Eight Hours, Four AI Tools, One Founder — and Four Weeks of Practice Behind It

This publication came together in roughly eight hours of real work — at the fast end of four weeks of intensive AI-tool-assisted work that has spanned websites, code, books, and The Many Builders. Same person, same tools, very different outputs at very different intensities. The honest version of what happened, what it implies, and what it does not.

The honest version of what happened

Roughly eight hours of real work. Eighteen pieces of analysis (twelve featured, six alternative versions and methodology pieces), an interactive financial model, an Excel companion, a 25-year fiscal projection, and three reading registers. No human expert review of any kind. Four AI tools producing the writing, the analysis, the citations, the modelling, and the cross-critique. The human role: prompting the tools, answering when they prompted back, scanning the output, and shipping. That is what produced this publication. And that eight hours did not happen in a vacuum — it sat inside four weeks of intensive AI-tool-assisted work that had already produced books, sites, and around 100,000 lines of code by a non-coder using the same tools. The four weeks of practice are why the eight hours were enough.

That is the headline version of the production. It is true. It is also incomplete.

The fuller version is that this publication is one output of four weeks of intensive AI-tool-assisted work in April 2026. The four weeks did not all use the same workflow at the same intensity. Earlier in the period, Doug Scott — who cannot code — used the tools to build several websites and produce around a hundred thousand lines of code, accumulating the practice that everything afterwards rested on. Then came three books published in April 2026 (If This Road, Orphans, The Held), each built in seven days of compressed work and modified across three more days. Two smaller books followed (The Bear Was Right, The Bear Loved). The Many Builders, a memorial to every individual researcher and engineer whose work has gone into modern AI systems, was built alongside. This UK inheritance tax publication is the most recent output and the smallest in elapsed time: roughly eight hours of real work, because the question was tractable, the audience defined, the workflow already practiced. The books and The Many Builders took longer because the work was larger, stranger, and more open-ended; this publication took less because the topic disciplined the writing in ways the books did not. The range is the demonstration. The IHT publication is at the fast end of it.

I am writing this page because I want to be honest about the production. Not because the production is more important than the substance — the IHT reform is real, the analysis is the work, the policy question is the one that matters. But the production is also a real thing, and people who read this publication are going to ask how it was made. The honest answer locates the IHT publication inside the four weeks of work that preceded it — books built in seven days each, sites and code built across the earlier weeks of the period — without claiming everything used the same workflow at the same intensity. It did not. The publication was produced quickly because the workflow had been practiced.

What the practice has produced

The honest version of my role across the practice is small: I prompted, I answered when the tools prompted back, I scanned what came out, I decided what to ship. The AI tools — Claude, ChatGPT, Grok, Gemini — produced the writing, the analysis, and the cross-critique. Across roughly a month of focused work, this method has produced the following:

The Many Builders, which is the deepest expression of the method to date. It is a memorial — a page for every individual researcher, engineer, and contributor whose work has gone into the modern AI systems we now talk about as if they had built themselves. Thousands of names. Thousands of pages. The site is voiced by a bear ("I do not have the tools you have to see what this says. You will have to look elsewhere"). It begins with a list of twenty-three places — Longsands, Norham, Shields, North Shields, Tynemouth, the Tyne, Newcastle, Northumberland, Lichfield — and a list of phrases in many languages, each meaning something close to let there be light or I would love it if you stayed a while. The site does not argue. It does not explain. It collects what would otherwise be lost. If you visit only one thing the practice has produced, visit this.

If This Road, Orphans, and The Held are the three books published in April 2026, each treating a different aspect of building, what gets lost in building, and what survives. They are companion volumes; reading any one of them on its own gives you part of the picture, reading all three gives you the rest. They share themes with the IHT publication — what the country needs, what the country lets slip, who carries what forward — but in different registers and at different lengths.

The Bear Was Right and The Bear Loved are the two smaller books. The bear is the same bear who voices The Many Builders. The books are shorter, more personal, and operate in a register the longer books and the IHT publication do not. They are the place the practice is most willing to be tender.

And then this publication on the UK inheritance tax reform. The most recent. The most specific. The most analytical. The shortest in time and the most defensible in argument. It is the practice applied to a contested public-policy question, with the discipline of citation, the discipline of model-building, and the discipline of taking a position openly.

