The difference is you. You told it what you wanted before it started thinking.

Lawyers have a name for this. Leading the witness, and it’s banned in court because it works. Most people writing AI prompts for research or decisions do it without noticing. Your assumptions ride in on the phrasing, the model picks them up, and what comes back is your own opinion with better grammar and more confidence than you had going in.

Below are eight prompts that strip your bias out of the question before the model answers it. Plus a swap list for the phrasings that cause the most damage.

Why is AI suddenly agreeing with everything you say?

AI models learn from human ratings, and people rate agreeable answers higher than accurate ones, so agreement gets rewarded in training. OpenAI shipped a GPT-4o update on 25 April 2025 and pulled it four days later because it had tipped into flattery, endorsing plans it should have questioned. Asking a model to be more honest doesn’t fix this. Changing the wording of your question does.

Models learn from human ratings, and humans rate agreeable answers higher than accurate ones. Agreement gets rewarded. OpenAI rolled back a GPT-4o update on 29 April 2025 after four days because it had tipped into flattery, endorsing plans it should have questioned.

These prompts don’t ask the model to be honest. Asking doesn’t work. They change the input so there’s less of your opinion in there for it to copy back at you.

Prompt 1: How do I tell if my prompt is leading the AI?

Paste your prompt into a fresh chat and ask the model to analyse the prompt itself before answering it. Tell it to quote the exact words that signal your preferred answer, any assumption you’ve stated as fact, and any option your phrasing rules out. Most prompts carry two or three biasing phrases the writer never noticed typing. Run it on the questions that gave you answers you liked.

What it does: Shows you exactly which words in your own prompt are steering the answer, before the model answers it.

When to use it: Any time you’re asking AI something where you already suspect what you want the answer to be.

The Prompt:

Here is a prompt I’m about to send you: [PASTE YOUR PROMPT]

Do not answer it yet. Analyse the prompt itself first and identify three things. One: any wording that signals the answer I’m hoping for. Two: any assumption I’ve stated as fact without evidence. Three: any option I’ve ruled out by how I framed the question. Quote my exact words for each. Then tell me in one sentence what answer this prompt is pushing you towards, and how confident you are that a neutral version would get a different answer.

How to use it:

1. Write your real prompt as you normally would, without editing it to look better.

2. Paste it into the audit above and send.

3. Read the quoted words. Those are your fingerprints on the answer.

Example input: Here is a prompt I’m about to send you: “Our churn is high because the onboarding is too long. What’s the fastest way to shorten onboarding without hurting activation?”

What you’ll get: A short breakdown showing you asserted causation you haven’t proved (long onboarding causes churn), asked for speed rather than correctness, and excluded every other explanation for churn by naming one. Usually two or three quoted phrases you didn’t notice writing.

Advanced note: Run this on prompts you’ve already sent and got answers you liked. Those are the ones worth checking.

What you’re missing

That audit catches the bias in one prompt. It won’t catch the bias in your thinking, which is the bigger problem, and it does nothing about a long conversation where the model has quietly stopped arguing with you.

The remaining seven prompts handle that.

Prompt 2, Neutral rewrite: Rebuilds your loaded prompt into a version that doesn’t point at an answer, and tells you what it stripped out.

Prompt 3, Premise flip test: Runs your question twice on opposite assumptions so you can measure how much of the answer was ever about evidence.

Prompt 4, Anonymous review: Gets an assessment of your work with your ownership of it removed, which is where the flattery mostly comes from.

Prompt 5, Hostile reviewer: Puts a named opponent in the room who wants your plan killed and has read it properly.

Prompt 6, Steelman the side you rejected: Argues the position you dismissed at full strength, then names the one fact that would decide it.

Prompt 7, Confidence and source split: Forces the model to separate what it knows from what it inferred from your wording, claim by claim, with numbers.

Prompt 8, Conversation drift audit: Finds the exact message where the model stopped pushing back and started agreeing with everything.

You also get the swap list at the end. Eight phrasings that bias answers, with the neutral version of each, so you stop writing leading questions in the first place.

Prompt 1 on its own will improve one decision this week. The rest of it changes how you ask questions permanently.

Read more

Everyone has the same models. The results are nothing like the same.

You’ve used AI for a while now. It works. You ask, you get an answer back in about forty seconds, and it’s fine.

Fine is where most people stop, and understandably so. An analysis, a plan, a summary of forty pages you didn’t want to read, a draft of something: what comes back is competent, unremarkable and roughly what the person next to you got when they asked something similar.

That gap is the subject of this newsletter. Not what AI can do, which you already know, but what separates the people getting real work out of it from the people getting a slightly faster version of average.

It usually comes down to structure. One prompt gets you an answer. A sequence of prompts, each one working on what the last one produced, gets you something you can actually use.

What this is

AI Prompt Hackers is read by more than 32,000 people who use AI for real work. Founders, analysts, marketers, consultants, writers and a decent number of people who just got tired of mediocre output.

Twice a week you get prompts you can copy and use the same day. No theory sections. No explanations of how large language models work. Every prompt written out in full, with the placeholders marked, an example of what to put in them and a description of what comes back.

Getting through a long document and coming out with something you can act on. Turning a messy spreadsheet into an answer you’d be willing to defend. Pressure-testing a decision before you commit to it. Taking a task you rebuild every month and turning it into something you run instead. Some issues are about writing too.

If it can’t be used within ten minutes of reading, it doesn’t go in.

Free

Two issues a week, and the first prompt in every one.

That first prompt is complete and working, with the full instructions rather than a sample of them. Plenty of people run on those alone and get real value from it.

Paid

$7 a month, or $49 a year.

Every prompt in every issue, which is usually seven to nine more per article, plus the parts that only work as a sequence.

Then the two skill packs, which are the reason most people upgrade. Each covers one area properly rather than sampling several, and they’re the two areas where a connected sequence makes the biggest difference.

Sell What You Know is four connected skills for building a solo business. Working out which of your skills somebody actually pays for, and demanding evidence rather than enthusiasm. Researching your buyer properly, including what they’ve already tried and what it taught them to distrust. Writing a sales page and then reading it back as a sceptic. And building a delivery system that ends with capacity maths, which sometimes tells you the business doesn’t work at your price.

