Chapter 3 · The Agent Playbook for Ecommerce Operators

The brand context file

One document, written once, pasted or uploaded into everything. It is the difference between AI copy that sounds like your brand and AI copy that sounds like every other brand in your category, and it is the reason most teams quietly give up in week three.

Published in full on the web, no email required. Roughly ten pages of the printed book. Last checked against tooling current in July 2026.

Here is the pattern behind almost every "we tried AI and the copy was generic" story. Someone opened a chat window, typed "write a product description for a cedar and sea salt bath soak", and got back a competent paragraph that could have belonged to any of four hundred brands. They concluded the tool was shallow.

The tool was not shallow. It was under-briefed. It had no idea that you never use the word "indulgent", that your customer is a 38-year-old who buys this as a recovery ritual after long runs rather than as a luxury, that the scent is described as "green" and not "fresh", that you cannot say "relieves muscle soreness" because you have no substantiation, and that your bullets are always five, always start with a verb, and never exceed 90 characters.

You know all of that. You have never written it down. That is the whole problem, and this chapter is the fix.

What this file is

A single plain document, 600 to 1,200 words at the start, containing everything about your brand that a talented freelance copywriter would need on their first day, and nothing they could look up themselves.

That test does real work. It cuts "we care about our customers" (everyone says it, it changes nothing) and keeps "never use the word 'detox' anywhere, in any context, including social" (specific, enforceable, changes output).

The seven sections

A complete example, annotated

This is a real-shaped file for a fictional but plausible brand: Salt & Cedar, a DTC bath and body company, roughly 140 SKUs, three people plus a fractional agency. Read it, then read the annotations, then write yours by substitution.

# Salt & Cedar: brand context for AI-assisted work

Last reviewed: 2026-07-14. Owner: Priya (brand). If something here is
wrong, fix it here first, then fix the asset.

## What we sell and against what

Bath and body products for physical recovery, not for pampering.
Bath soaks, muscle balms, body oils, and one bar soap. Price band
$18-$44. Our real competition is not other bath brands: it is a hot
shower and an ibuprofen. We are the deliberate version of the thing
people already do badly after hard physical days.

We are not a spa brand, not a clean-beauty brand, and not a wellness
brand. We do not use the word "wellness".

## Who buys

Segment 1 (about 60% of revenue): "the recreational athlete".
Trains 4-6x/week, 30-50, buys for themselves, repeat rate is high,
discovers us through running and lifting content. Their objection is
that this looks like a gift-shop product and they do not want to
smell like a hotel lobby. They respond to specificity: gram weights,
ingredient percentages, how long to soak.

Segment 2 (about 25%): "the giver". Buys the gift set in Q4 for
someone in segment 1. Never used the product. Their objection is
whether it will seem thoughtful rather than generic. They respond
to the story of why the formula exists.

The remaining 15% is unclassified and we do not write for it.

## Voice

Direct, technical, unhurried. Closer to a good gear review than to
beauty copy. We assume the reader is competent.

Rules:
- Short declarative sentences. Average under 18 words.
- Second person. "You soak for twenty minutes." Not "one soaks".
- Concrete over evocative. "300g of magnesium sulphate" beats
  "mineral-rich".
- We explain mechanism when we know it, and say nothing when we
  don't. Never gesture at science we cannot cite.
- Humour is dry and rare. One dry line per email at most, never on
  a PDP.
- No exclamation marks anywhere, including SMS.
- British English on the .co.uk store, US English everywhere else.

Register calibration:
- WRONG: "Indulge in a moment of pure bliss with our luxurious
  restorative soak."
- RIGHT: "Two cups, twenty minutes, water as hot as you can stand.
  Most people use it the evening of a long run."

## Words we never use

indulge · indulgent · pamper · luxurious · bliss · self-care ·
wellness · detox · toxins · glow · radiant · elevate · elevated ·
curated · artisanal · game-changing · must-have · nourish (of skin)
· "clean beauty" · "chemical-free" · "all-natural"

Constructions we never use:
- Rhetorical questions as openers ("Tired of sore muscles?")
- "Introducing..." as an email subject
- Three-adjective stacks ("rich, creamy, decadent")
- Long dashes. Use a comma, a colon, or two sentences.

## Product truth

The authoritative source for every product fact is the spec sheet
in Google Drive: "PRODUCT SPECS master (do not edit)". Fill weights,
INCI lists, scent notes, pH, shelf life, and allergen flags live
there and nowhere else.

STANDING RULE: never state a number, ingredient, percentage,
certification, sourcing origin, or shelf life that is not in that
sheet. If a fact is needed and missing, write [NEED SPEC: fill
weight] in the draft. Do not estimate. Do not infer from a similar
product.

Things we are frequently asked and get wrong internally:
- The bath soak is 900g, not 1kg.
- The balm is not vegan (beeswax). The soak and the oil are.
- We are not certified organic and never claim to be. Some
  ingredients are organically farmed; that is not the same claim.
- Manufactured in Portugal, formulated in the UK. Never "made in
  the UK".

## Claims policy

We may say:
- What is in it, and how much, from the spec sheet.
- What people do with it: "most customers soak for twenty minutes".
- Sensory description: scent, texture, temperature, how it feels.
- Magnesium sulphate content, and that it dissolves fully in hot
  water. Both are measurable.

