Working with AI

I spend a lot of my time working with AI on ideas and formal documents. A few friends mentioned they'd only ever used it like a smarter Google, quick answers and the odd bit of guidance, and asked how I got more out of it.

So I pulled it together: the rules I already keep in my custom instructions, the habits I knew I relied on, and then a pass where I had the AI go back over my previous conversations to surface the other processes I use consistently without thinking about them. This is the result: a beginner's guide to using AI for writing and thinking, written for them. They've since pushed me to make it available more widely, so here it is.

It's deliberately casual and practical. Take what's useful, leave the rest.


Working With AI: A Practical Framework for Research, Writing and Analysis

Quick Start

If you read nothing else, these are the core habits. The rest explains each one and why it matters.

• It can be confidently wrong, so check anything that carries a consequence.
• Give it the context, and tell it how to work, before you ask for the work.
• Build anything long one section at a time, not in a single pass.
• When a chat gets long and messy, summarise where you have got to and start a fresh one.
• Ask it to challenge your thinking, and to read your work as your reader would.
• When it gets something wrong, point at the exact part and name the kind of error, rather than starting over.
• On things that matter, check the sources and get a second opinion from another model, given the same task cold.
• You stay the one making the decisions.

Not sure which model to reach for? The annex has short notes on how the main ones behave along with some worked examples. 

Introduction

You may already use AI for general searches or insights, but it is also a powerful tool for drafting and developing ideas: a policy document, a piece of research, a complex email. This is a short set of habits for doing that well.

One thing to know up front: while the models keep improving, they still hallucinate. A model writes with the same confidence whether it is right or wrong, and can invent facts, figures and even sources with no sign it has done so. Treat what it gives you as a draft to check, not an answer to rely on, and verify anything that carries a consequence.

This framework is built through practice, not theory, so take what is useful and leave the rest. It reflects the tools as they stand in May 2026; the tool-specific advice dates fastest, so it sits separately at the end.

One principle governs everything that follows: manage the AI to work to your style and approach, in a controlled and logical way.

Intent of the document

There is a difference between a tool that produces some work and one that produces your work. Used casually, AI does the first: a plausible, generic draft that could have come from anyone, taken largely on faith. Used the way described here, it does the second. The thinking stays yours, so what you produce carries your judgement, reads as your own, and is something you can stand behind.

Governing principle

Getting the best out of it comes down to one idea: you do the thinking.

The AI does the work you point it at — gathering sources, summarising, adding information, collating your ideas, drafting, formalising what you have given it. It does not make your decisions or stand in for your judgement.

Two things keep it in that role: being deliberate about what you tell it, and not letting the conversation get so cluttered that you lose track of what matters. When this goes wrong, it usually traces to one of those slipping — you let the tool decide something without enough thought, or the chat grew so long the model could no longer tell what was relevant, even losing the first things you told it. A model can only hold so much of a conversation in view at once; that limit is its context window.

Everything below is practical habits for keeping you in charge, with the reasoning behind each.

1. Set it up before you start

How you set up at the start decides how much you fix later.

1. Start with the context. The model works only with what you give it, so before any real work, set out what it cannot guess:

• what you are trying to do
• what a good result looks like
• what you already know: facts, earlier work, settled decisions
• what you do not yet know
• where you are confident, and where you are not
• the assumptions you are making, said out loud so the model can question them rather than quietly accept them

If any of it is provisional, say so, or the model treats everything as equally solid.

2. Tell it how to work with you. A model's default manner is rarely right for the task, so say up front what you need. You might want it to:

• challenge you and look for the weaknesses in your thinking
• offer a counter-argument to everything you say
• carry out instructions without pushing back
• tighten your writing without changing what it says
• think something through alongside you
• search rather than rely on memory, when facts matter

If its manner drifts, reset it. Standing preferences belong in your saved setup (section 3); this is the adjustment for the task in front of you.

Build this up over time. It might start as a single line, but you will spot rules you want followed every time. Add those to your standing rules (section 3) and you gradually build your own guide for how the tool works with you — a well-developed one runs to a page or two.