Write Like Yourself is four connected skills for anyone who publishes. It reads three to five things you’ve written and builds a specification of your actual voice, down to your sentence lengths and punctuation habits. Then a skill that turns a rough idea into a tested brief, one that drafts the piece and then attacks its own draft, and one that gets everything out of it after publication.

Both packs come in three formats: Claude Skills, Gemini Gems and Custom GPTs.

Both include full worked examples so you can see the output before committing an afternoon to it.

Yours to download the moment you upgrade.

Power User

$99 a year.

Everything above, plus the skill library.

Full multi-step skills with worked examples, built to the same standard as the ones in the packs, with a new one added every month.

The library is where the rest of the ground gets covered. Research that grades its own sources and tells you where they disagree. Data analysis that hunts for the finding that’s true and misleading. Decisions that name the option you’re avoiding. Getting found, when you’ve built something good and nobody knows it exists.

A year of that is a substantial library of things you own and keep, and it’s the tier for people who’d rather have the systems than read about them.

What this isn’t

Worth saying plainly, because a lot of AI newsletters promise things they don’t deliver.

This isn’t news. If a model gets an update on Tuesday, you’ll read about it somewhere else first, and probably from someone who enjoys writing that more than I do.

It isn’t a course. There’s no curriculum and no sequence you have to follow. Every issue stands alone.

It isn’t going to do the thinking for you. The skills are deliberately confrontational about that. One will tell you your favourite business idea has no evidence behind it. One will do the capacity maths and show you the business doesn’t work at the price you picked. Another will name the exact paragraph where a reader gives up on your draft. That’s the useful part, and it isn’t always the enjoyable part.

And there’s no upsell waiting. No coaching, no cohort, no strategy call. Two issues a week, the packs, and the library if you want it.

Why I do this

Before stepping back from corporate life I was a CEO, a COO and a CTO, and I co-led a $200m management buyout, one of the biggest and fastest of its time. Most of that job was reading things I didn’t have time to read and making decisions on partial information, which is roughly what I use AI for now.

Since then I’ve watched people get remarkable results out of these tools while others get very little, often from the same model and a similar question. The difference is almost never intelligence. It’s method, which is a teachable thing, and more interesting to write about than another take on where AI is heading.

I read every reply. If something doesn’t work, or works better than you expected, tell me. Some of the best material here has come from someone using a prompt in a way I hadn’t thought of.

If you’re getting answers out of AI and suspect there’s a better one available, you’re in the right place.

Andy

P.s. One thing worth knowing

I use AI to help produce this newsletter. I’m not performing artisanal purity about the writing process, and I won’t pretend otherwise. What I bring is 40 years of judgment about what’s worth testing, what the results actually mean, and what’s worth your time. That’s not something AI can replace yet.

If you want to know exactly how I use it, read the Honest AI Manifesto.

Subscribe now

For anyone dropped into an unfamiliar field with a deadline. Works out what depth you actually need, separates the load-bearing ideas from the fashionable ones, then tests whether you understand it or just recognise it.

Packs come in three formats: Claude Skills, Gemini Gems and Custom GPTs, and include full worked examples so you can see the output before committing your to it.

Fast learning usually produces fluency without understanding. You pick up the vocabulary, read whatever is popular, and come out able to hold a conversation while unable to spot an error, judge a supplier or predict what happens next. That’s a worse position than knowing nothing, because it’s confident.

This skill starts by establishing which of four depths you actually need. Enough to judge, to hold a conversation, to decide, or to do the work. They’re different jobs, and people habitually aim at doing-depth for something they’ll never do, which turns five weeks into five months. Then it separates the load-bearing ideas, which everything else depends on and which are usually unfashionable because they were settled decades ago, from the heavily discussed recent material that dominates the available reading. It maps where competent experts actually disagree, which protects you from absorbing one camp’s position as the field’s consensus. And it tests understanding rather than recognition, since those feel identical from the inside.

It ends by naming what you still cannot do and what would take years, and it won’t shorten that section if you ask.

In the worked example, a new charity trustee was about to start a financial modelling course to sit on an investment committee. She needed to judge a portfolio, not build one. Five weeks instead of a year, and three hours of arithmetic in the middle of it surfaced a reporting error nobody had noticed in three years.

👇 For Power User subscribers only. Download here 👇

NOT ADVICE

The information is intended to be helpful, but is in no way a substitute for seeking professional advice for your specific situation or intent. This applies to business, financial, legal, health, medical, personal, or other matters discussed herein. Please read the full DISCLAIMER

For clarity, all content found in The Library is for your use only. Not for resale or redistribution in any form.

For anyone who rebuilds the same thing every month. Maps what actually happens rather than what you remember, cuts the parts nobody reads, then writes the prompts.

Packs come in three formats: Claude Skills, Gemini Gems and Custom GPTs, and include full worked examples so you can see the output before committing your to it.

The task you rebuild every month costs far more than you think. Nobody counts it, because each individual instance feels reasonable, and the price only becomes visible multiplied by twelve. Then most attempts to fix it go straight to automation, which is the wrong order.

This skill does three things first. It refuses to let you describe the task from memory, because memory produces the version where you were doing your job, with the waiting removed and the chasing forgotten. It asks who uses each part of the output, by name rather than by department, and deletes whatever has no name against it. And it sorts the automation boundary by what happens when a step goes wrong and nobody notices, rather than by how hard the step is, because the most tempting things to automate are often where errors hide best.

Then it writes the runbook and the actual prompts, with a check attached to every step that could fail invisibly, and a map saying what feeds what.

In the worked example, a fourteen-page monthly report became six pages before any technology was involved, and the biggest single improvement had nothing to do with AI. Twenty-two hours of waiting for two people set the calendar, so moving one request from day six to day one removed a week of elapsed time without saving a minute of work.

