The Source Packet That Makes a Draft Publishable

Building a landing page you can defend

The chapter walks through fixing a real failure mode in AI-written marketing copy: a fluent draft that invents prices, invents customers, and asks for the wrong action. Working with an invented small business, Ledgerlane Bookkeeping, it builds a short "source packet" — product facts with real prices and explicit exclusions, an audience paragraph, a one-sentence goal for the specific piece of copy, and a real sample of the business's own writing standing in for a tone description. That packet goes into a project, a container that holds standing instructions and reference files so every conversation starts from the same material. The chapter uses Claude Projects as the walkthrough tool, noting the file formats and size limits involved in uploading documents to Claude, and flags the account requirements for the alternative, ChatGPT Projects, plus Gemini's equivalent "Gems."

It then runs the same request twice — once with no project and no files, once inside the built project — and lays the two drafts side by side. The bare request produces confident, well-formed prose full of unsourced claims: a made-up company name, an invented discount, a wrong price. The project-backed draft ties every number to a specific line in the product facts, and the chapter inspects it anyway, checking claim by claim, hunting for anything stated with no source, comparing tone against the brand sample, and checking the closing line against the stated goal. Two real problems turn up — a dropped condition on the onboarding timeline and a missing mention of what the service doesn't do — and get fixed in one further pass. The chapter closes by separating out what the model contributes, what the project layer contributes, and what only the supplied packet contributes, and by warning that an assistant's confident description of "your typical customer" is not evidence of anything.

Where the assistants changed this month

A short update flags a change to how Claude's memory behaves following Anthropic's move to unify memory across chat and Cowork, covered in Anthropic's rundown of Claude's personalization features. Memory now writes during a session rather than in a daily batch, and it's sorted into editable topics. The practical point: brand voice and facts you'd defend to a client belong in files you control, not in whatever a chat happens to remember, and it's worth a quick audit of what's been saved.

An assistant will write you a landing page in eight seconds. That is not the hard part. The hard part is that the page it hands you says things about your business that are not true, in a voice that belongs to nobody, ending with an invitation nobody asked for. You can feel that something is wrong with it, but you cannot always say what, and you certainly cannot show a client what.

This chapter is about closing that gap, once, on one small piece of work. You are going to hand an assistant a short stack of facts about a business, ask it for one section of a landing page, look hard at what comes back, fix what is wrong and save the result. By the end you should have a piece of copy that could go on a real page for a real purpose, and you should be able to point at every claim in it and say where that claim came from.

Marketing with an AI assistant sits between two older ways of working. On one side is writing everything yourself, which is slow but where every sentence is yours and you know what it rests on. On the other side is the kind of tool that fills a template for you, fast but generic, because it does not know anything about your offer beyond the few words you typed in a box. A general assistant that you can give documents to sits in between, and that is its advantage: it writes at speed, but it can write from your material instead of from the average of everything on the internet. The whole skill is in what you hand it. That is why the first working session is about a stack of paper, not about clever wording.

So let me introduce the business we are going to work on.

Ledgerlane Bookkeeping is a small firm doing monthly bookkeeping and tax-ready accounts for one- and two-van trade contractors: electricians, plumbers, a roofer with a mate. Its customers lose receipts in the footwell of the van, do their paperwork at eleven at night on a Sunday, and file late. Every fact I give you about Ledgerlane is illustrative. It is not a real company, its prices are invented for teaching, and you are not expected to adopt it or to work in bookkeeping. What transfers is the shape of the work: a service with a defined buyer, a specific problem, a price, and things it deliberately does not do. If you sell software, run a yoga studio or freelance as a designer, you have all four of those. You will build the same stack out of your own facts and follow the same steps.

Here is the stack. There are four items in it, and I am going to call it a source packet, which just means the small set of documents you hand the assistant before you ask for anything. Each item exists because something specific goes wrong when it is missing.

