315. Drafting with AI without losing the voice¶

Ask a language model for the benefits of a vector plug-in and it will give you ten. They will be fluent, plausible and identical to the ten it would give any other company. That list is useful, but not as a draft. It is a map of what everyone already says: the floor the copy has to rise above.

A model writes the statistical middle of what it has read. The house voice is what the middle leaves out: the named object, the condition that changes the outcome, the honest limit, the detail someone noticed because they looked. This chapter describes how to use models for marketing work without letting the middle replace the voice, and without letting a fluent sentence carry an invented fact into print.

What models are good at, and what stays human¶

Sort the work from objective to subjective before deciding what to delegate.

Usually delegate	Usually keep
Sorting support mail into themes	Choosing the main claim
Listing obvious benefits, to see what to avoid	Writing the headline and opening
Drafting FAQ answers from supplied facts	Deciding the offer and its terms
Variants of an approved line, for testing	Choosing which variant ships
Checking a draft against a list of rules	Quotations, testimonials, case studies
First drafts of specification prose	Launch announcements and founder notes

License and legal wording is never paraphrased by a model. Experience is human: a sentence that reports what someone saw, tried or decided must come from that person.

Know what good looks like before delegating. Write the first version of a page type by hand, then use a model to draft the next ones against it. A writer who has never written a release note cannot tell whether a generated one is good.

Give the model material, not adjectives¶

“Write a friendly, professional, exciting product page” produces the middle. Adjectives describe a voice; examples demonstrate it. A working brief supplies:

the fact sheet: names, versions, platforms, formats, prices, terms, dated;
the relevant changelog and manual pages;
a few approved pieces in the same register and surface;
the house rules, or the relevant skill;
the reader, the surface and the one thing the piece must achieve;
an instruction to mark every missing fact instead of supplying one.

A fictional working brief:

Write the opening section of a product page for owners of version 4. Use only the attached fact sheet, the version 5 changelog and the two approved pages. If a fact you need is not in them, write [NEED: what is missing]. Do not state a price, date, compatibility or performance claim that is not in the fact sheet. Before drafting, ask me up to five questions about the reader and the offer.

The last line matters. A model that interviews the writer first turns missing information into questions instead of guesses.

Markers make gaps visible¶

Use the same markers in every brief, so any draft can be scanned:

Marker	Meaning
`[NEED: …]`	A fact is missing
`[CLAIM: …]`	A claim needs evidence before publication
`[DECIDE: …]`	A policy or offer decision belongs to a person
`[ASK ME: …]`	The writer's own input was too vague to use

The guide's other placeholders ([VERIFY CLAIM], [CONFIRM OFFER], [CONFIRM LABEL]) work the same way. A draft built from thin material that contains no markers at all is a warning sign: the gaps have been filled. Count the markers before reading the prose.

Instead of asking “are you sure?”, ask the model to label each sentence with the document it came from, or “not in sources”. Then verify or delete every sentence in the second group.

Stage the work¶

One prompt that researches, outlines, drafts and polishes at once produces a draft that is hard to check. Separate the stages:

Facts. Gather and verify the facts; write the list of claims the piece may make.
Outline. Agree the structure before any prose.
Draft. Write against the outline, section by section.
Repetition pass. Remove points made twice.
Voice pass. Revise in manageable chunks against the approved examples.
Review. A separate session checks the draft against the rules.
Approval. A named person checks every fact and approves.

Tools save drafts; they do not publish or send. Keep a short log of the prompt, the sources, the reviewer and the date for anything substantial.

Ask for the check as well as the rules¶

“Follow these rules” asks the model to hold every rule in mind while it writes. A long rule set does not survive that: models drop instructions as their number grows, and the ones they drop are rarely the ones you would choose. Add a second instruction. After each section, the model goes through a short checklist one item at a time, names the passage that passes or fails each item, repairs the failures and checks the repaired passages again. Three rounds are enough; what still fails gets a marker instead of a claim that it passed.

The list should be short and observable. “Every price appears in the fact sheet” can be checked; “the copy is engaging” cannot. The checklists give the house version of the loop, and the writing skills end with one.

A separate reviewer¶

The session that wrote a draft is a poor judge of it. Review in a fresh session that sees only the draft, the fact sheet and the house rules. Ask it to flag, not to rewrite, and to name the rule behind each flag. Ask it to read as a doubtful buyer:

You own version 4 and doubt you should pay for version 5. Read the draft. Which sentence fails to answer you, and which claim would you not believe?