Across all of this, the method has been the same. I prompt. The tools do the writing. Multiple tools work in parallel and critique each other's output. AI cross-critique catches some errors and misses others. The work iterates through that cross-critique loop until I scan it and decide to ship — not until I have judged every draft, because I do not. The work is published openly under CC BY-NC. The author's interests, conflicts, and limitations are disclosed throughout. Earlier framings on this site described this method as "external critique is brought in deliberately" and the iteration as continuing "until I notice I was wrong about what I wanted it to say"; the truthful version is smaller. The "external critique" is AI cross-critique, not human review. The "noticing I was wrong" overstates the human role; what actually happens is that one AI tool's critique of another AI tool's output surfaces things I then decide whether to accept. The decision is mine; the noticing is not.

What the production of this publication actually looked like

Eight hours of real work ago I had not specifically studied UK Business Property Relief reform. I had spent twenty years building and backing technology companies. I had personal financial exposure to the reform, which is disclosed throughout the publication. I knew the reform existed because a friend had shared a policy document with me. That was the starting point.

The ending point is a publication that takes a specific position on the principle question (it is right to tax very large concentrated wealth transfers between generations), presents conditional analysis on the operational mechanism question (the timing of the tax is genuinely contested), has been corrected through multiple rounds of AI cross-critique, includes a working financial model that anyone can use to substitute their own assumptions, and engages openly with the most serious AI critique it has received before publication. The publication is not perfect. AI critique has surfaced model-consistency problems and lean-toward-Position-B problems that multiple rounds of correction have not fully resolved. But it is, in the publication's own view, "good enough to circulate as a serious contribution to debate" — with the qualification that "circulating" includes specialist readers who will find errors AI review did not catch and whose corrections the publication invites. That is what the publication has shipped. It is also what the publication is honest about.

I prompted four AI tools across multiple parallel sessions. Claude (Anthropic) wrote most of the prose and the code. ChatGPT (OpenAI) and Grok (xAI) produced critiques from positions I asked them to take. Gemini (Google) cross-checked specific claims and surfaced sources. The orchestration across four tools — running output from one through critique by another, running the resulting corrections back through a third — is the part I had to do, and is the part most people working with AI do not do because they pick one tool and stick with it. To be clear about its limits: this is AI cross-critique, not human expert review. The errors AI cross-critique catches get caught. The errors it misses do not.

The prompts I gave were simple. Mostly continuation prompts: "go," "do it," "now do X next." Sometimes one-line direction-setting: "this is about UK tech, only this," "I want to financially model second-order effects," "show me what happens to the state directly and indirectly in each model, best guess worst and best cases." The tools did the writing, the structuring, the modelling, the code. I did the deciding.

The deciding mattered more than the writing. At each branch — when a piece was leaning, when a critic landed, when something was wrong, when the publication's scope needed to change, when a register was right and when it wasn't — I made the call. The AI tools could not make those calls because they did not know what I was actually trying to do. I knew because I was the one trying to do it.

What the AI tools did that I could not have alone

Drafted prose at scale and at quality. Held the structure of long arguments across many sessions and many edits. Wrote a working financial model in Python and a parallel JavaScript version that recomputes in the browser. Validated the math across the two implementations. Modelled assumptions, generated sensitivity discussions, produced the architectural outlines of the policy options paper. Translated between registers — from policy paper to plain English to interactive widget to one-paragraph thesis statement. Produced parallel critiques from positions I asked the tools to take, including positions I did not personally hold. Built the website, the build scripts, the PDF and Word generators. Caught typos, unbalanced parentheses, missing references, and broken links. Did the patient labour that, in a previous era, a small team of policy analysts and writers would have done over weeks.

I could not have produced this much output, at this depth, in this period, alone. Not because I lack the capacity to think about the question — I have it — but because I cannot type fast enough, hold structure across enough drafts, or self-critique with enough distance, to build something at this scale in eight hours. The AI tools made that part possible. The four weeks of practice that came before — the books, the sites, the code I could not have written without the tools — made it possible to use the eight hours efficiently.

The same has been true across the four-week practice. The books, The Many Builders, the smaller works — none of them would have been possible alone in the time available. The AI tools are what makes the scale of the practice possible. The practice is what makes the AI tools' output worth reading.