The first item is the product facts. This is the least glamorous and the most useful. For Ledgerlane it reads roughly like this. Ledgerlane does monthly bookkeeping for sole traders and small limited companies in the building trades, one or two vans, up to about eight hundred transactions a year. One van is a hundred and eighty pounds a month. Two vans is two hundred and sixty. That covers photographing receipts into the app as they happen, a monthly reconciliation of the bank feed, quarterly value added tax return preparation, and year-end accounts handed over tax-ready. It does not cover payroll. It does not cover chasing customers for unpaid invoices. It does not give tax planning advice about whether to incorporate; Ledgerlane will say when that question is worth asking and refer it out. Onboarding takes about two weeks from signing to the first clean month, because the previous year's records usually need untangling first.

Notice how much of that is about limits. Price, scope, and the three things the service refuses to do. Those are the sentences an assistant cannot possibly guess, and they are exactly the sentences that stop a draft from over-promising. Without a product facts document, an assistant fills the gaps with what similar businesses tend to say, and what similar businesses tend to say is where invented claims come from.

The second item is the audience, and one honest paragraph is enough at this stage. Ledgerlane's buyer is a contractor doing between sixty and two hundred thousand pounds a year, working on the tools himself five days a week, who has no bookkeeper and no real system. He is not confused about accounting in principle. He is out of time. His paperwork is a carrier bag and a banking app, and the thing that finally makes him act is a late-filing penalty or an accountant's year-end bill that arrived larger than expected. He does not read the word "solutions" and feel anything.

That paragraph does two jobs. It tells the assistant who is being spoken to, which changes vocabulary and length. And it tells the assistant what the buyer already knows, which stops the draft from explaining bookkeeping to a man who has been invoicing for fifteen years.

The third item is the goal, and it is one sentence about this specific piece of writing, not about the business. Ours is: the top section of the Ledgerlane landing page, for contractors arriving from a search for help with late accounts, whose single job is to get them to book a twenty-minute call. Written down like that, the goal quietly settles a lot of arguments. It tells you the section is short. It tells you the call to action is a call booking and not a newsletter and not a free trial. And when you come to inspect the draft, it gives you something to measure the last line against.

Skip the goal and you get copy that is pleasant and asks for nothing, or asks for four things.

The fourth item is a brand example: one clearly labelled piece of writing whose tone counts as acceptable. Not a description of the tone, a sample of it. Adjectives like "friendly but professional" mean almost nothing, because every business on earth would tick that box. A paragraph shows what you actually mean. Ledgerlane's example is a bit of its own onboarding email, which I am labelling as illustrative too:

"Send us the last twelve months of bank statements and whatever receipts you still have. If there are gaps, say so now rather than later — we would rather know. Most of the mess we see is normal, and none of it is interesting to us. We just want the numbers to be right."

Short sentences. Plain words. No exclamation marks. It says the unglamorous thing out loud. Whatever your own business sounds like, find one real piece of writing that sounds like it and put that in the packet instead of a list of tone words.

Four files, then, and each one is short: the product facts, the audience paragraph, the goal, and the brand example. That is the packet. Now we need somewhere to put it.

The place to put it is a project: a container inside the assistant that holds standing instructions and reference files, so that every conversation you start inside it already knows your material. Both Claude and ChatGPT call this a project, and Gemini's version is called a Gem. I am going to work through Claude Projects here, because on Claude a free account can create up to five projects, so nobody needs to buy anything to follow along. If you would rather use ChatGPT Projects, know the account requirement first: creating a project there needs a paid plan such as Plus, Pro, Team, Business, Enterprise or Edu, and sharing a project with teammates is limited to the Business, Enterprise and Edu plans. Gemini's Gems are available on personal Google accounts for users thirteen and up, and on supported work or school accounts with access to the Gemini apps.

We also need to name the assistant, because we will be referring to it for a long time. Ours is called Ida. The name is a convenience, nothing more. It does not change the model or grant it anything; it just means I can say "give Ida the packet" instead of a longer phrase every time.