The writer keeps the flags that are real and ignores the rest. A reviewer that finds nothing is not proof of a good draft.

Before a person reads for voice, read once as a named, doubtful reader and mark every sentence that pushes; edit on a different day from drafting.

Models invent references with confidence: links that do not resolve, papers that do not exist, specifications from another version. The reviewer opens every link and checks every citation, figure and specification against its source before a person reads the draft for voice. Checking accuracy is not the whole review; the register is read by a person too.

A text goes out under a person's name only after that person has read and changed it. An article carries its real author's byline, not the name of whoever is most senior.

Numbers that circulate without a source¶

Some figures travel from book to book and slide to slide without a traceable origin, or contradict their own references. A model will offer them confidently. Do not quote them: the billion dollars Netflix's recommendations are said to save, the 35 percent of Amazon's sales attributed to its recommendations, the "rule of seven" exposures before a purchase, "95 percent of decisions are subconscious", and percentages for employee reach or influencer returns that the cited study does not contain. If a figure matters, find the primary source and cite it; otherwise leave the sentence without a number.

The strip pass¶

Search every AI-assisted draft for the places invented facts hide:

numbers, percentages and counts;
“guarantee”, “free”, “always”, “lifetime”, “instantly”, and “will” in a promise about results;
dates and version numbers;
social proof: “trusted by”, “loved by”, “thousands of”;
superlatives and adjectives any competitor could use.

Each hit stays only with a source. “Trusted by foundries worldwide” goes unless a named, permissioned foundry can be cited. A delay notice must not acquire a new estimated date on the way through a model.

Five failures that look like success¶

Before passing a draft on, check for:

an invented detail that sounds like the product;
a guess presented as a fact;
quiet truncation, where part of the requested content is missing;
a dropped instruction;
an answer to a different question from the one asked.

Each produces fluent, confident text, which is why they pass a quick read.

Build the voice from real edits¶

A voice document written in adjectives (“warm, expert, witty”) gives a model almost nothing. A better one is built from evidence:

Collect pairs of first drafts and their edited, approved versions.
Ask a model to list how the edits differ from the drafts.
A person approves each candidate rule and writes it as an observable behavior: “name the object in the first sentence”, not “be concrete”.
Date each rule. Delete rules nobody needs. Add a rule only when a mistake recurs.
Keep a review list of struck phrases with the reason and the usual repair.

That list is a prompt for attention, not a ban. As chapter 309 explains, a blacklist that forces a worse synonym makes the writing worse. “Unlock” is fine for a lock.

Keep the always-loaded voice core short, since every line competes for the model's attention. Load channel-specific files, price tables and compatibility matrices only for the tasks that need them. The house skills follow the same idea: a core for the rules that always apply, with worked cases kept in reference files.

Rotate the examples¶

Approved examples age. Prices change, versions change, the voice develops. Refresh the example set every few months with recently approved, human-edited pages, never with recycled model output, which would teach the next draft to imitate the previous model rather than the house.

Practice: brief, mark, review¶

Use this fictional packet:

A vector plug-in adds a tool that rounds a corner while keeping the sharp node underneath.
It works inside a host illustration app.
No price, trial terms or measured saving are supplied.

Write a brief for a model to draft a 120-word product-page section for illustrators comparing ways to round corners. Include the materials you would attach, the markers, and the instruction to ask questions first. Then write the reviewer prompt for a separate session.

Finally, take this fictional generated sentence and run the strip pass on it:

Trusted by thousands of illustrators, our revolutionary corner tool saves you hours every week and works flawlessly with every version of your design app.

List each hit, what evidence it would need, and the sentence that remains when only supported claims are left.

Checklist¶

The writer knew what a good version looked like before delegating.
The model received facts, approved examples and rules, not adjectives.
The brief required markers for missing facts and asked for questions first.
Markers were counted; a marker-free draft from thin material was treated as suspect.
Stages were separate, with an agreed outline before prose.
The brief asked for an item-by-item check after each section, with repairs and at most three rounds.
A fresh session reviewed the draft against the rules and named each rule.
The strip pass removed every unsourced number, guarantee and piece of social proof.
The five silent failures were checked.
A named person wrote or approved the lead claim, the offer and any quotation.
Voice rules come from real edits, are observable and dated; examples are refreshed with human-approved text.
Every link, citation, specification and source in the draft was opened and checked by a person; models invent references.