# AI product descriptions that sound like your catalogue · ScriptGrain

> Build a voice profile from product descriptions you already like, brief each product with true facts, and score every draft against it.

Canonical: https://scriptgrain.com/ai-product-descriptions-in-your-brand-voice

# AI product descriptions that sound like the rest of your catalogue

*By Jack Stovell [X](https://x.com/StovBuilds) · [Instagram](https://www.instagram.com/stovbuilds) · [LinkedIn](https://www.linkedin.com/in/jackstovell/) · [YouTube](https://www.youtube.com/@StovBuilds) · [TikTok](https://www.tiktok.com/@stovbuilds) · published 2026-11-18 · updated 2026-11-18*

You add a new product on a Friday afternoon. The photos are good, the stock is counted, and you need a description by the end of the day. So you ask a tool for one, and what comes back is tidy, friendly and correct in its grammar. And it sounds like it was written by someone who has never read a word of your shop.

You can't quite say why. The other descriptions on the site are plain and a bit dry, perhaps, or chatty with the odd aside. The new one is neither. Put it between two old ones and the join shows.

This page is about closing that gap. It covers how to give a tool your existing descriptions as the sample, how to brief it, how to check the result, and what it costs in credits. ScriptGrain is the tool, so the details are about ScriptGrain, and I'll say plainly where it stops being useful.

## Why a new description reads like a stranger wrote it

A model asked for "a warm, professional product description" gives you its default. That default is an average of a great deal of writing, and averages have no habits. Your catalogue has habits. Maybe you favour short sentences, or you never open with the product name, or you use "you" a lot and "we" almost never. None of that is written down anywhere, and you probably couldn't list it if asked.

Adjectives don't fix it. "Make it punchier" works for a paragraph and then the defaults come back. What works is measurement: how long your sentences are and how much that length varies, how often you use contractions, and how formal you sound. How you open and which pronouns you reach for count too. Those are countable, and a draft can be held to them.

That is the idea behind a voice profile. It is built from writing you provide, and it records 45 attributes in 8 groups. Sentence length and its spread are in there, along with contraction rate, formality, openings and pronoun mix. You can read every attribute, so nothing about your voice is hidden from you.

## Your existing descriptions are the sample

For a shop, the sample is the descriptions already live on your site that you like. Not the ones you wrote at midnight and meant to redo. The ones you'd be happy to see copied.

The minimum is one sample of 50 words or more. That will build a profile, but a thin one. The [recommendation](https://scriptgrain.com/train-ai-on-your-writing-style) is three or more pieces and about 3,000 words in total, and you can add up to 20 samples.

Here's the catch with product descriptions: they're short. A single 150-word description is a small sample, and small samples give noisy measurements. So give it many. Twenty descriptions of about 150 words each comes to about 3,000 words, which is the target. It sounds like a lot until you remember you probably have more than that on the site.

Variety helps too. Include a product with a long story behind it and a plain accessory that needed two lines. The profile can then see what stays constant across both, and that constant is your voice. If every sample is the same kind of product, you risk teaching it the product rather than the voice.

Choose ones written without a committee. A description that went through four rounds of sign-off tends to lose its habits along the way, and you'd be handing over the committee's voice.

Building usually takes under a minute. The median is 39 seconds, measured on 28 September 2026.

### Getting the samples in

You can paste the text in. On paid plans you can also import samples from a public URL, which saves the copying if your descriptions are on live pages.

There is a second route called Brand Mimic. It builds a profile from up to 8 public pages of a website, and it needs at least 2 usable pages. It is for sites you own or have permission to use, and the request has to state that consent. Mimic is on paid plans, with a monthly allowance of 1 on Writer, 3 on Operator, 10 on Studio and 50 on Agency.

## Briefing a description

Product Description is one of ScriptGrain's 16 content types. The length runs from 60 to 400 words and the default is 150. Its structure leads with the benefit to the reader, then the concrete features that deliver it, then a light close. Benefit first, not a spec dump.

That order matters for a shop. It stops the draft opening with a list of dimensions and materials, which is how a lot of descriptions go wrong.

Now the part people skip. ScriptGrain does not know your stock, sizes, materials, prices or delivery terms. The brief supplies the facts. If you leave something out, the tool won't fill it in from knowledge of your shop, because it has none.

A company context can be added as well: your facts, your terminology, your audience. The generation prompt tells the model to use only facts present in the company context or the brief, and never to invent specifics such as pricing, metrics or capabilities. That is a useful guard, and it is not a substitute for reading what comes back.

A good brief is short. Four to six lines does it:

1. What the product is.
2. Who buys it.
3. Two or three features that are true.
4. The one thing to avoid.

That last line earns its place. "Don't call it luxury." "Don't mention the old version." One sentence of restraint saves a rewrite.

There's also an outline step, and it is free for this type. If you want to see the shape before spending a credit on a full draft, start there. You can also write up to three variants from one brief, each costing its own credit, which is handy when you can't decide between two angles.

[British or American English](https://scriptgrain.com/british-english-ai-writing) is a setting applied when the draft is written (it's called english_variant on the API), so "colour" and "color" are handled before you ever see the text.

## Checking the draft against the catalogue voice

Every draft is scored from 0 to 1 against your profile. On-voice drafts typically land between 0.85 and 0.95. Partial matches sit between 0.5 and 0.8. A different voice scores under 0.45.

