Gnothi AI for Marketers

The offer page's booking button gets a destination: a six-field Tally form, a confirmation and a first follow-up email, with every line traced to a source and invented promises replaced by placeholders for the founder. The form is then checked for phones, keyboards and screen readers, and weighed against native CRM forms.

This episode is narrated by AI-generated synthetic voices.

Episode page & show notes

Visit website

Six fields, each tied to a decision

The "Book a diagnostic call" button on the Ledgerlane offer page finally leads somewhere: a short form, a confirmation message and a first follow-up email. Every field has to feed a decision. Name, email and phone let the founder reach the lead. A one-van or two-van choice maps straight onto the two price rows. A trigger question (penalty letter, big year-end bill, something else) shows what the call is about, and an optional box comes last. Income, postcode, business name and "how did you hear about us" are left off. A qualifying note above the fields says what Ledgerlane covers and what it does not, including that penalty appeals are not part of the service.

Why Tally over Calendly, Typeform and Google Forms

Calendly's free plan has fixed intake questions. Its answer-based logic and redirects need a paid plan, as its guide to routing forms and pricing page show. Tally's free plan includes branching logic and redirects, as its own comparison with Google Forms describes, and it is hosted in Belgium, as its GDPR page states. Typeform's free plan caps answers at ten a month, according to Typeform's pricing. A test entry produces a stored record and a notification email to the account owner.

Catching invented promises

Claude drafts the labels, the confirmation and the email inside the Ledgerlane Marketing project, and tags each sentence to a source. The illustrative fault is a confirmation that promises a call "within 24 hours" for "a quick 15-minute chat". Both numbers are invented, because call length and reply time are still open owner questions. The fix is a marked placeholder for the founder to fill, not a softer guess. The follow-up email restates the price, the inclusions and exclusions, and the two-week setup. It asks only whether the lead is VAT-registered.

Checking it on a phone and without a mouse

Each field needs a visible label tied to it in the page's code, as WCAG's guidance on labels or instructions requires. Errors must appear as text, not just colour, and focus must not be hidden behind the sticky button. Text needs a contrast ratio of 4.5:1. The email and phone fields need the right keyboard hints. Tap targets must meet the WCAG 2.2 24-pixel minimum. Lighthouse, axe or WAVE catch only part of the problems that matter, so the manual checks are the real inspection.

Native forms versus a separate builder

A native HubSpot form turns a submission straight into a contact record and matches it against existing contacts by email. In 2026 HubSpot added in-CRM meeting booking in its July updates and spam controls for native forms in its September updates. Klaviyo added multi-step landing pages in its February release. Tally records will need a connection that checks for duplicates once a CRM arrives, and that requirement goes in the fault note.

Ledgerlane's offer page gets built section by section: the order of the promise, one evidenced objection and two labelled hypotheses, and a draft from the assistant with three sourcing faults fixed. The page is checked at phone width, and the saved reply instruction is tested in a fresh chat.

This episode is narrated by AI-generated synthetic voices.

Episode page & show notes

Visit website

A page in the order a reader needs it

Ledgerlane's offer page is written for a problem-aware trade contractor reading alone on a phone. It opens with the penalty-or-bill moment in contractors' own words. Copy research such as Copyhackers on writing landing pages recommends mirroring the visitor's symptoms above the fold. The next section covers the monthly timing that sets Ledgerlane apart from the once-a-year accountant. The price appears earlier than the usual pattern puts it: £180 a month for one van, £260 for two. Nielsen Norman Group on showing prices finds that buyers who see no price tend to assume it is too high. The exclusions sit beside the price, so a reader who needs a penalty appeal learns early that Ledgerlane doesn't handle one. Proof is limited to facts a reader can check, and the page has no invented testimonials. The button reads "Book a diagnostic call". The line under it about call length or reply time stays blank because no fact supports it.

