AI Guides › Workbench
By Nigel Guy · 7 min read
Most people fix "AI-sounding" drafts by adding adjectives: "warm, punchy, conversational." Everyone in your niche asks for the same, so the model averages them and you get the voice of nobody. Each draft is fluent, so it feels fine, but you still rewrite every one by hand or paste the same style notes into every new chat.
The rule: describe your voice from evidence (your real posts and what already works in your niche), write that description down once as a Claude Skill, and judge the skill against a sample you held back, not against how it feels.
The chain has three links: a niche scan (what lands with your audience now), a voice card (what you actually do on the page) and a SKILL.md (the card packaged so Claude loads it when needed).
| Tool | What it does here | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Claude chat with web search | Runs the niche scan | Free (listed on every plan) | Seeing what currently lands | Public pages only; no engagement data unless you paste it |
| Claude chat or a Project | Builds the voice card | Free (Projects capped at five) | The voice breakdown | Mediocre samples, mediocre card |
| skill-creator skill | Turns the card into a skill | Included with Skills | Not hand-writing files | Ask to see every file it writes |
| Hand-written SKILL.md (ZIP upload) | Same skill, by you | Free | Full control | Folder structure must be right |
| Claude Code skills folder | Same file, locally | Pro or above | Terminal drafters | No sync with claude.ai |
So the whole chain can cost £0. Pro is listed at US$20 a month (US$17 billed annually); the page showed no sterling price when checked, so confirm the £ figure at checkout.
Skills run inside Claude's code environment, so they do nothing until code execution is on. According to Anthropic's help centre at time of writing:
Anthropic's developer docs still mention an older Settings > Features location; trust the help centre if they disagree.
You want the shapes of post your audience already rewards, not anyone's phrasing. Use a fresh chat with web search on.
You are a content analyst who studies [PLATFORM] posts in the [NICHE] niche, written for [AUDIENCE].
Goal: show me the patterns in posts that currently get strong responses, so I can decide which ones suit my own voice. A good result is a set of patterns I can act on, not a list of creators.
Steps:
1. Search public examples from the last [TIME WINDOW]. If you can't see engagement data, say what you used to judge a post as working.
2. Group findings into 5 to 8 patterns: opening, length, structure, point of view, type of proof, ending.
3. Describe each in your own words with one or two source links.
4. Flag patterns that look overused.
Output: a table with columns Pattern | What it looks like | Why it seems to work | Overused? | Sources.
Constraints: no made-up numbers, no guessed follower counts, no copying anyone's sentences.
If [NICHE], [PLATFORM] or [AUDIENCE] is vague, ask me before you search.
Before you answer, check that every pattern has at least one real link and that nothing in the table is a guess presented as fact.
Fill in: your platform, niche, audience and a time window such as "three months".
Most of the value is here. Paste 8 to 15 pieces you wrote yourself, unedited by others. Hold two more back; you'll test the skill against them.
You are an editor who describes a writer's voice precisely enough that another writer could copy it.
Goal: build a voice card from the samples below. I'll judge it by whether someone who has never read my work could use it to write something I'd sign without edits.
My samples: [PASTE 8-15 PIECES, separated by ---]
Patterns from my niche scan that I'm willing to use: [PASTE OR "none"]
Steps:
1. Read every sample before you write anything.
2. Describe what I actually do: rhythm, openings, endings, repeated words, avoided words, evidence, humour, jargon, punctuation, formality.
3. List 5 to 10 things I never do, with evidence from the samples.
4. Pick two short passages that best show the voice and say why.
5. Mark any trait that only shows up in one sample as "weak signal".
Output: a voice card with the headings Rhythm, Openings, Endings, Vocabulary (use/avoid), Evidence and tone, Never does, Best passages.
Constraints: only what the samples show; no flattery or labels like "authentic".
If there are fewer than 8 samples or they're mostly one format, tell me before you start.
Before you answer, check that every claim on the card points to something in my samples.
Fill in: your samples, plus any patterns from the scan you want to keep.
If the card says you "use vivid language" without an example, send it back.
Anthropic's anthropics/skills GitHub repository includes a skill-creator skill, and its README says the example skills are already available on paid claude.ai plans. Check it's switched on in Customize > Skills, then, in the chat holding your voice card:
Use the skill-creator skill to turn the voice card above into a Claude Skill.
Requirements:
- Skill name: [SHORT-NAME], lowercase with hyphens.
- Description under 200 characters: what it does, and to use it for [CONTENT TYPES] on [PLATFORM].
- Body: the voice card, the "never does" list, the two best passages as examples, and a final check against "never does".
- Leave out weak signals.
Ask if anything is unclear. Then show me every file in full before I install it.
Fill in: a short skill name, plus the content types and platform you want it used for.
A skill is a folder containing SKILL.md: YAML frontmatter, then Markdown instructions.
---
name: my-voice
description: Writes LinkedIn posts and newsletter drafts in my voice. Use when I ask for a draft, rewrite or hook for my audience.
---
# My voice
[Voice card goes here]
## Never
[Never-does list]
## Examples
[Two best passages]
## Before you hand over a draft
Check it against the Never list and remove anything that breaks it.
Per Anthropic's docs at time of writing: name is up to 64 characters of lowercase letters, numbers and hyphens, without "claude" or "anthropic"; the help centre caps description at 200 characters for claude.ai uploads (the API allows more). Zip the folder so it sits at the ZIP's root, then in Customize > Skills click the add button, choose Upload a skill and select the ZIP. The help centre writes the filename as skill.md, the developer docs as SKILL.md; if an upload is rejected, try the other casing.
Put the folder at ~/.claude/skills/my-voice/SKILL.md (all projects) or in a project's .claude/skills/. It doesn't sync with claude.ai, so update both copies.