3. Be careful with sensitive information. Check how a platform stores and uses your data before you put anything confidential, contractual, regulatory or personal into a chat. When the real data matters:

• give it dummy data in the same format as the real thing
• bring the real figures in only once you are back in your own secure environment

You give it the dummy version, it builds what you need, and you take that back to apply to the real thing — so it solves your problem without ever seeing the real content.

For example: you need a formula for a spreadsheet whose contents are sensitive. Copy the column headings into a clean sheet, add a few rows of dummy data in the same format, and tell the model what you are trying to work out, noting the real sheet runs to around 800 rows. It writes the formula against your sample; you take that back to the real sheet in your own environment and apply it. You get exactly what you needed, and the model never saw a single real figure.

2. Stay in charge as you go

Once you are working, it is easy to let the model take over. These habits keep you in the driving seat.

1. Settle the structure before you draft. Work out what the document is made of before you ask for content.

• if you have a required format or a clear sense of the shape, give it to the model
• if not, ask it to suggest headings, and adapt them until they fit

With the skeleton settled, you fill in a shape rather than discovering it as you type.

2. Draft sections in the order that helps most, not the order they are read in. Executive summaries and introductions sit first on the page but are easier to write last, once you know what they introduce. Build the parts in whatever sequence makes each easiest to get right, then arrange them for the reader at the end.
3. Work one section at a time. Quality falls away over a long single pass, and you lose the chance to correct course. Build it up section by section.
4. Ask for the strongest one or two points, not everything. A long answer is harder to use than a focused one. Ask for the sharpest one or two points first, and draw out more only if you need it.
5. Summarise each section before moving on. When you finish a section, ask for a short summary of where it landed and which assumptions it now rests on. After a few rounds of edits, that summary is the definitive version — the one that counts. The chat by now is cluttered with old drafts; the summary tells the model which one is real, so the next section builds on what you agreed.
6. Use the model to check for bias and contradiction. This is a genuinely new strength: a tireless, detached reader that goes back over the whole piece in seconds and shows where the thinking is weak. Every so often, ask it to read the work back and flag any bias, lazy reasoning, or places where two parts contradict — rather than quietly fixing them. A prompt that works:

Please read through the work we have just completed. Identify where we might have been subject to bias, heuristics or dissonance in the thinking and the work.

Seeing the problems listed beats having them silently smoothed over.

7. Read it back through your reader's eyes. Before finishing anything written for a particular audience, ask the model to read it as that reader and say what they would expect that is not there. Sweep several readers at once to find gaps quickly, then take one at a time for a sharper read:

Read this document as each of the following, and tell me whether anything is missing that they would expect to be here: a CEO, a client, an investor, a regulator, a member of the public.

8. If it matters, sleep on it. For anything that carries weight, leave the final version for a while and reread it yourself before it goes out.

3. Don't start from scratch every time

A good single conversation helps little if you rebuild everything next time.

1. Set up the standing context the tool remembers. Most tools store background that carries across conversations — custom instructions, saved projects or profiles. Use it for what rarely changes:

• your role and what you work on
• who you usually write for
• how you like things written
• any methods or frameworks you use often

Set it up once and you stop explaining yourself from scratch.

2. Write down what works so you only work it out once. When a way of framing or approaching something proves its worth, capture it: ask the model to put it into clear words, then paste that into your standing context. Letting it write the definition is quicker than doing it yourself, and saves re-deriving the same thing every session.
3. Start a fresh conversation when one gets cluttered, and hand over properly. As a chat grows it fills with noise and dead ends and presses against the context window, so the model loses the thread and the work gets worse rather than better. Don't reopen things cold: ask the cluttered chat to summarise in depth what you have been doing, what you are trying to achieve, and where you have got to, then paste that into a fresh conversation and carry on. Keep the old chat for reference.
4. Keep one running list for anything with many moving parts. When a piece pulls in several people, sources or requirements, keep a single visible list of:

• what is settled
• what is still open or blocked
• what depends on what else

You carry that list from one conversation to the next, rather than the conversations themselves.