Objections, evidenced and hypothetical

Three doubts are answered, and each FAQ question is phrased the way a visitor would ask it, in line with NN/g on FAQs that deliver value. "My accountant does this once a year" comes from an evidenced category. "Software is cheaper" and "Is it too late?" are labelled hypotheses because nothing in the evidence file quotes them. The "too late" answer gives the HMRC deadlines and penalty stages with their conditions and promises no outcome. It names what the page can't know, and that is why the call is the next step.

Fixing the assistant's draft

Each sentence carries a source tag. Three sentences fail:

One revision pass fixes all three. It also confirms the page implies no result from HMRC.

Checking at phone width

The chapter walks through Chrome DevTools device mode at widths of 320–390 px and notes that the tool only imitates a phone. The checks cover line breaks and the WCAG minimum target size of 24 px, or 44 px at the stricter level. A sticky button must never cover the exclusions or the penalty condition.

The saved instruction in a fresh chat

From early September to 12 October, Anthropic's documentation shows no change in how Projects apply instructions and files in new chats. That check excludes the September coordinator update and memory changes. A new inquiry from a two-van electrician about a large year-end bill passes all four checks. That shows the instruction carries over for this kind of message, not for every case.

A trade contractor's HMRC penalty letter becomes the test case for adapting a defended marketing message to one customer without adding promises, guesses or unsupported VAT claims. It ends with how a Claude Projects change spreads saved instructions across parallel threads.

This episode is narrated by AI-generated synthetic voices.

Episode page & show notes

Visit website

Reading the brief against one penalty letter

A one-van contractor with no bookkeeper writes to Ledgerlane after an HMRC late-filing penalty letter, asking for help and a price. Their message holds four facts, and anything else in a reply has to come from the brief, the product-facts register or the evidence file. The brief already covers this contractor in three places: the penalty trigger, a competitor that is the once-a-year accountant plus inertia, and the one-van price of £180 a month. It leaves three gaps: the penalty details, VAT status (the register's VAT row is still open with the founder), and the contractor's deadlines.

Deadlines and penalties go in as register rows, conditions included

The assistant doesn't recite tax rules from memory. Each deadline and penalty stage becomes a row taken from HMRC's guidance on Self Assessment penalties, with the condition kept in the same cell as the figure. "The penalty is £100" is false for many people. "A return up to three months late gets a fixed £100 penalty" is true. A Making Tax Digital row records the thresholds, including the drop to over £20,000 from April 2028. The reply may say the rules exist. It may not say they apply to this contractor, whose income is unknown.

The reply must also fit the CAP Code's rules on misleading advertising. It can't guarantee or imply that a penalty will be cancelled. It can't leave out that interest keeps building up during an appeal. Penalty appeals aren't one of Ledgerlane's inclusions, so the reply only points to the appeal instructions in HMRC's letter.

Auditing the assistant's draft

The request asks for a bracketed source tag after every changed sentence and a list of unknowns at the end. The first draft reads well and has four faults:

After one revision, the reply passes the new-claims audit and the voice check. It closes by asking the contractor three questions.

Situation versus person

The offer-page section speaks to everyone with a penalty letter. The personal reply uses only the facts this contractor supplied. A correct guess still teaches the reader that you invent things. The method is saved as a reusable "situation adaptation" instruction, with a two-van tax-bill inquiry as practice.

Parallel threads in Claude Projects

The chapter reports that from 17 September 2026, a coordinator can send project work to parallel threads, and each thread inherits the project's instructions and files when it starts. One stale file now reaches every thread, so update the register before starting a job. To check that the instruction works without this chapter's chat history, test it in a fresh chat against four checks. If a check fails, fix the instruction, not the draft.

Ledgerlane's facts sheet becomes a dated, owned register, and its brand guide is rebuilt from accepted and rejected examples. Both are tested on a fresh FAQ answer that slips in an unsupported VAT claim. The chapter closes with how Custom GPT retirement and Gemini file-reading failures affect keeping reference files current.