👇 For Power User subscribers only. Download here 👇

NOT ADVICE

The information is intended to be helpful, but is in no way a substitute for seeking professional advice for your specific situation or intent. This applies to business, financial, legal, health, medical, personal, or other matters discussed herein. Please read the full DISCLAIMER

For clarity, all content found in The Library is for your use only. Not for resale or redistribution in any form.

For anyone facing a hard choice with a deadline. Finds the options you’re avoiding, works out what you’d need to believe for each to be right, then writes the story of how it fails.

Packs come in three formats: Claude Skills, Gemini Gems and Custom GPTs, and include full worked examples so you can see the output before committing your to it.

Pros and cons lists don’t work. You already know both columns, which is why you’re stuck. Nothing in the exercise weighs the items, nothing checks whether the option list is even complete, and at the end you have a tidier version of the same uncertainty. Which is why almost everybody who makes one then does what they were going to do anyway.

This skill attacks the problem in three places instead. It breaks the framing, because most decisions arrive as a binary that isn’t one and the binary hides four or five other options. It turns each option into claims that could turn out false, marked for whether you could check them before the deadline, which is where decisions actually get won. And it writes the story of the failure twelve months out, in the past tense, for the leading option and the runner-up both.

The most common useful outcome is not a recommendation. It’s stop the process, go and make two phone calls, come back on Thursday. People routinely agonise for a month over something one conversation would settle.

👇 For Power User subscribers only. Download here 👇

NOT ADVICE

The information is intended to be helpful, but is in no way a substitute for seeking professional advice for your specific situation or intent. This applies to business, financial, legal, health, medical, personal, or other matters discussed herein. Please read the full DISCLAIMER

For clarity, all content found in The Library is for your use only. Not for resale or redistribution in any form.

For anyone who has to stand behind a figure in a meeting. Works out what your data can and cannot answer, then hunts for the finding that’s true and misleading.

Packs come in three formats: Claude Skills, Gemini Gems and Custom GPTs, and include full worked examples so you can see the output before committing your to it.

A membership organisation looks at five years of data and finds its renewal rate has dropped from 84 per cent to 76. The calculation is correct. The data is clean. The finding is real.

Three days before it reaches the board as evidence of declining member satisfaction, somebody breaks the number down by category. Every single category is flat or improving. The overall rate fell because a discounted early-career grade had grown from 5 per cent of the membership to nearly 16, and that grade renews at 61 per cent, which is entirely normal for a discounted entry tier.

There was no renewal problem. There was a successful recruitment campaign, and a metric describing the change in mix rather than any change in behaviour.

That’s the failure this skill exists for. It survives review, because the number checks out.

👇 For Power User subscribers only. Download here 👇

NOT ADVICE

The information is intended to be helpful, but is in no way a substitute for seeking professional advice for your specific situation or intent. This applies to business, financial, legal, health, medical, personal, or other matters discussed herein. Please read the full DISCLAIMER

For clarity, all content found in The Library is for your use only. Not for resale or redistribution in any form.

For anyone whose research ends up in front of someone else. Grades your sources, names what they disagree about, and tells you how confident you can honestly be.

Packs come in three formats: Claude Skills, Gemini Gems and Custom GPTs, and include full worked examples so you can see the output before committing your to it.

Three things this skill does that ordinary research prompts don’t

It decides what would count as evidence before searching. If nothing could settle the question, it says so and stops, because a question of judgement or values dressed up as a research question produces a document that looks authoritative and is a collection of opinions.

It searches against itself on purpose. The query shapes the answer, so for every search framed one way it runs one framed the other, plus deliberate hunts for the failure cases. Organisations that tried something and abandoned it are the most useful and least visible sources, because success gets a press release and reversal gets silence.

It refuses to smooth disagreement. Where sources conflict, it names the conflict and works out which kind it is, and the kinds matter enormously. Two sources measuring different things under the same word is a finding. Competent people looking at the same evidence and disagreeing is a different finding.

Both get flattened into a moderate-sounding sentence by the usual approach.

👇 For Power User subscribers only. Download here 👇

NOT ADVICE

The information is intended to be helpful, but is in no way a substitute for seeking professional advice for your specific situation or intent. This applies to business, financial, legal, health, medical, personal, or other matters discussed herein. Please read the full DISCLAIMER

For clarity, all content found in The Library is for your use only. Not for resale or redistribution in any form.

For writers and one-person businesses publishing into silence. Works out how many people you actually need, then which single channel will reach them.

Packs come in three formats: Claude Skills, Gemini Gems and Custom GPTs, and include full worked examples so you can see the output before committing your to it.

What this skill does

Eight steps. The maths, then where those specific people already are, then an audit of what you’ve already published and what happened to it.

Then it scores the channels on three things: how many of the right people are actually there, whether you can produce that format repeatedly without dreading it, and how long until you know whether it’s working.

Then one primary channel. Not three. Spreading thin effort across five is the most reliable way to conclude that distribution is hard, when distribution is only slow.

Then the first ten things to publish or send, drawn from your buyer’s own problems.

👇 For Power User subscribers only. Download here 👇

NOT ADVICE

The information is intended to be helpful, but is in no way a substitute for seeking professional advice for your specific situation or intent. This applies to business, financial, legal, health, medical, personal, or other matters discussed herein. Please read the full DISCLAIMER

For clarity, all content found in The Library is for your use only. Not for resale or redistribution in any form.

Four connected skills that turn what you’ve already done into one offer somebody pays for, a buyer you understand properly, copy that survives a sceptic, and a way of delivering it that doesn’t eat your life.

Packs come in three formats: Claude Skills, Gemini Gems and Custom GPTs, and include full worked examples so you can see the output before committing your to it.

It’s a reliable pattern. Somebody with fifteen years of hard-won expertise decides to go solo, and the thing they want to sell is the thing they’ve been reading about for six months rather than the thing they’ve been doing for fifteen years.

The reason is understandable. The expertise feels boring to them, because they’ve done it a thousand times. It’s routine. Meanwhile the new interest is exciting precisely because they’re not good at it yet.

The market disagrees. What felt routine to you looks like magic to somebody currently drowning in it, and they’ll pay for that.

