# How We Built a Brand-Voice Claude Skill That Ships 50 Product Descriptions in an Afternoon

**Author:** John Morabito (Founder, /winston)
**Published:** June 11, 2026
**Reading time:** 16 minutes
**Canonical:** https://www.winstondigitalmarketing.com/playbooks/brand-voice-claude-skill-product-descriptions/

Voice samples in, on-brand descriptions out, compliance gate built in. The full architecture of the Skill, the part most teams skip (negative examples), and why the compliance layer is the whole game in a regulated category like cannabis.

## The problem this solves

Every brand with a catalog has the same backlog: dozens or hundreds of product pages with thin, templated, or missing descriptions. The copywriter quote is real money and a six-week timeline. The intern-with-ChatGPT version is fast and reads like it. In a regulated category, the intern version is also a liability, because nobody taught the model what the brand is legally not allowed to say.

A brand-voice Claude Skill fixes all three at once: a packaged, versioned unit that loads the brand's actual voice, the category's generation rules, and a hard compliance gate every time. Write it once, batch fifty descriptions in an afternoon. Our first batch needed human edits on about 30 percent of descriptions in month one, and the edits fed back into the Skill, so the rate drops every cycle.

What a Skill is, in one paragraph: a folder. A SKILL.md instruction file plus reference files (voice spec, terminology, compliance rules, examples) that Claude loads when the task matches. Unlike a pasted prompt, it is versioned, complete every time, and enforces its checklist as part of the workflow. Skill vs Custom GPT decision framework: https://www.winstondigitalmarketing.com/playbooks/custom-gpts-vs-claude-skills/

## The architecture: four files

### 1. The voice corpus

15 to 30 samples of the brand's best existing writing. Not the brand guidelines PDF; the actual writing. Quality over volume: five great samples beat thirty mediocre ones, because the model reproduces whatever you feed it, including the mediocrity.

Annotate each sample with one line on why it works ("self-aware humor, never explains the joke"). The annotations matter more than the samples. They turn imitation into rules.

### 2. The voice spec

A one-page distillation, four sections:

- **Voice traits.** Three to five, each falsifiable, with a sample sentence. Not "friendly and approachable."
- **Sentence mechanics.** Average length, contraction policy, person, punctuation. The mechanical layer makes fifty descriptions feel like one author.
- **Vocabulary.** Words the brand owns and words it never uses. The never-list does more work than the always-list.
- **Negative examples.** The same product written in the brand's voice and in generic e-commerce voice, with "never produce the second one." The highest-leverage file in the Skill. Models learn boundaries from contrast, not adjectives.

### 3. The generation rules

Per-category structure so output drops into the product template without reformatting:

```
For each product:
- Hook line (under 15 words, no product name repeat)
- Body (60-90 words: what it is, who it is for,
  what it is like; one sensory detail, never two)
- Specs line (strain/dosage/format pulled from the
  data sheet verbatim, never paraphrased)
- No superlatives without an attribute behind them
- Output as JSON: {sku, hook, body, specs}
```

"Specs verbatim, never paraphrased" exists because the model will otherwise round 23.4% THC to "about 23%," and in a regulated catalog that is a labeling problem, not a style problem.

### 4. The compliance gate

The file that earns the fee. Cannabis is the strictest version, so it makes the best example. Per state where the client sells:

- **Prohibited claims.** No health or medical claims. "Relieves anxiety," "helps you sleep," "reduces pain" are violations in most markets, even when the reviews say exactly that.
- **Prohibited framing.** Nothing that could read as appeal to minors: no candy comparisons, no cartoon energy.
- **Required language.** State-mandated disclaimers attached by SKU type and destination state.
- **The gate itself.** The Skill's final step re-reads every description against the prohibited list and flags or rewrites before output. Violations get caught by the model, not by the one human reviewer skimming description forty-seven of fifty.

Related platform-rules problem: https://www.winstondigitalmarketing.com/playbooks/cannabis-brand-social-media-rules/

## The afternoon, hour by hour

| Hour | What happens | Human involvement |
|---|---|---|
| 1 | Product data exported as CSV. Skill loaded, calibration batch of 10 run. | High: review all 10, mark edits |
| 2 | Calibration edits fed back into spec and rules. Re-run the 10, diff against edits. | Medium: confirm fixes took |
| 3 | Remaining 40 run in batches with updated Skill. Compliance flags reviewed. | Low: spot-check |
| 4 | Full set human-reviewed, approved descriptions loaded to catalog. | High: final pass, ship |

The calibration loop in hours one and two is the difference between this and bulk slop. You measure the Skill against a human edit on a small set, correct, and only then scale. Same logic at article scale: https://www.winstondigitalmarketing.com/playbooks/agentic-content-pipeline-10-articles-a-day/

## What month one actually looked like

Roughly 30 percent of descriptions needed a human revision in the first month. Mostly voice drift on edge-case products and over-cautious compliance rewrites that stripped personality along with the violation. Every revision became a new negative example or rule. That feedback loop is the asset: a freelancer's learning stays in the freelancer; the Skill's learning compounds in a file you own.

## Where this goes next

The same architecture (corpus, spec, rules, gate) generalizes past product descriptions: category pages, email flows, social captions, and the crawlable menu-mirror pages for Dutchie-embedded dispensary sites (https://www.winstondigitalmarketing.com/playbooks/dutchie-iframe-seo-problem/), where every product page needs compliant, on-brand copy at volume.

Service: https://www.winstondigitalmarketing.com/services/ai-marketing/custom-gpts-and-skills/
Audit: https://www.winstondigitalmarketing.com/contact/#audit