop give me the prompt that I can copy and paste.

You can also write prompts for other people — one person on a team who's good at prompting can multiply the whole team's effectiveness.

Ask AI to name the thing. When I don't know what something is called: "What is the category of software I should Google, to answer this question?" AI is very good at translating "the thing I vaguely want" into "the name of the industry that sells it".

Paste the criticism along with the work. When acting on feedback, don't paraphrase it — paste the original work and the verbatim feedback. AI does far better with both sides of the argument.

Name the framework. AI knows Nonviolent Communication, Difficult Conversations, SWOT, MECE, RICE, Jobs-to-be-Done, and hundreds of others. Naming one saves you a thousand words of explanation.

Dump raw data and let AI process it. Some of my highest-value uses treat AI as a processing engine: a full email thread ("summarize the bugs so I can hand them to a coding agent"), a complete database schema, raw JSON, a meeting transcript in another language. You don't need to pre-process. Paste the whole thing — AI is better at finding signal in noise than you are at guessing what to include. I've pasted 85,000 characters of conference talk descriptions with the instruction "Categorize these into themes and give me a table."

Let AI decide what questions to ask about your data. When you have data but don't know what's interesting about it:

"I want to know how long my son is spending on Anki sessions each day, and how much time per card. I don't know exactly what I should be analysing, like percentiles or mean or whatever. Please formulate some specific questions and give me the answers. The more questions the better."

Use photos and screenshots as input. Photograph a confusing error screen, a foreign-language menu, a whiteboard, a math problem, a prescription. "Explain this", "OCR please", "what should I click?". It sounds minor until you start doing it; then your phone becomes a different tool.

Recursive decomposition for planning. Don't just ask for a plan — ask AI to break it down, then break it down again, then review whether the steps are right-sized. Then: "Make a todo.md that I can use as a checklist." The output goes straight to a coding agent, or to me.

Use the right tool for the job

I use several AI tools, because they're good at different things — Claude for conversation, writing, and stress-testing arguments; ChatGPT for Deep Research, voice, image generation, and data analysis with file uploads; Gemini for Deep Research and for ingesting YouTube videos and podcast audio; coding agents (Claude Code and friends) for execution on real codebases.

Within each tool, match the model to the task: reasoning models for hard analytical problems, fast models for quick lookups, research mode for anything that needs current information. Using a heavy reasoning model for "what is a MOSFET?" is wasteful; using a lightweight model to evaluate a major strategic decision is penny-wise and pound-foolish.

Two capabilities most people haven't tried:

Feed it video and audio directly. Gemini accepts YouTube URLs and MP3s. Don't ask for "a summary" — ask for something specific: "I just started learning piano. I need this turned into very detailed instructions." Or: "Create a dense, detailed, actionable, no-fluff article based on what Simon said in this video." The medium changes; the techniques don't.

Voice mode for learning. I've had extended voice conversations about Ito's Lemma and abstract algebra while walking. It's like a tutor who never gets impatient. I've also used it for live translation at an event held in a language I don't speak.

If you catch yourself writing the same instructions repeatedly, make a Project (a persistent context with its own system prompt). I have about 30 — one turns every message into an Anki-ready flashcard; one always outputs single-file Python scripts with inline dependencies; one applies my formatting preferences to any braindump. Rule of thumb: any instruction you've written more than twice should become a Project.

What most people get wrong

Anti-pattern	What to do instead
One-shot query, then close the tab	Use follow-ups. The conversation is the tool.
Asking "please write X" from scratch	Write a rough draft first, then "improve this"
Spending 10 minutes polishing a prompt	Typos don't matter. Write fast, iterate more.
Treating the first answer as final	Push back, ask for alternatives, say what's wrong
Giving generic tasks ("summarise this")	Say who the audience is and what you'll do with it
Re-explaining the task on every message	AI remembers. Just say "shorten" or "more specific"
Asking AI to agree with you	Ask it to find flaws, steel-man the opposite
"Is this good?"	"Which parts would a critic pick on?"
Pre-processing data before pasting it	Paste the raw data. AI extracts signal better than you guess.
Using the same model for everything	Reasoning models for hard problems, fast models for lookups
Only saying what you want	Also say what you don't want
Typing when you could photograph	Screenshots and photos are input

Closing thought

The gap between a mediocre AI user and a great one isn't about clever prompt-engineering tricks. It's three simple shifts:

1. Treat it like a colleague. Give context, share constraints, explain who will read the output, push back when the first answer isn't right.
2. Have a conversation, not a query. The real value is in the back-and-forth — the follow-ups, the refinements, the things you didn't know you didn't know.
3. Feed it everything. Stop pre-filtering. Paste the raw data, upload the messy spreadsheet, photograph the error screen.

If you do only these three things, you will immediately be in the top 20% of AI users in any organisation. Everything else is a refinement of those ideas.