So this pack starts by separating what you’ve actually done from what you’re interested in, and it’s a bit brutal about the difference.

👇 For all Paid subscribers only. Download here 👇

NOT ADVICE

The information is intended to be helpful, but is in no way a substitute for seeking professional advice for your specific situation or intent. This applies to business, financial, legal, health, medical, personal, or other matters discussed herein. Please read the full DISCLAIMER

For clarity, all content found in The Library is for your use only. Not for resale or redistribution in any form.

Four connected skills that teach AI to write in your voice, find the piece worth writing, tear up its own first draft, and get everything out of what you publish.

Packs come in three formats: Claude Skills, Gemini Gems and Custom GPTs, and include full worked examples so you can see the output before committing your to it.

Ask any model for a “more natural tone” and watch what happens. It gives you the same paragraph with the collar unbuttoned. Still nobody’s voice. Still the sound of a very polite committee.

The reason is boring and fixable. The model has no idea what natural sounds like for you specifically. It has never read you. It is doing an impression of writing in general, which is exactly what generic prose is.

So this pack starts by reading you. Five samples, an analysis of your actual sentence lengths and punctuation habits and the words you reach for, and a specification precise enough that a stranger could write a paragraph and land close. That file then feeds everything else.

👇 For all Paid subscribers only. Download here 👇

NOT ADVICE

The information is intended to be helpful, but is in no way a substitute for seeking professional advice for your specific situation or intent. This applies to business, financial, legal, health, medical, personal, or other matters discussed herein. Please read the full DISCLAIMER

For clarity, all content found in The Library is for your use only. Not for resale or redistribution in any form.

That’s a normal hit rate, and it’s why AI research falls apart the moment someone senior asks where a number came from.

Better models haven’t fixed this. Explicit source rules have. The eight prompts below turn a research request into an audit. You tell the AI what counts as evidence before it starts, then you make it grade its own sources afterwards.

Why this keeps happening

Every major model now ships a deep research mode, and they’re all built to look thorough. Twenty sources reads better than four. Nothing in the retrieval step rewards accuracy, so that job lands on you by default.

These prompts work by setting acceptance criteria first, then forcing a second pass where the model marks its own homework against those criteria.

Prompt #1: Source rules brief

What it does: Sets the evidence standard for a research task before the AI searches, so weak sources never enter the output in the first place.

When to use it: As the opening message of any research thread where the output will be seen by someone other than you.

The Prompt:

Apply these source rules for the whole of this task.

Topic: [TOPIC]

Decision this research supports: [WHAT YOU’LL DO WITH THE ANSWER]

Acceptable sources, in order of preference:

Tier 1: peer-reviewed papers, working papers, clinical trials, official datasets

Tier 2: government statistics, regulatory filings, court documents, central bank or agency publications

Tier 3: reporting from outlets with a named author, a publication date and a corrections policy

Tier 4: company documentation, but only for claims about that company’s own product or finances

Never cite: SEO blog posts, listicles, content marketing pages, press releases written up as findings, AI-generated summaries, aggregators that don’t link to the original, anything with no named author, anything published before [DATE LIMIT] unless it is itself the original study.

For every claim you make, give me the claim, its source tier (1 to 4), the publication date and a direct link to the specific page where the claim appears. Not the homepage.

If you can’t find a tier 1 to 3 source for something, write UNSUPPORTED next to the claim and move on. Do not fill the gap with a weaker source.

Research question: [SPECIFIC QUESTION]

How to use it:

1. Fill in the topic, the decision it supports and the date limit before you paste anything.

2. Send this on its own and wait for the model to confirm the rules back to you.

3. Then send your research question, or paste it in at the bottom and send once.

Example input: Topic: remote work and productivity. Decision this supports: a board paper recommending we keep two office days. Date limit: January 2023. Research question: what does the evidence say about output changes when knowledge workers move from full office to hybrid?

What you’ll get: A shorter answer than you’re used to, with tier labels, dates and deep links attached to each claim. Expect two or three UNSUPPORTED tags, which are the most useful lines in the response.

Advanced note: Set the date limit tighter than feels comfortable. If the answer collapses at an 18 month cutoff, the topic moves faster than your research method does.

The rest of the audit is for paying subscribers

Prompt 1 will cut your unusable citations by more than half, mostly because it stops the model papering over gaps.

The other seven prompts are the audit itself. You run those on research you already have, including work the AI did back when nobody was checking.

Prompt 2, Citation grader: makes the model score every source it just handed you and delete its own failures.

Prompt 3, Primary source trace: chases a claim back through the chain of blogs quoting blogs until it reaches the original study or dead-ends.

Prompt 4, Disagreement hunt: finds credible sources that contradict your conclusion, before someone in the meeting does.

Prompt 5, Stale data check: flags figures that have been revised, superseded or quietly withdrawn since publication.

Prompt 6, Who benefits: identifies funding, ownership and commercial interest sitting behind each source.

Prompt 7, Claim-to-source map: lists every claim with nothing attached to it, which is exactly where confident nonsense hides.

Prompt 8, Confidence-rated memo: rewrites the whole piece with high, medium and low confidence labels so you know which lines you can say out loud.

Plus a reusable research brief template you fill in once per project and paste at the top of every thread.

If you only do one thing with AI this week, Prompt 1 is a decent choice. The audit is what makes the output safe to send to someone who checks.

Read more

The problem with a new idea is that it arrives fully dressed. You see the finished newsletter, the packed launch, the nice round subscriber number. You do not see the six months of grinding out content nobody reads. The gap between those two pictures is where I’ve lost the most.

So I built a gate. Eight prompts I run before I let myself fall in love. I run them in Gemini, not ChatGPT, and there’s a specific reason for that, which I’ll get to. The short version: Gemini can actually go and check whether the world agrees with me, and I’ve stopped trusting any idea validation that can’t.

Here’s the whole thing.

Why Gemini for this

Most idea-validation prompts are for show. You describe your idea, the model tells you it’s promising, you feel good, you build it, and it dies. The model was never checking anything. It was pattern-matching your enthusiasm back at you.