Setting up in Claude goes like this. Open Projects in the left sidebar and choose to create a project, then give it a descriptive name — Ledgerlane Marketing will do. On the right you will find a Project Knowledge panel, and inside it the option to set project instructions. Those instructions are the standing brief: they apply to every conversation you start in this project, so you write them once instead of retyping them. Ours say something close to this. You are Ida, a marketing writer for Ledgerlane Bookkeeping. Every factual claim about the service — price, scope, timing, what is included — must come from the product facts file. If a fact is not in the files, do not invent it: say what is missing and ask. Match the sentence length and vocabulary of the brand example. Write for the audience described in the audience file, and do not explain bookkeeping basics to them.

That last-but-one instruction matters more than it looks, and here is the mechanism behind it. Uploaded files become the assistant's factual ground truth, but the model does not stop using what it learned in training. It keeps using that general knowledge for grammar, for structure, and for anything your files leave out — quietly, and in the same confident voice. The only thing that changes that behaviour is an explicit instruction to work from the supplied sources and to say when something is absent. So write that instruction in.

Then upload the four files through the add-content control in the Project Knowledge panel. Claude accepts the ordinary document formats you already have: portable document format files, Word documents, plain text, comma-separated values, rich text, and standard images among others; spreadsheets in the Excel format need code execution turned on to be read properly. The cap is thirty megabytes per file, and there is no fixed limit on how many files, only the practical limit of how much text the model can hold at once. On paid tiers, once your knowledge grows past that, Claude switches automatically to retrieval indexing, which means it searches the files and pulls in the matching passages rather than holding everything at once. For four short documents none of that is in play. Worth knowing where the ceilings are on the other two, though: ChatGPT Projects allow up to twenty files in a project, with a generous half-a-gigabyte-per-file cap, and Gemini Gems allow up to ten attached source documents.

Now, before trusting any of it, we do the demonstration. Two requests, same asset, and the difference is the lesson.

The first request goes into an ordinary chat with no project, no files, no instructions. It is the bare version most people actually type: write me a landing page hero section for a bookkeeping business for tradespeople. Here is the sort of thing that comes back, and I am inventing this draft to show you the failure mode:

"Stop Drowning in Paperwork. Focus on What You Do Best. At BookkeepingPro, we're passionate about empowering hardworking tradespeople with seamless, stress-free financial solutions. Trusted by thousands of contractors nationwide, our award-winning team handles your books so you can get back on the tools. Plans start from just $99/month. Get started today — your first month is free!"

Read that with the product facts next to you and count what is wrong. The business is not called BookkeepingPro. Nobody said thousands of contractors, or nationwide, or award-winning, or anything about a free first month — those are four claims with no source, and any one of them could be a legal problem as well as an embarrassment. The price is wrong twice: wrong number, wrong currency. "Seamless financial solutions" is the borrowed voice, the sound of every service company at once, and it is the exact register the brand example rejects. And the call to action, "get started today", does not match the goal, which was to book a twenty-minute call.

Here is the thing to understand about that draft: it is not badly written. The sentences work. The rhythm is fine. Which is the trap. The generic request produced competent prose with nothing underneath it, and competent prose is hard to argue with until you have something to check it against.

Now the same request inside the project. In a new chat in Ledgerlane Marketing, the ask is short, because the instructions and files carry the rest: using the product facts, audience and goal in the project files, draft the top section of the Ledgerlane landing page — headline, one short paragraph, three brief lines on what is included, and one call to action. Match the brand example. After the draft, list each factual claim and the file line it came from.

Here is the draft that comes back:

"Your accounts, done every month, so year end stops being a shock. Ledgerlane does the monthly bookkeeping for one- and two-van trade contractors. You photograph receipts as you get them, we reconcile the bank feed every month, and your year-end accounts arrive tax-ready instead of arriving as a carrier bag. Monthly reconciliation of your bank feed. Quarterly value added tax return prepared. Year-end accounts handed over tax-ready. From a hundred and eighty pounds a month for one van. Most books need untangling before the first clean month, so we usually take about a fortnight to get started properly. Book a twenty-minute call and we will tell you what your first month would involve."