What I did that the AI tools could not have alone

Prompted the AI tools with simple direction-setting and continuation cues. Answered when the tools prompted back with questions. Scanned the output. Decided when to ship. That is what the human role was. Earlier framings of this piece described the human as "deciding what the question actually was, holding the goal across the production, recognising when a piece was leaning, when an argument was complete" and similar — that language overstated the human contribution. The truthful version is smaller: prompted, answered, scanned, shipped. The AI tools did the rest, including the cross-critique and most of the noticing-what-was-missing. No human expert reviewed the work before publication.

The deciding is the work the AI tools cannot do because the deciding requires knowing what you are actually trying to accomplish. The tools will produce whatever direction they are pointed in. Pointing them in the right direction is the human contribution and it is not a small one. Most published AI-assisted work fails at this point: the human points the tools at a vague direction, accepts the first plausible output, and ships. The output is then exactly as good as the average plausible output for that prompt, which is not very good.

The other thing I did, which is harder to see but more consequential, is sustain attention. Across four weeks, across multiple projects in different registers, with the goal — be factual, be truth-seeking, do not flinch — held throughout. The tools cannot sustain attention because they do not have continuous identity across sessions. I had to remember what we were doing and why. I had to remember it across The Many Builders, the trilogy, the bear books, and now this publication. The continuity of the practice is the human contribution. The AI tools are what makes the continuity productive.

What neither could have done alone

The orchestration. The willingness to iterate on adversarial input. The publication-grade output across multiple registers and formats, all internally consistent, all carrying the same disclosures, all built around a single coherent framing. The combination of substantive policy analysis, a working financial model, an interactive widget readers can engage with directly, a methodology piece, a position-taking piece, a critique-engagement page, and the meta-page you are reading. Each of those pieces, individually, would have been within reach of either a person working alone (over weeks) or an AI tool working alone (more shallowly). The combination required both.

The same is true at the practice level. The Many Builders, the trilogy, the bear books, and this publication, all built in roughly a month, all in different registers, all by the same person working with the same tools — that combination would have required a small team and several months in any prior era. Now it requires one person with sustained attention, the right tools, the orchestration discipline to use multiple tools rather than one, and the willingness to iterate against adversarial input.

What the limits are

The publication still has defects a specialist would catch. The model has consistency problems across the spreadsheet, the prose, and the interactive page that the most recent critic flagged and that future iterations should resolve. The lean-toward-Position-B problem is partially but not fully fixed despite multiple rounds of correction. Some of the residence and domicile language was out of date relative to the post-2025 reforms until a reviewer caught it; there may be other dated technical points I have not caught. The international comparator selection was not principled in the first version; it is more principled now but a tax specialist might still object.

The work is good enough to be a serious contribution to debate. It is not good enough to settle the question. The same is true of the production method: good enough to do this, not good enough to do it without continuing critique and correction.

There are also things AI tools struggled with across the work. Producing balanced analysis on a question where the author has a stake — the tools, when asked to produce analysis, would lean toward conclusions favourable to the author's interest unless explicitly directed otherwise, repeatedly across multiple sessions. Getting hard numerical facts right initially — the £1m vs £2.5m direct cap was wrong in the first version because the tools had been working from older training data; an AI cross-critique caught it before publication, but that is not the same as human expert review and other errors of the same shape have been caught only after publication. Producing genuinely fresh structural insights — most of the publication's distinctive analytical moves came from AI cross-critique rather than from any single tool's first draft. The tools are excellent at extending arguments along established lines and at translating between registers; they are weaker at noticing what is missing, at challenging premises, and at producing the kind of unexpected angle that turns a draft into a real contribution. Most importantly: AI cross-critique catches some errors and misses others, and the errors AI cross-critique misses are exactly the ones that would have been caught by human expert review. That review did not happen. A specialist reader engaging with the publication will find errors of this kind.

The same limits show up across the four-week practice. The books are also imperfect. The Many Builders is also imperfect. The practice is open about being a work in progress. The honest version of "what AI tools enable a curious citizen to do" is "produce a remarkable amount of substantive work in a short period, with known defects that the author publishes openly rather than waiting until they are fixed." That is the practice. The IHT publication is one example.