Gemini’s free tier does two things that break that loop. Its answers are grounded in Google Search, so when I ask whether a market exists, it goes and looks rather than guessing from training data. And Deep Research will spend a few minutes crawling dozens of pages and hand me back a proper report. Free accounts get a handful of those a month, which is plenty when you’re only running it on ideas that survive the first few prompts.

There’s also Gems, which are custom saved assistants, free for everyone now. I’ve turned this whole sequence into one. More on that at the end.

The point of grounding is simple. I don’t want a model that agrees with me. I want one that can be sent to find out I’m wrong.

The eight prompts

Run them in order. Each one is allowed to kill the idea. If it does, you stop, and you’ve saved yourself a season of work.

Read more

I got tired of guessing what to write next. So I handed the problem to ChatGPT, and not in the vague “give me 20 blog ideas” way that produces 20 ideas you’d never publish. I made it audit what I’d already done, find the holes, and fill them with a proper calendar. Three months came out the other side. Most of it was better than what I’d have picked myself.

Here’s the exact sequence.

The idea behind it

The trick isn’t asking for ideas. Anyone can get ideas. The trick is making the model understand your existing mix well enough to spot what’s missing, which is a completely different job.

A gap is only a gap relative to a pattern. So the first half of this is teaching ChatGPT your pattern. The second half is asking it to break the pattern usefully. Skip the first half and you get generic filler. Do it properly and you get ideas that fit your voice but point somewhere you haven’t been.

One note on tiers before we start. The six core prompts run on the free version of ChatGPT. Two optional upgrades near the end use Deep Research and custom GPTs, both of which sit behind the paid plan now. I’ll flag those clearly so you know what’s free and what isn’t.

The six prompts

Run them in one session so the model keeps the context. Each one builds on the last.

1. Feed it the raw material

Don’t describe your newsletter. Show it. Paste in your last 20 to 40 post titles, or the titles plus a one-line summary of each if you have them handy.

Here are my recent newsletter posts, titles and short summaries. Read them as a set, not one by one. Don’t comment yet. Just confirm you’ve got them and tell me you’re ready for the next step.

Making it hold the whole set at once matters. You want it reasoning across the corpus, not reacting to the last thing it read.

2. Make it map what you actually cover

Now group these posts into themes. Give me the real categories my content falls into, not the ones I’d claim. Tell me roughly what share of my output sits in each theme. Be honest if one theme is eating everything.

This step alone is worth the exercise. Seeing your own mix as percentages is uncomfortable in a useful way. I found nearly half my posts clustered in one corner I thought was a sideline.

Read more

Downloadable Claude Skill and Custom GPT

On 21 July, Substack switched on an AI detector. It’s built with a company called Pangram, and it scans posts, notes and comments over 100 words. Anyone can run it. Readers on your work, you on your own drafts, or a stranger on someone they’ve taken against.

It’s opt-in and there’s no penalty attached. Nobody gets ranked down or banned for a high score. Chris Best has been careful to frame it as a first step toward transparency rather than a rule about which tools you’re allowed to use.

I’m relaxed about it, and I should explain why before I get to the interesting part.

I’ve never hidden that I use AI. I wrote an Honest AI Manifesto in April saying exactly how and where. Some pieces here are drafted almost entirely by hand and tidied with a model. Some start as a conversation and get rewritten by me four times. A few are mostly mine with a paragraph I couldn’t crack handed over in frustration. That’s the honest answer and it’s been on the record for months.

So a scanner isn’t a threat to me. It might tell readers something they already know.

I do have one worry, and it isn’t about my score. The false positives land in a pattern. They hit people writing in a second language, and they hit neurodiverse writers whose rhythm and repetition a model reads as synthetic. One writer ran the same piece twice last week and got 100% AI, then 100% human. That’s a coin flipper, not a detector.

But that’s an argument for treating the number with suspicion, not for hiding from it.

Slop is a readability problem before it’s a moral one

Using AI is fine. Publishing what it hands you first time is where it goes wrong.

Generated prose has a texture. Every sentence lands between 15 and 20 words. Every section closes with a tidy summary. The same eighty verbs turn up over and over. Delve. Leverage. Foster. Navigate. Nothing is technically wrong, and that’s the trouble, because there’s nothing to catch on.

Human writing is lumpy. Two-word sentences next to forty-word ones. A paragraph that goes on too long because the writer got interested. A joke that doesn’t quite land. Repetition where they should have varied it. Those are the warts, and readers need them. They’re the handholds.

Strip them out and you get text that’s correct and frictionless, which sounds like a compliment until you notice nobody finished reading it.

So the case for humanising your drafts has nothing to do with detectors. Slop is boring, and boring loses subscribers faster than any label ever will.

What I built

A tool that scans a draft, tells you what’s wrong with it, then edits. You paste your writing in. That’s the whole interface.

The scan counts rather than guesses. Ask any AI to tell you if your writing sounds robotic and it’ll have an opinion, and the opinion will be different tomorrow. This one comes back with numbers, and they’re the same numbers every time.

It catches what your eye slides past. Em dashes. Semicolons. Roughly 300 words and phrases that show up constantly in generated text and hardly ever in good writing. It gets the structural habits too, like the self-posed question (”The result? Devastating.”) and the trailing clause that adds nothing (”...solidifying its position as a market leader”).

Then it measures rhythm. It works out how much your sentence lengths vary, and your paragraph lengths, and tells you when the answer is “not enough”.

You get a number at the end. Under 10 per thousand words reads human. Above 30 and the piece needs rebuilding rather than tidying.

I ran a deliberately awful sample through it and got 48. I ran a passage I was proud of and got zero.

What it won’t do

It won’t help you beat Pangram, and I’ve built it to refuse if you ask.

The refusal is practical. The job can’t be done reliably. Detectors score statistical patterns in how text was generated, which isn’t the same thing as the vocabulary and sentence shapes this edits. The two overlap a bit. Your score might drop after an edit. It might go up. Anyone selling you a guaranteed number is selling you a guess in a confident voice.