We may never say, in any channel, without written sign-off from
Priya AND a substantiation file:
- That any product treats, relieves, prevents, reduces or heals
  anything: soreness, inflammation, cramp, DOMS, pain, anxiety,
  sleep problems, skin conditions.
- Any absorption claim ("magnesium absorbs through the skin").
  We have no evidence and the literature is contested.
- Comparative performance claims against a named competitor.
- Any "clinically proven", "dermatologist tested", or percentage
  of users who reported anything. We have run no study.

Customer reviews may describe outcomes in their own words. We never
edit a review, never write one, never solicit only positive ones,
and any review used in an ad keeps the reviewer's own wording.

Affiliate and creator content must carry a clear disclosure of the
paid relationship in the post itself, not in a bio or a linked page.

## Format conventions

- PDP title: [Product name] + descriptor + size.
  e.g. "Recovery Soak, magnesium and cedar, 900g"
- Bullets: exactly 5, each starts with a verb or a number, max 90
  characters, no terminal full stop.
- Meta description: 150-158 characters, includes the product name
  once, no ellipsis.
- Sizes in metric first, imperial in brackets on the US store.
- Prices without decimals when whole: $32, not $32.00.
- Sentence case for all headings and subject lines. Never Title
  Case, never ALL CAPS.
- Email subject lines: under 45 characters. Preview text is a
  separate sentence, never a continuation of the subject.

## Open questions

- We have no agreed line on sustainability claims. Until we do,
  say nothing about packaging beyond "the tub is HDPE #2".
- Segment 2 messaging is under-tested. Treat any AI output aimed
  at givers as a hypothesis.

Why each section is shaped that way

The positioning names an alternative. "Our real competition is a hot shower and an ibuprofen" is the single most useful line in the file. It gives every piece of copy something to push against. Positioning statements without a named alternative produce copy that floats.

The segments carry objections. Demographics generate nothing. An objection generates a headline, an FAQ, a bullet, and a subject line. Two segments with one sharp objection each will outperform six personas with lifestyle detail.

The voice section ends in a rewrite. Adjectives about tone are the weakest possible instruction. A wrong-to-right pair is worth ten of them, because the model can pattern-match against a concrete example. If you write nothing else in the voice section, write that pair.

The banned list is enforceable. Every other section requires judgement to audit. This one requires a search box.

Product truth includes a placeholder convention. [NEED SPEC: fill weight] is the mechanism that converts an invented number into a visible gap. Without a stated alternative to guessing, a model will guess, because producing a complete draft is what it was asked to do. Give it a legal way to not know.

The claims policy is written as may / may never / needs sign-off. Not as principles. A marketer under time pressure needs a lookup table, not an ethical framework.

Open questions are in the file. This is the section teams omit, and it is the one that prevents the model from confidently filling a gap you have deliberately left open.

What to leave out

Building yours in an afternoon

You do not write this from scratch. You mine it out of assets you already have, then correct the draft. That reverses the hard part: editing a wrong description of your voice is far easier than describing it from nothing.

1. Gather source material into one folder: your five best-performing product pages, three emails you were proud of, your last three ad scripts, your returns policy, and (importantly) two pieces of copy you disliked and rejected.
2. Run the extraction prompt below. Give it the good and the bad; the contrast is what makes the voice rules specific.
3. Correct the draft ruthlessly. Expect the banned list to be wrong: it will suggest bans on words you actually like and miss the ones you hate. Ten minutes of hand-editing here is worth more than any amount of prompt tuning.
4. Add the sections a model cannot infer: product truth, claims policy, open questions. These come from you and your compliance reality, not from your copy.
5. Test it with the drift test below before you rely on it.

Testing it: the drift test

Before you trust the file, prove it changes behaviour. This takes ten minutes and it is the only way to know whether you wrote a document or a decoration.

1. Pick a product you know cold. Ask for five PDP bullets with no context file attached. Save them.
2. Start a fresh conversation. Attach the file. Ask for the same five bullets with the same product data.
3. Put the two sets side by side. If a stranger could not tell which set came from a briefed model, the file is not doing its job.

The second version names the two inputs, cites the format section rather than restating it, gives a legal way to not know a fact, and ends with a self-audit that makes your review a thirty-second scan instead of a fact-checking exercise.

That trailing "list which field each bullet drew on" is a small habit with outsized returns. It converts review from "does this sound right" into "does this match the source", which is a question you can answer quickly and correctly.

Keeping it alive

Add a line for exactly four reasons, and no others.

• An output made the same mistake twice.
• Someone on the team corrected an AI draft in a way they will have to correct again next week.
• A product fact changed: a reformulation, a new size, a new certification, a discontinued claim.
• Your compliance position changed: you obtained substantiation for something, or you lost the right to say something.

Review it quarterly against your live PDPs, and treat the last change date at the top of the file as load-bearing. A file last reviewed fourteen months ago is actively dangerous: it will assert a fill weight that changed in the reformulation, and it will do so in every channel at once.

What to remember

• Generic AI copy is an under-briefing problem, not a model problem. The fix is one document, 800 to 1,200 words, attached to everything.
• The highest-yield sections are the banned word list, the wrong-to-right register pair, and the standing rule against inventing product facts.
• Give the model a legal way to not know a fact. [NEED SPEC: x] turns an invented number into a visible gap.
• Write the claims policy as may / may never / needs sign-off. A lookup table beats a principle when someone is publishing at 4pm.
• Mine the file out of your existing copy, including copy you rejected, then correct it by hand. Do not write it from scratch.
• Prove it works with the drift test, name an owner, and put a review date at the top.