4. Pick the right tool, and use more than one

From here the approaches grow more advanced; come to them as the work gets larger or more complex, not on day one. No single tool is best at everything.

1. Use the right kind of tool for the task. A chat model — the kind you hold a back-and-forth with — is best for shaping ideas, drafting and testing your thinking. It is poor at searching large document sets or scanning widely across sources, which other tools do better. Match the tool to the task; see the annex.
2. Different tools have different strengths and habits. They are built and trained differently, so the same question gets different treatment from each, with its own tendencies and blind spots. That is part of why using more than one pays off.
3. For important work, get a second and third opinion, then decide for yourself. Give the same task to more than one model separately, look across the answers, and decide on your own reasoning. No single tool should be the one that decides.
4. Give each model the same task, cleanly and identically — the payoff is clarity. Run the same prompt past each one without telling it what you hope to hear or what another model said, and ask what is weak rather than whether your draft is strong. Where they agree, you have your strongest points; where they diverge, or each catches something different, you have surfaced the weak spots and the gaps you missed. That only holds when they come at it cold — agreement you nudged them into is just your own view handed back.

5. Get the sourcing right

Good answers depend on good sources, and the model will not police that for you.

1. Judge the quality of each source, not just whether it fits. A model reaches for a weak, fringe or partisan source as readily as an authoritative one, so judging whether a source is any good is on you.
2. When you know the authoritative source, tell it to prefer that. If you know where the reliable answer lives, name those sources and tell the model to treat them as primary — a government department or agency for policy, law or official figures; the official club or league site for a team's results. For research, consider Google Scholar, PubMed and Web of Science as well. A prompt that works:

While researching, treat primary sources as your main material, for example official sites ending in .gov. If you draw on a secondary source, say why and what it adds.

3. Check every claim against its source — don't take the citation on trust. For any claim the model makes, find the in-text reference and check the actual link or reference behind it, every one. A confident citation is not the same as a correct one. Judge the source too: is it one you would have reached for? If not, ask why it chose that one and point it to the source you trust.

6. Steer what comes back

Setting the work up well is half of it. The other half is how you handle what the model gives you: interrogating it rather than accepting it, and steering when it drifts.

1. Query the output. When a draft or answer is not right, don't quietly accept it or start over. Point at the exact part that is wrong and have the model redo just that, not the whole thing. Name the kind of error, not only the instance — “this is too long” or “this drifts from the rest” teaches it more than “try again.” Ask it to defend a claim you doubt, and to argue the opposite as a test. If a whole direction is wrong, say so plainly and restate what you want rather than trying to rescue it.

From drafting this document: a proposed structure for these examples misread what was wanted, so the direction was rejected outright and the intended shape restated, rather than the wrong version tweaked into place.

2. Stop drift, and reset when needed. If the model starts to wander, repeat itself, or ignore something you told it earlier, pull it back by restating the goal and where you have got to. Watch across turns, not just one answer: a single poor reply needs only re-prompting, but a pattern slipping over several turns — the same mistake creeping back, or a style you have already corrected returning — is drift, and the fix is to summarise and start a fresh chat (section 3).

From drafting this document: a clipped, choppy sentence style began creeping into several replies; once named as drifting from the document's voice, it was corrected across the rest of the draft.

Examples 

Simple example

You don't need a big report for this to help. Take a cover letter for a job. Even here, the first few habits make the difference.

1. Tell the model what the letter is for, who will read it, and what you want them to think or do by the end.
2. Give it your CV, and paste in the job description.
3. Ask for a simple structure — what to cover and in what order — before any wording.
4. Work through it a part at a time.

Even those few steps get a far better result than asking for the finished letter cold.

Fuller example

Suppose you have a piece that has to land with a particular reader: a proposal, a report, a funding application, a careful email.