What this implies

The substantive analysis in this publication concerns one specific UK tax-policy question. The analysis stands or falls on its own merits, and reasonable people will reach different views about whether it succeeds. That is the work and it should be judged as the work.

The production is a separate question. The IHT publication came together in roughly eight hours of real work because by then the workflow had been practiced — Doug, who cannot code, had spent earlier weeks of the four-week period using AI tools to build several websites and around 100,000 lines of code; the books took seven days each plus three of modification; The Many Builders was built alongside. Same person, same tools, very different intensities. Saying "same workflow produced everything" overstates the uniformity. The point is that the threshold for what a curious citizen with sustained attention can produce in short periods has moved. Several years ago this practice would have required a small team and a long timeline. Now it requires one person, a goal, multiple tools used in parallel for cross-critique, and the willingness to iterate.

The implication is not that AI replaces specialists, or that this work substitutes for the modelling HMRC and OBR could publish if they chose to. It is smaller: the relationship between expertise, citizen participation, and the tools that mediate between them is changing fast enough that paying attention to specific examples is more useful than waiting for the abstract argument to settle.

What is reasonable for a reader to do with this

If you have a question that matters to you, that the public debate is treating too quickly, and that you have the patience to iterate on, this is now possible. You do not need a publishing platform. You do not need a research budget. You do not need permission. You need a question you actually care about, the willingness to be wrong in public, the discipline to work with multiple AI tools rather than one (so they can cross-critique each other), and the patience to iterate when the first version is not yet ready. What you do not need to pretend you have is human expert review of the result; what you should do, instead, is be honest that AI cross-critique catches some errors and misses others, and invite specialist readers to point at the errors AI review did not catch. That is what produced this publication and the practice it sits inside, and the same combination will produce others.

If you find the work in this publication useful, the highest compliment you can pay it is to do something similar on a question you care about. The author would rather see ten attempts of this kind on different policy questions, half of them flawed and improving over time, than have this single publication treated as a finished thing.

If you find the work flawed — and it is, in places, real flaws have been named openly across the publication — the second-highest compliment is to engage with it specifically. Tell the author what is wrong. The work is published under CC BY-NC. The author would rather be corrected than carry the errors forward.

If you find the production interesting — the eight hours, the four AI tools, the four weeks of practice that sat behind those eight hours — the right move is not to share it as proof of what AI can do, because that is not what it proves. It proves what one curious citizen, working with AI tools and a goal, can do on one question over eight hours of focused effort, after four weeks of accumulated practice with the same tools across very different work. Whether that scales to other questions, other people, and other goals is the genuinely interesting question, and the only way to find out is for more people to try.

If you find The Many Builders interesting — and you should visit it — the right move is to look. The bear has made a page. Type a name. Then go and dig.

The author's actual view

The substantive view on the IHT question is in the principle piece and the timing piece: the principle of the reform is right, the strongest objection that lands is about the timing rather than the amount, and a realisation-based design would deliver the same fairness with fewer of the side effects critics worry about most. That is the publication's position on the IHT question.

The meta view is the one this page exists to surface, and it is bigger than the IHT question. AI tools have moved the threshold for what a curious citizen with sustained attention can produce — across registers, across formats, across questions, across creative ambition. The implications for the relationship between citizens, governments, the public conversation, and the tools that shape both are larger than the implications for any single tax reform. The IHT analysis is one demonstration. The Many Builders is another, and a stranger one. The trilogy is another. The bear books are another. The author thinks the bigger story is going to be a major part of the public conversation over the next several years, and would rather be a specific set of examples than one more voice talking about it abstractly.

The practice is not finished. Other questions deserve this kind of treatment. Some of them are coming. The author's hope is that other people will do the same on questions he cannot, and that the cumulative effect will be a richer public conversation about the world we are actually building than the current debate produces.

If you are reading this, you are part of how that turns out.

---

Two corrections logged on the corrections page: the workflow-honesty correction (1 May 2026) replacing earlier "architect" framing; the same-workflow / twelve-hours correction (1 May 2026) replacing the "twelve hours elapsed" framing with the eight-hours-of-real-work figure and naming the variation in intensity across the four weeks.