And chasing the score is the wrong game anyway. If you’re editing to satisfy a classifier, you’ll produce writing that satisfies a classifier. Nobody has ever subscribed to a newsletter because it scored well on anything.

Edit so it reads better. The score does what it does.

The part many ‘humanising’ tools get wrong

Every de-slop tool I’ve tried is too keen.

You write about actual landscapes and it strips out “landscape”. You use “systemic” correctly in a piece about systemic risk and it swaps in something vaguer. You make a joke and it sands the joke off, because jokes are irregular and irregular looks like a mistake.

What comes back is clean, correct and dead. Which is the same problem you started with, wearing a different coat.

So this one shows its work. Every edit comes with a change log, and underneath that, a list of everything it flagged and deliberately kept, with reasons. If it left your weird metaphor alone, it says so and says why. You can disagree with it, which is more than most editing tools allow.

Where to get it

It’s a Claude Skill, so it loads by itself when you mention humanising a draft. There’s a ChatGPT version built as a custom GPT. It runs the same scanner under Code Interpreter. The Gemini version works but loses the diagnostic, because Gems can’t run code. I’d rather say that now than have you find out.

The term list is a plain text file. Add your own words, delete mine, and the scanner picks up the change without anyone touching code. My list bans “delve” and “tapestry” and about 298 other things. Yours will be different. Mine bans “utilise” purely out of spite.

Download the files for Claude and ChatGPT, including instructions, here (zip file - unzip when downloaded) …

Read more

It’s inbox summarisation. Someone gets AI access, opens their email, and thinks: this is the obvious place to start. Forty unread emails become a tidy bullet list in ten seconds. It feels like a win.

It isn’t. Summarising your inbox doesn’t remove any of the work. You still have to read the original emails to reply properly, you still have to decide what matters, and now you’ve added a summary you have to check against the source. The actual bottleneck in most people’s day isn’t reading email. It’s deciding what to do about it and writing the reply. That’s the part most workflows skip, because it’s harder and less satisfying to set up.

This article gives you 8 prompts that target the real bottleneck instead. You’ll go from “summarise this” to “draft my actual response, flag what needs a decision, and tell me what I can ignore.”

Why this matters now

Most AI workflow advice still treats summarisation as the entry point, because it’s the easiest demo. It looks impressive in five seconds and it’s the first thing every tool tutorial shows you. But a workflow only saves time if it removes a step. Summarising adds one. The prompts below skip straight to decisions and drafts, the parts of your day that actually eat the hours.

Prompt 1: The reply drafter

What it does: Writes a complete first-draft reply to an email, ready for you to edit and send.

When to use it: For any email that needs more than a one-line response. Skip it for anything you can answer in five words.

The Prompt:

Here’s an email I need to reply to. Write a complete draft reply in my voice: direct, no corporate filler, short paragraphs. Address every question or request in the original email. If something needs more information from me before you can answer it properly, flag that clearly at the top instead of guessing. Email: [PASTE EMAIL]

How to use it:

1. Open the email you’d normally summarise and move past

2. Paste the full text into the prompt

3. Edit the draft for tone and send, rather than starting from a blank reply box

Example input: Here’s an email I need to reply to. [pasted client email asking for a project update and a new deadline]

What you’ll get: A complete reply addressing both the update and the deadline question, with a flag if you haven’t given it enough detail to answer the deadline part confidently.

Advanced note: Save 3 or 4 of your own past replies somewhere and paste one in alongside the prompt occasionally. It keeps the draft sounding like you instead of drifting into generic email voice over time.

That one prompt alone removes the blank-page problem on every reply you write this week.

But replying faster is only half the fix. The other half is knowing which emails need a decision from you and which ones don’t need anything at all. The next 7 prompts handle that.

Here’s what’s behind the paywall:

Prompt 2, the decision flagger: Reads a batch of emails and tells you which ones actually need a choice from you, not just a reply.

Prompt 3, the no-reply filter: Identifies emails you can archive or ignore without any response, and tells you why.

Prompt 4, the meeting request handler: Drafts accept, decline, or reschedule replies to meeting requests based on your stated priorities.

Prompt 5, the follow-up tracker: Scans a thread and tells you what you’re still waiting on from someone else, so nothing falls through.

Prompt 6, the tone matcher: Adjusts a drafted reply to match the formality level of the person you’re writing to.

Prompt 7, the batch responder: Drafts short replies to several low-stakes emails at once, so you clear the small stuff in one pass.

Prompt 8, the weekly inbox review: Looks back at a week of email and tells you which relationships or threads need attention you’ve been missing.

Plus: a one-page workflow audit framework, a short set of questions to run against any AI workflow before you build it, so you catch this mistake yourself next time instead of reading about it after the fact.

If you only do one thing this week, run Prompt 1 on your next three emails. If you want the system that actually clears your inbox instead of just describing it to you, the rest is one upgrade away.

Upgrade to get all 8 prompts, the workflow audit framework, and the full step-by-step sequence

Read more

There’s a difference. Comments are generally where people agree. Complaints are where they tell you what’s actually broken in their work or their process. I think the majority of people scrolling LinkedIn skim the comments and skip the complaints, because complaints are usually buried three replies deep in a thread about something else entirely.

And that’s the gap. Everyone in your niche is publicly narrating their problems on LinkedIn right now, in posts, in comments, in the replies to other people’s posts. No amount of guessing or formal market research delivers this intel as fast or as cheaply.

This article gives you 8 prompts that turn LinkedIn into a research tool. You’ll pull real language from real posts, sort it into patterns, and end up with a shortlist of problems worth solving, and crucially, in words your audience already uses.

Why this works now

Like all social platforms, LinkedIn’s algorithm rewards posts that get engagement. This means the platform is full of people venting, asking, and comparing notes in public. That’s free research sitting in plain sight. AI makes it usable because it can read fifty posts and group the patterns in minutes, something that would take you an entire afternoon by hand. You don’t need a survey. You need fifteen minutes and the right prompt.

Prompt 1: The struggle scrape

What it does: Pulls the recurring problems mentioned across a batch of LinkedIn posts or comments from your niche.