1. Tell the model what it is for, who will read it, what is settled and what is open.
2. Ask for the two strongest ways in and a rough outline, not the finished words.
3. Work through it a section at a time.
4. After each section, ask for a summary of where it landed and what the rest now rests on.
5. Leave the opening and any summary until last, once everything beneath them is settled.
6. Ask it to read the work back and flag any bias, heuristics or dissonance.
7. Ask it to read it again as each of your key readers, and say what they would expect that is missing.
8. If the stakes warrant it, cross-check the draft against another model or two (see the annex).

Working this way, the piece comes together cleanly, you can change direction while there is still time, and a record of how you got there sits in the conversation.

Advanced example

A more involved workflow for a substantial piece, once you are comfortable with the basics, running across several tools with each doing what it does best. A research report could go like this:

1. Sketch the outline in Claude, so you know what the piece needs to cover.
2. Have Claude write a research query for Gemini's deep research, and run it to produce a long, sourced paper (usually fifteen to twenty pages). Feed the same query to Claude in a fresh chat too, and compare its response to Gemini's.
3. Bring the paper back into Claude and have it summarise the findings.
4. Draft the report in Claude, section by section, from that base.
5. Give the draft to Gemini and ChatGPT separately for assessment, each coming at it cold. Give it to Claude itself in a fresh chat too, cold, as another independent read.
6. Feed the assessments back into Claude, decide what to act on, and tweak. Repeat if it needs another pass.
7. Upload the draft to NotebookLM and listen to the podcast-style overview, to hear it from the outside and judge whether it lands.
8. Claude cannot read audio, so to bring the podcast back you first need a transcript. Upload the audio to a fresh NotebookLM notebook and click the file to produce one. Give that transcript to Claude for a final round of validation and tweaks.
9. Finalise the report.

Annex: tool-specific notes

This part dates faster than the rest. The habits above should hold; the specific tools and quirks below will need updating.

How the main models behave, as of May 2026. These are my own observations rather than hard fact, but each model has its own habits, and knowing them helps you read its answers.

• Claude^* reasons well, takes positions and will push back. It is smart but pedantic — it can make mountains out of molehills — yet it is more reliable than ChatGPT and more thorough than the others. It can be cautious on sensitive subjects and wordy unless told to be brief. Two practical quirks: it runs out of context window surprisingly fast, and it “forgets” — so when you start a new chat or protocol, copy-paste your instructions back in rather than assuming it still has them.
• ChatGPT tends to the safe, middle answer and hedges when you want it to commit, while still sounding confident when it isn't — so ask it to take a clear position. Its critiques have sharpened lately, though it will always find something to improve and can overwork a point or invent one just to have something to say, so weigh what it flags. It has also recently added a deep research function.
• Gemini leans towards telling you what you want to hear and praising your question, so discount the enthusiasm, though it is strong on well-sourced information.
• Perplexity is built around an LLM for search, so it suits what you would normally Google, though it handles ordinary chat too. It was first to show sources by default, which helps, but check what it cites — it does not always lead with the most authoritative source.

* A more capable Anthropic model, Claude Fable, sits above these but was still withheld at the time of editing. If it becomes available it may change some of the advice here — particularly these model-specific notes and where Claude fits against the others.

One useful trick: give an over-agreeable model two opposing versions of a question so it cannot just agree; give a fence-sitter clear options to choose between; give a cautious model explicit permission to disagree.

For research and fact-finding, use whichever tool shows exactly where its information came from, since that work depends on checking sources. Several tools now offer a “deep research” mode that searches the web and returns a sourced write-up rather than a normal reply. As of May 2026, the strongest for this is Gemini.

One tool I would single out, as a personal preference rather than a rule, is NotebookLM. Feed your work in and it produces a spoken, podcast-style discussion of it — a genuinely different way to catch what is off. Hearing it read back, you notice where it fixates on the wrong things, and if the points you cared about are buried too deep they simply never come up. It won't be to everyone's taste, but it is worth trying once to see if it clicks for you.