This episode is narrated by AI-generated synthetic voices.

Episode page & show notes

Visit website

Turning the facts sheet into a register

A standard only helps when a draft can be checked against it. Ledgerlane's product-facts sheet, which lists the one-van and two-van prices, four inclusions, three exclusions and onboarding "about two weeks from signing", becomes a register. Each fact gets its own row with a claim, value, source, named owner, date last checked, and status. Conditions like "from signing" stay inside the value. A header gives the file's owner and its next full check. Clearly labelled sections, as Anthropic's prompting guidance recommends, let the assistant see where one fact ends and the next begins.

Why a stale file slips through

A Claude Project treats every uploaded file as available knowledge. A newer upload does not replace an older one. When two files disagree, Claude may blend them, hedge, or pick one. The result is a confident email quoting last year's price. There are three fixes. Keep one current register. Delete the old file before uploading the new one. Tell the project instructions to use only current rows, to ask when sources conflict without dates, and to list the rows each draft used. A fresh-chat audit then confirms the project is clean. Instructions guide the model, but only removing the old file removes the risk.

A brand guide built from what passed and failed

Adjectives like "friendly" and "clear" rule nothing out. Paired examples do. Two pieces of accepted work are included: the plain onboarding email and the corrected hero section. Two rejected pieces sit beside them. One is the "stress-free, hassle-free" ungrounded draft. The other is an on-voice benefit sentence with no customer evidence behind it. The assistant drafts seven rules, each tied to an example. Two get cut: "light humour" and "warm, approachable tone". Five checkable rules remain. The guide is laid out as a dated header, a who-we-sound-like description, phrases to use and avoid, the rules, and the side-by-side pairs.

Testing both on a fresh request

A fresh chat produces an FAQ answer for a two-van contractor. It passes on voice, but it states the price "including VAT", and no register row supports that. The phrase is removed. A new VAT row is added with its status left open, assigned to the founder to confirm.

Keeping files current across assistants

OpenAI is retiring Custom GPTs on 11 December 2026 and moving them to Plugins and Skills (migration FAQ). Skills are called up only when ChatGPT judges them relevant, so voice rules may not apply every time. Rerun the fact-listing audit after migrating. Gemini Gems can silently fail to read complex PDFs or oddly named files. Plain text with clear headings is safer.

Ledgerlane's accurate hero copy speaks to nobody, so this chapter makes the audience decision out loud and turns it into a six-part brief with evidence line numbers behind every element. It covers picking a well-supported trigger situation over a thin one, naming the real competitive alternatives, splitting proof from inference, and the three citation faults found in the first draft — then how to reuse approved copy across channels without altering a single claim.

This episode is narrated by AI-generated synthetic voices.

Episode page & show notes

Visit website

Choosing one situation and declining the others

Ledgerlane's hero section is accurate and aimed at nobody. The missing piece is a decision no evidence file can make: which situation the message speaks to, and what it stops speaking to.

Three independent lines in the evidence file describe the same trigger — months of nothing, then a penalty notice or an unexpectedly large tax bill, then someone looking for help that week. One line stands alone: a contractor who hands his paperwork to his sister-in-law once a quarter. The chapter explains why a message built on a single line fails unreadably, and sets a promotion rule — two independent sources to reach emerging support, three to carry a message of its own.

The competitive set is five options: spreadsheet or glovebox, the once-a-year accountant, a spouse, self-managed software at ten to thirty-five pounds a month, and doing nothing. The incumbent is the year-end accountant plus inertia, which makes the difference a matter of when rather than how good — the fourth piece of positioning most drafts skip.

Boundaries are named: accountant and bookkeeper are unprotected titles, chartered designations are not, and anyone providing bookkeeping or tax services as a business must register for money-laundering supervision before trading.