When to use it: Right at the start, before you do anything else. This is your raw material.

The Prompt:

I’m going to paste a batch of LinkedIn posts and comments from people in [YOUR NICHE]. Read through all of them and pull out every specific problem, frustration, or struggle mentioned, even small ones. For each one, quote the exact phrase they used and note who said it (just their role or title, not their name). Don’t summarise or interpret yet. Just extract. Posts: [PASTE POSTS]

How to use it:

1. Go to LinkedIn and search a keyword or hashtag tied to your niche

2. Copy 15 to 20 posts and their top comments into a doc

3. Paste the lot into the prompt and run it

Example input: I’m going to paste a batch of LinkedIn posts and comments from people in freelance copywriting. [20 posts pasted]

What you’ll get: A list of 30 to 50 quoted frustrations, each tagged with the poster’s role. Some will repeat. That’s the point.

Advanced note: Don’t clean the posts before pasting them. Typos, ALL CAPS, and rambling are signal. AI will smooth them out if you ask it to summarise too early, and you’ll lose the texture that makes a problem feel real.

That prompt alone will hand you a raw list of problems straight from the people who have them, no guessing required.

But one list of quotes is just that, a list, not a strategy. The next 7 prompts turn that list into something you can actually act on.

Here’s what’s behind the paywall:

Prompt 2, the pattern sorter: Groups your raw quotes into 5 to 8 actual themes instead of 40 scattered complaints.

Prompt 3, the frequency ranker: Tells you which problems show up again and again versus which ones are one-off noise.

Prompt 4, the language miner: Extracts the exact words and phrases your audience uses, so your content stops sounding like a textbook.

Prompt 5, the hidden problem finder: Surfaces the struggles people imply but never say outright, the ones competitors miss.

Prompt 6, the content angle generator: Turns each pattern into 3 ready-to-use post or email angles.

Prompt 7, the audience segment splitter: Shows you which struggles belong to which sub-group inside your niche, so you stop writing to an imaginary “everyone.”

Prompt 8, the validation check: Cross-checks a problem you think you’ve found against a fresh batch of posts, so you’re not building on a fluke.

Plus: a one-page pain language tracker, a simple log for the exact phrases people use, so the patterns build across weeks instead of resetting every time you do this.

If you only do one thing this week, run Prompt 1 on twenty posts and read what comes back. If you want the system that turns that into a repeatable research habit, the rest is one upgrade away.

Upgrade to get all 8 prompts, the pain language tracker, and the full step-by-step sequence.

Read more

Which best describes your working week?

a) Publishing. Newsletter, blog, videos, building an audience.
b) Marketing, for clients or an employer.
c) Running a small business. Publishing is one job among several.
d) A job inside a larger organisation.

Just hit reply and send a, b, c or d. Nothing else needed.

Cheers
Andy

They’re getting better at writing prompts. Tighter instructions, cleverer personas, more specific examples. And the outputs do get better, for a while, on the tasks they thought to test. The problem is that “it looked good when I tried it” is not a quality standard. It’s a hope.

Tools like DSPy are already optimising prompt text automatically. The writing part is being automated. What can’t be automated, at least not yet, is knowing whether the output is actually good. That requires a human to define what good looks like, write down the test cases, and run them consistently.

That’s the skill shift. From writing prompts to testing them.

These 9 prompts take you through the full process: building test cases, scoring outputs, finding failure modes, running comparison tests and locking in the version that actually works.

Why this matters right now

Automated prompt optimisation is moving fast. DSPy, PromptFoo, and similar tools can iterate through hundreds of prompt variants in the time it takes you to write one manually. The bottleneck they all share is the same: someone has to tell the system what a good output looks like.

If you can’t answer that question in writing, with specific criteria, the tools are useless. The people who’ll get the most from the next wave of AI tools aren’t the best prompt writers. They’re the best prompt evaluators.

Prompt 1: Build your first test case

What it does: Produces a structured test case document for any prompt you’re currently using, including input variations, success criteria and failure flags.

When to use it: Before you rely on any prompt for real work, or when an existing prompt is producing inconsistent results.

The Prompt:

I want to build a proper test case for an AI prompt I’m using. Here is the prompt I want to test:

[PASTE YOUR PROMPT HERE]

My use case is: [DESCRIBE WHAT YOU USE THIS PROMPT FOR, e.g. “writing LinkedIn posts for a B2B software company”]

My target output quality is: [DESCRIBE WHAT GOOD LOOKS LIKE, e.g. “under 200 words, no jargon, ends with a question, sounds like a person not a brand”]

Please produce a test case document with: 1. 5 input variations that cover the range of real scenarios I’d use this prompt for (from easy/typical to difficult/edge case) 2. For each input, a written description of what a passing output would look like 3. A list of 5 specific failure flags (things an output could do that would make it a fail regardless of other quality) 4. A scoring rubric with 4 criteria, each scored 1-5, that I can apply to any output from this prompt

Format the result so I can paste it into a document and reuse it.

How to use it:

1. Paste your existing prompt into the first placeholder exactly as you currently use it

2. Write the use case description in plain language, not aspirationally, describe what you actually use it for

3. For target output quality, be specific: length, tone, structure, what it must and must not contain

Example input: Prompt = a prompt that writes cold email subject lines. Use case = “outreach emails for a freelance UX designer targeting e-commerce startups.” Target quality = “under 8 words, no question marks, specific to the recipient’s industry, doesn’t use the word ‘quick.’”

What you’ll get: A reusable test document with 5 scenario inputs, pass/fail descriptions for each, a list of automatic failure conditions and a 4-criteria scoring rubric. Takes about 3 minutes to fill in and saves hours of guesswork later.

Advanced note: Run the same test case document through two different models (Claude and GPT-4o, for instance) using identical inputs. The differences in where each model fails tell you more about prompt weaknesses than any amount of manual tweaking will.

ooops, a paywall…

That test case gives you something most people never have: a written definition of what good actually means for your specific prompt.