Then the claim list: one hundred and eighty pounds for one van, from the pricing line of the product facts. Receipt photography, monthly reconciliation, quarterly value added tax preparation and tax-ready year-end accounts, all from the inclusions line. Two weeks to the first clean month, from the onboarding line. One- and two-van contractors, from the audience file.

That is better, and the improvement is not luck. Every number in it exists in a file. But do not accept it yet, because a draft that cites its sources can still cite them wrongly, and the inspection is the part of this work that actually makes copy publishable.

Inspect in four passes, and treat it as a real operation rather than a read-through.

The first pass is claim tracing. Take each factual statement in the draft, one at a time, and find the line in the product facts that supports it. Not the file — the line. The price traces cleanly. The inclusions trace cleanly. The two-week onboarding traces to the onboarding line, and here it is worth being exact, because the draft said "about a fortnight" while the facts say about two weeks from signing to the first clean month. A fortnight is two weeks, so the number holds, but the draft has dropped "from signing", which is the thing that makes the two weeks make sense. That is a small softening rather than a false claim, and I would put the condition back.

The second pass is looking for claims with no source at all, and there is one. Nothing in the draft is untrue, but nothing in the product facts says what the service does not do, and the section does not mention it either. Payroll, invoice chasing and incorporation advice are all outside the scope. A contractor reading this could easily book a call believing payroll is included. That is not a factual error in the draft; it is an omission that will produce a wasted call and a disappointed prospect. Worth fixing on the page rather than in the call.

The third pass is tone, checked against the brand example rather than against your taste. Read them next to each other. The brand example uses short sentences, plain words, no exclamation marks, and says the unglamorous thing plainly. The draft mostly matches, and "instead of arriving as a carrier bag" is exactly the register — concrete, slightly dry, no flattery. One line drifts: "so year end stops being a shock" is a promise about feeling rather than a statement about work. It is mild. It stays, because it earns the reader's attention and does not claim anything checkable.

The fourth pass is the call to action against the goal. The goal was a booked twenty-minute call. The draft says book a twenty-minute call and tells the reader what the call will produce. That matches, and matching is not automatic — this is the check that catches a draft ending in "sign up now" for a page whose only conversion is a phone call.

So the inspection found two things: a dropped condition on the timing, and a missing scope line. One revision pass, and I ask Ida for exactly that, naming the fixes: restore that the two weeks runs from signing, and add one plain line saying payroll, chasing unpaid invoices and incorporation advice are not included, drawn from the scope lines in the product facts. Change nothing else.

The revised section reads:

"Your accounts, done every month, so year end stops being a shock. Ledgerlane does the monthly bookkeeping for one- and two-van trade contractors. You photograph receipts as you get them, we reconcile the bank feed every month, and your year-end accounts arrive tax-ready instead of arriving as a carrier bag. Monthly reconciliation of your bank feed. Quarterly value added tax return prepared. Year-end accounts handed over tax-ready. From a hundred and eighty pounds a month for one van, and about two weeks from signing to your first clean month, because last year's records usually need untangling first. We do not run payroll, chase your unpaid invoices, or advise on whether to incorporate — we will tell you when that question is worth asking. Book a twenty-minute call and we will tell you what your first month would involve."

That is publishable for its stated purpose. Be precise about what the inspection establishes, though, because this is where people overclaim. It establishes that every factual statement on the page is supported by Ledgerlane's own product facts, that the scope is stated rather than implied, that the voice matches a piece of writing the business has already approved, and that the section asks for the one action the page exists to get. It establishes nothing about whether contractors will respond to it. Whether "year end stops being a shock" is the thing that moves this buyer, or whether the late-filing penalty should be named right there in the headline, is a question about real customers, and no amount of inspection against your own files can answer it.