The benefit is written for a problem-aware buyer in the vocabulary of the lines themselves. Exclusions work as a filter that improves who books. Proof splits into three piles, with the Self Assessment and CIS penalty timetable — one hundred pounds at a day late, ten pounds daily from three months, VAT registration at ninety thousand, Making Tax Digital from April 2026 — counted as proof of understanding the customer's year, not a claim about Ledgerlane. Client counts, testimonials and tax-saved figures stay out.

The draft came back with three faults: an uncited benefit sentence about confidence, the single-source line promoted into proof, and a citation claiming monthly visibility the lines never mention. All three, and the revision, are shown.

Getting more out of the approved copy

Adapting the hero section to other channels under claim parity: format and pacing change, every number, timeframe, qualifier and exclusion holds. Channel limits set the shape — email preview truncation, the See more cut, two hundred and eighty characters on X. The assistant is asked for a claim breakdown before the variant. When a mandatory qualifier will not fit, link to full terms or pick a different claim.

A landing-page draft built from traceable facts still doesn't tell you if it sounds like a real contractor, so the chapter builds a sourced file of verbatim customer language, separates observation from inference, and sets up a reusable brand brief so an AI assistant stops inventing personas and starts citing evidence.

This episode is narrated by AI-generated synthetic voices.

Episode page & show notes

Visit website

The founding charter

Ledgerlane has a hero section built from traceable numbers, but nothing yet tells you whether the words in it sound like a contractor's own week. Asking the assistant for a customer persona produces a fluent, plausible "Dave, forty-one, runs a two-van electrical firm" — and none of it was said by anyone. It was predicted, not gathered, and researchers call the error reification: treating an inferred sketch as data. The fix is a rule that governs everything else here — a line only counts as evidence if you can name the exact source.

That means going to where UK trade contractors actually write about money in their own words: review sites like Trustpilot and Google Business Profiles (bimodal, capturing fury and gratitude but nothing ordinary), trade forums such as electriciansforums.net threads on tax and bookkeeping and the ContractorUK accounting and legal board, search phrasing, inbound call notes, competitor FAQ pages, and short customer interviews. Each source is named and its bias made explicit — forums skew toward people trying to avoid paying anyone, reviews toward extremes, inbound notes toward people who already bought.

Interviews go wrong through politeness and bad self-prediction, so the chapter turns to the approach from Rob Fitzpatrick's Mom Test method for customer interviews: ask about a specific past occasion, not a hypothetical future, and check whether the person already spent effort on the problem.

The evidence gets logged verbatim — typos and jargon intact — using in vivo coding, Saldaña's coding method for qualitative researchers, keeping the customer's own labels rather than paraphrasing them. Every line then gets marked observation or inference, and inferences get a line-count threshold before they're trusted: three independent sources make a pattern, two an emerging finding, one a hypothesis that stays labelled no matter how appealing it is.

Explain AI-marketing changes and practical shortcuts

The assistant gets asked to group the evidence lines by situation, name each cluster from the words inside it, and cite line numbers — removing its room to invent. The result is checked the same way the hero section was: search for the quoted words, and count the cited lines against the actual file, watching for a smuggled-in extra quote that sounds right but isn't there.

A reusable brand brief then gets written once into the project's instructions layer — voice rules, a banned-words list, one accepted and one rejected example — while facts stay in the project's files, since Claude Code's CLAUDE.md hierarchy treats instructions and reference material as separate layers that can silently conflict without a stated precedence rule.

A walkthrough of fixing AI-generated marketing copy for a small bookkeeping business by building a short packet of real facts, audience details, a stated goal, and a genuine tone sample, then loading it into a project and inspecting the resulting draft claim by claim before publishing it, followed by a short note on a recent change to how Claude's memory works.

This episode is narrated by AI-generated synthetic voices.

Episode page & show notes

Visit website

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 notes the per-plan file limits of the alternative, ChatGPT Projects, plus Gemini's equivalent "Gems."

It runs the same request twice and lays the two drafts side by side: first in a bare chat with no project and no files, before the packet exists, then inside the built project. 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.