Receipts · Article 1 of 2
How these articles were made
A library arguing for AI disclosure owes you its own. Here is the machine that helped draft these pages, the linter that polices its voice, and the humans who sign.
The answer to our own question
Earlier in this library we argued that how do you use AI is now a normal client question, and that the firms who dread it are the ones without an answer.
Fair is fair. Here is ours, for the words you are reading right now, because a library that preaches disclosure and practices silence would deserve every hard question it got.
These articles were drafted with AI. More precisely, they were drafted in working sessions between the named authors and Anthropic’s Claude, inside a system we built for the purpose, and what follows is a tour of that system: what the machine does, what it is forbidden to do, and why the accountable names on these pages are human ones. We publish the tour because our own doctrine requires it, and because the system is a fair sample of what we build for clients.
The press came before the paper
The first thing built was not an article. It was the machine that makes articles.
Each piece in this library exists as a single master file, plain text with a small amount of structure, holding the argument and nothing else. A pipeline we wrote turns every master into three finished formats at once: the web page, a plain-text edition, and a branded Word document typeset to the same specification as our client deliverables. Change a sentence in the master, run the build, and the correction is true in every format simultaneously. No version of an article can quietly drift from the others, because there is one source and three projections of it.
The pipeline went together quickly, and the reason is the honest part, which is that almost none of it was new. The Word-document engine implements a formatting specification we had extracted, painstakingly, from a real engagement deliverable weeks earlier, including the file-format rules that cost us an afternoon to learn the hard way. The writing rules it enforces had already been written down for other purposes. Building the press was mostly assembly.
That is the pattern we keep finding with AI systems generally. The speed you see on the day comes from preparation nobody saw in the weeks before.
The linter that polices our voice
Before any article renders, the build runs every master through a style gate: a list of things our house voice forbids, checked mechanically, on every article, every time.
The list bans a certain famous punctuation mark beloved of machine prose. It bans the consultant clichés you already know, the breathless hype words, the stock invitation to put time on a calendar, exclamation marks, and even the wrong kind of quotation marks.
And it bans three ordinary adverbs that no marketing style guide would think to forbid, because they are not marketing words at all. They are the small verbal tells a language model reaches for when it wants to sound convincing. We noticed our drafting partner reaching for them, so we added them to the gate.
Here is the part we enjoy. This article cannot show you the list. Printing the banned punctuation mark, or any of the banned words, would cause the very build that produced this page to flag this master, and the flag would sit in the build output until a human dealt with it.
The rule is not a memo asking writers to be careful. It is a check the text physically passes through, and the fact that we have to describe the list rather than quote it is the cleanest demonstration of enforcement we can offer.
The humans are the gate
The machine drafts. It does not decide, and it does not sign.
Every article carries a named byline, and the byline is the accountability, exactly as our oversight article demands of any firm using these tools. Nothing here publishes until the named author has read it, verified it, and approved it, and the review standard is the one we prescribed in that article, applied to ourselves. Citations get pulled rather than pattern-matched. The vendor documents quoted in this library carry the dates we read them. Claims we could not verify are labeled as what they are, or cut. Where an article reports numbers from third-party coverage rather than from our own verification, it says so in the same sentence.
The division of labor is the one we recommend to clients, and it is not complicated. The machine is tireless and fast, and it reads everything you put in front of it, at any hour, without complaint and without fatigue. Humans hold the judgment, the relationships, and the signature. Drafting is where that division pays. The signature line is where it must never blur.
The receipt
Descriptions of process are cheap, so we close with the property that is not.
The day after this library was first drafted, we re-ran the entire pipeline from its packaged sources on a different computer, and compared the output to what had shipped, file against file, to the byte. Every article reproduced identically.
That check does two jobs at once. It proves the pipeline actually produces the published text, and it proves nobody quietly hand-edited anything outside the system afterward, because a single altered character would have surfaced in the comparison.
We run that kind of check because measurement you can trust is the theme of this whole shelf of articles, and because trust the process is a request while a reproducible build is a receipt. If your firm publishes under its own name, with or without AI in the drafting chair, the standard travels. Know how your words get made. Write the rules down somewhere a machine can enforce them. Keep a named human on the signature, and be able to show your work.
That is our answer to the question we told you to prepare for. We are glad to take follow-ups.