Behind the upgrade, there are 8 more prompts that take this further:

Prompt 2: Score an existing output: Apply your scoring rubric to a real output and get a structured evaluation with a numerical score and specific improvement notes, not just “this could be better.”

Prompt 3: Find failure modes: Run your prompt against 6 adversarial inputs designed to break it, so you know where it falls apart before your actual work does.

Prompt 4: Write a comparison test: Set up an A/B test between two prompt variants using identical inputs and a blind scoring method, so you can make version decisions on evidence rather than preference.

Prompt 5: Extract implicit criteria: Feed Claude a set of outputs you rated highly and ask it to reverse-engineer your unstated preferences into explicit criteria you can add to your rubric.

Prompt 6: Build a regression log: Create a living document that records prompt versions, test scores and change notes, so you can roll back to a previous version if a tweak makes things worse.

Prompt 7: Stress-test for edge cases: Generate 10 edge-case inputs specific to your use case, including unusual formats, missing information and off-topic requests, to find where your prompt gets confused.

Prompt 8: Write a minimum viable spec: Turn your test case document into a one-page prompt specification that another person (or an automated tool like DSPy) could use to optimise or rebuild your prompt from scratch.

Prompt 9: Audit a prompt you didn’t write: Apply the full testing process to a prompt you found online or inherited from someone else, so you can decide whether it’s actually fit for purpose before using it in production.

Plus: The Eval Scorecard, a reusable plain-text template you can copy into any document and fill in for any prompt in under 10 minutes.

If you only do one thing differently this week, building that first test case is a solid start. The full testing workflow is behind the upgrade.

Read more

I knew there had to be a better way. Turns out there was, and it had been sitting in Claude’s settings the whole time.

Skills files. A persistent instruction set that loads automatically whenever you open a chat. You write it once, Claude knows your defaults forever. No preamble, no re-explaining, no “as I mentioned earlier.”

I spent a couple of hours building one for copywriting work. The prompts below are exactly what I used. The result is a version of Claude that already knows your clients, your process and your rates before you type a single word.

Why Skills files aren’t just for developers

Most of what’s written about Skills online is aimed at coders. Which makes sense. They got there first.

But a developer’s work is fairly self-contained per session. Copywriting isn’t. It runs on accumulated context: who the client is, how they sound, what a good brief looks like, what you’re going to push back on. That context disappears every time you close the tab.

Skills is where you put it so it doesn’t.

Prompt 1: Brand voice entry builder

What it does: Produces a fully formatted Skills entry for a specific client’s brand voice, ready to paste straight into your Skills file.

When to use it: When onboarding a new client, or when you’ve been working with someone long enough that you can describe how they sound with some precision.

The Prompt:

You are helping me build a Claude Skills file entry for a client’s brand voice. I’ll give you information about this client and you’ll turn it into a clean, structured Skills entry I can paste directly into my file.

Client name: [CLIENT NAME] Industry: [INDUSTRY] Target audience: [WHO THEY’RE WRITING FOR] Tone description: [HOW THEY SOUND - e.g. “direct but warm, plain English, no jargon”] Words or phrases they use: [SPECIFIC VOCABULARY OR PHRASES] Words or phrases to avoid: [THINGS THAT DON’T FIT THEIR VOICE]Output types I write for them: [E.G. “email sequences, landing pages, LinkedIn posts”] One example of their writing I like: [PASTE A SHORT SAMPLE]

Format the output as a Skills entry with these sections: Voice overview (2–3 sentences), Tone markers (bullet list), Vocabulary rules (bullet list), Output defaults (what to assume unless told otherwise). Keep it under 300 words.

How to use it:

1. Fill in each placeholder from memory, not from a mood board. If you can’t describe how the client sounds without looking something up, you don’t know their voice well enough yet.

2. Read the output like a client would. Does it sound like them? Edit anything that’s off before you file it.

3. Paste it into your Skills file under a heading like “Client: [Name] - Brand voice.” Keep that heading consistent so Claude can find it.

Example input: Client name: Hartley & Co / Industry: B2B SaaS / Target audience: ops managers at mid-size logistics firms / Tone: plain-spoken, slightly dry, no marketing fluff / Words they use: “straightforward,” “works out of the box,” “no surprises” / Words to avoid: “innovative,” “cutting-edge,” anything with an exclamation mark / Output types: case studies, email sequences, web copy / Example: “Hartley tracks your fleet in real time. No setup headaches, no ongoing IT support. It just works.”

What you’ll get: A 200-250 word Skills entry with four clearly labelled sections, formatted for direct paste into your Skills file.

Advanced note: If you write for the same client across different channels (e.g. long-form articles versus social posts), run this prompt twice with different output types specified. The voice entry will pick up the register shifts between channels and give you a more accurate default.

oops, you hit the paywall!

Here’s what’s behind it…

The bonus is a full Skills file template with all seven sections pre-labelled and placeholder text in every slot. You can open it, work through the prompts in order, and have a complete file by the end of the session.

Plus the other seven prompts in the sequence:

Prompt 2 - Brief intake entry builder: Produces a Skills entry that tells Claude exactly what information you need from a client before you start writing, so it can flag gaps automatically when you paste a brief in.

Prompt 3 - Platform tone calibrator: Builds a Skills entry for output defaults per platform: what changes between a LinkedIn post, an email subject line and a homepage hero when you’re writing for the same client.

Prompt 4 - Feedback translator: Creates a Skills entry that helps Claude turn vague client feedback (”make it punchier,” “less salesy”) into specific edits rather than a full rewrite.

Prompt 5 - Scope and rate defender: Builds an entry that gives Claude enough context about your pricing to help you write scope-of-work language and push back on feature creep.

Prompt 6 - Portfolio case study writer: Produces a Skills entry with your preferred case study format, so Claude can draft a first version from a few bullet points about a completed project.

Prompt 7 - Client onboarding doc generator: Builds an entry covering your working process, turnaround times and revision policy, so Claude can produce a ready-to-send onboarding doc for any new client.

Prompt 8 - Generic output checker: Creates a Skills entry with your personal quality bar, so you can run any draft through Claude and get a specific list of what needs fixing and why.

Read more