The score comes with per-feature deltas, which show the draft's value next to yours. So you don't just learn that a draft is off. You might see, for example, that its sentences run longer than yours, or that it has no contractions where you use them freely. That tells you what to fix.

One thing the score is not: a fact check. It is a style check only. It checks no facts at all. A draft can score 0.93 and still say the throw is machine washable when it isn't. A person must check every spec, claim and price before publishing, every time.

Scoring any text against a profile is free (the [method is published](https://scriptgrain.com/reference/voice-measurement-framework)), including descriptions written by someone else. That gives you a handy trick: score your existing catalogue. If most of it comes back high, your voice is consistent and the profile will be sharp. If the scores are all over the place, you've learnt something about your site before generating anything.

### When the score is close but not there

Polish revises a draft toward a target score, with a default of 0.9, over up to three measured passes. It changes how things are said and never what is said, so your facts stay put. It costs 1 credit, and it is free when the draft already passes.

Then edit it yourself. You know the product. When you're done, save the final version: "Mark as final" in the app, or save_edit over MCP or the API, which is free. A daily job looks at your saved edits and suggests changes to the profile. Nothing changes until you accept. So the profile can follow your voice as it shifts, but only with your say-so.

## Working through a batch of products

There is no shop-platform integration and no bulk import of a catalogue. For each product you give a brief, and you copy the finished text into your shop. That is the honest workflow, and for a catalogue of a few dozen items it is manageable.

Here is a worked batch, invented for illustration. A shop sells three products: a waxed canvas tote, a ceramic pour-over coffee set and a wool throw. The shop owner has one profile, built from twenty existing descriptions.

The brief for the tote might read: a waxed canvas tote bag; bought by people who carry laptops and shopping in the same bag; has an inside zip pocket and long shoulder straps; avoid calling it waterproof. The pour-over set and the throw each get four to six lines in the same pattern.

One profile, three briefs, three drafts, each scored. If one draft scores lower than the others, the deltas show which habit slipped, and a Polish pass or a quick edit brings it back in line.

Then each description is checked by a person against the real product, copied into the shop, and saved as final so the profile can learn from the edits.

If you'd like the drafts somewhere other than the app, finished drafts can be sent to a Notion page or database with headings and lists intact. That suits a shop keeping its product copy in one place before it goes live.

One limit matters here. ScriptGrain matches the habits of the writing you give it. It cannot match a voice that your own catalogue does not have. If your existing descriptions are inconsistent, the profile averages them, and an average of three voices is a fourth voice nobody recognises. So choose the samples you would want copied. If you have two distinct tones on the site, build from one of them.

## What costs what

A draft of up to 5,000 words is 1 credit, whatever the length. So a 150-word description is 1 credit. Variants cost a credit each. Polish costs 1 credit and is free when the draft already passes. The outline step is free for this type, and so is scoring. Saving your final edit costs nothing either.

The free plan gives you one free draft for the life of the account, plus one profile and one full analysis, with no card needed. That's enough to build a profile from your catalogue and see whether the first description sounds right.

On paid plans, Writer includes 40 credits a month, Operator 75, Studio 400 and Agency 1,500 or more. Credits reset each cycle. Top-ups are available on Operator and Studio only, and the details are on the [pricing page](https://scriptgrain.com/pricing).

Profile limits are worth a look if you run more than one brand. Writer has 1 voice profile, Operator 3, Studio 10 and Agency unlimited. A shop with several distinct brands needs one profile per brand, because a single profile would blur them together.

## Using it from ChatGPT, Claude or your own code

If you already work in an assistant, you don't have to leave it. From ChatGPT, Claude, Cursor or any MCP client, the profile is read live over the MCP server. It has 41 tools and is available on every plan, Free included. One detail to remember: ask the assistant to score the draft, since it scores only when it calls the tool. It won't do it on its own.

If you'd rather write code, the REST API takes a profile_id, a brief and content_type "product_description" in a POST /v1/generations request. The limit is 60 requests a minute per key. A developer can loop over a product list and make one generation per product, which gets you closer to a batch without a catalogue import. The facts still come from the brief you feed in for each product, and a person still checks the output.

## Questions

### Does ScriptGrain know my products?

No. It does not know your stock, sizes, materials, prices or delivery terms. The brief supplies the facts, and a company context can add terminology and audience. The generation prompt tells the model to use only what is in the brief or the company context.

### Does a high score mean the description is correct?

No. The score is a style check only and checks no facts. Check every spec, claim and price yourself before you publish.

### Can it import my whole catalogue?

No. There is no shop-platform integration and no bulk import. You give the brief for each product and copy the finished text into your shop.

### What if my existing descriptions don't sound alike?

Then the profile averages them. Score the catalogue first, since scoring is free, and build the profile from the samples you would want copied.

### Do I need more than one profile?

Only if you have several distinct brands. Each brand needs its own profile, and the number you can hold depends on your plan.

### Can I choose British or American spelling?

Yes. It's a setting applied when the draft is written, called english_variant on the API.

## Related

- [AI that writes like you](https://scriptgrain.com/ai-that-writes-like-you)
- [AI writing in British English](https://scriptgrain.com/british-english-ai-writing)
- [Train AI on your writing style](https://scriptgrain.com/train-ai-on-your-writing-style)
- [Pricing](https://scriptgrain.com/pricing)