Last step, and it takes thirty seconds. Save the finished section, and save the brief that produced it in the same place — the request you typed, plus a note of which files were in the project and the two fixes the inspection found. Keep the brief for three reasons. It is a record of where each claim came from, so in four months when someone asks whether the page can say a hundred and eighty pounds, you have the answer without reconstructing it. It is reusable: the next section of the page, and the email version of this offer, start from that brief with the goal line changed rather than from nothing. And it tells you what to revisit when a fact changes. Raise the one-van price and the brief shows you every asset built on the old number. Throw the brief away and you keep only the output, which means every future check starts from scratch.

Now, three different things contributed to that page, and being able to name them separately is what lets you diagnose a bad result instead of just retrying it.

The first is the underlying model: the general language ability, trained on an enormous amount of text. That is what produced fluent English, the rhythm of a headline, the instinct to put the price after the inclusions. It is also what produced "seamless financial solutions" in the first draft, because on the general average of the internet, that is what a bookkeeping page says. The model supplies capability and the pull toward the average. It supplies no knowledge of you.

The second is the assistant layer: the project, the standing instructions, the uploaded files, the interface. This is the part that makes your material available and keeps it available across conversations, and it is what turns one careful setup into every future chat starting from the same footing. It is worth knowing that this layer has two ways of working. When your files are small enough, their text sits directly in the model's context. When they get large, the system switches to retrieval — searching an index and pulling matching excerpts into the request. Either way, the assistant layer decides what the model can see; it does not decide whether the model prefers it.

The third is the supplied context: the packet. This is what made the second draft specific and defensible. A hundred and eighty pounds. Two weeks from signing. No payroll. Short sentences and no exclamation marks. None of it came from the model, and none of it came from the project container. It came from you writing down four short things.

Which is why, when a draft comes back wrong, you can usually tell which layer failed. Fluent but invented facts means the context is thin or the instruction to stay inside it is missing. Correct facts in the wrong voice means the brand example is missing or is a description instead of a sample. Correct everything but the wrong ask at the end means no goal was stated.

One more thing to be clear about, because it will save you from a genuine mistake. If you ask Ida to describe Ledgerlane's typical customer, she will produce a convincing paragraph about a forty-two-year-old electrician who worries about penalties. That paragraph is not evidence. It is the model's general knowledge of tradespeople, written in the voice of a research finding, and it will feel more solid than it is. A persona an assistant writes tells you what such a person is commonly said to be like. It does not tell you what your customers say, what words they use, or what would make them trust you enough to book the call.

So the next step is finding out. Real customer language, from interviews, reviews, support questions and what people actually type into a search box. That is where a message you can defend begins, and it is where the work goes from here.

Where the assistants changed this month

One recent change touches the setup we just built, so it is worth knowing before you go and build yours.

On the twenty-fifth of August, Anthropic unified Claude's memory across ordinary chat and Claude Cowork. Two things changed in how that memory behaves. Updates now get written during the session itself rather than being gathered up in a daily batch, and the remembered material is sorted into topics you can open and edit, under the memory section of settings. It is switched on by default for the Free, Pro and Max plans, and switched off by default for Team and Enterprise. Sensitive categories, such as health or personal beliefs, are excluded by default.

The marketing decision this touches is where your brand context actually lives. Until now, if you wanted Claude to know how Ledgerlane writes, you put it in the project: instructions and files, deliberately placed, easy to point at. With memory on, things you happened to say in conversation can persist too, and can carry into later drafting work without you re-uploading anything. That is convenient, and it is also how a rule you invented on a Tuesday afternoon for one email ends up quietly shaping copy a month later. Persisted chat remarks are not a reviewed brand document. The packet is.

The smallest useful action is an audit, and it takes a few minutes. Open the memory section of your settings, look through the topics that are active, and read them as if they were instructions someone else wrote for you — because in effect they are. Confirm that anything about voice or brand still reflects what you would actually approve. Delete the leftovers: notes from a one-off task, an old price you were testing, a preference about a campaign that finished. Then keep the standing brand rules where you can see them, in the project instructions and the packet files, rather than trusting them to accumulate on their own.

Two habits to keep separate, then. Facts and voice you would defend in front of a client belong in files you control and can hand to someone else. Everything memory accumulates is convenience, and convenience gets audited.