AI Guides › Workbench
By Nigel Guy · 6 min read
Most solo owners treat Claude as one general assistant and re-explain their business in every chat. The "department" version is worse: they download a stack of skills from a list, switch them all on, and never check what any of them does. A skill is not a team member. It is a folder of instructions that Claude loads when your request matches it, and it is only as good as the job you point it at.
The rule: give each department one skill you have read, tested on one real task, and kept only if the output was better than your plain prompt.
A skill is a folder containing a SKILL.md file. The file starts with a short header (a name and a description) followed by instructions, and the folder can also hold reference files and scripts. Anthropic's documentation describes three layers: Claude always sees the name and description, reads the body of SKILL.md only when your request matches the description, and opens bundled files or runs scripts only when the instructions call for them. That is why you can have many skills installed without clogging the conversation. It also means the description matters most: it is what Claude matches your request against.
These are real folders in Anthropic's anthropics/skills repository (checked 2026-10-04). The repository says its skills are for demonstration and education, so test before you rely on them. I verified the folder names and the document skills' behaviour in the docs; I could not read every SKILL.md for the others, so open each one before you trust my one-line summary.
| Department | Skill | What it does | Best for | Catch |
|---|---|---|---|---|
| Finance | xlsx |
Builds and edits spreadsheets, analyses data, makes charts | Forecasts, expense sheets, tidy-ups of messy exports | Check every formula; it does not know your VAT position |
| Admin and operations | docx |
Creates and edits Word documents | Contracts drafts, SOPs, proposals | A drafted contract is still a draft; get it checked |
| Sales | pptx |
Creates and edits PowerPoint decks | Pitch decks, client updates | Layout is generic until you supply your own brand details |
| Marketing | brand-guidelines |
Applies a brand's colours and typography to output | Consistent look across documents | In the repo this is an example built around Anthropic's own brand; you will need your own version |
| Internal comms | internal-comms |
Helps write structured internal updates | Status reports, team and client updates | Read its SKILL.md to see which formats it covers |
Cost at time of writing: the pre-built document skills (xlsx, docx, pptx, pdf) are built in and need no separate fee, but skills need code execution switched on, which sits inside your Claude plan. Pro is $20 a month, or $17 a month billed annually (about £15 to £16 at time of writing); Max starts at $100 a month (about £75); Team standard seats are $25 monthly or $20 annual. Check the £ price at checkout. Availability on the Free plan is unclear: Anthropic's help centre says skills are available on Free, Pro, Max, Team and Enterprise, while the pricing page comparison suggests they are not on Free. Check your own Settings before assuming.
Anthropic's repository held 19 skill folders when I checked, including skill-creator, doc-coauthoring, webapp-testing, mcp-builder and claude-api alongside those above. Beyond that, anyone can publish skills, and there is no official count. Treat third-party collections as software from strangers.
In claude.ai (Free, Pro, Max):
Team and Enterprise: an owner must first check that Code execution and Skills are enabled under Organization settings > Plugins & skills (Policy tab). Members then use Customize > Skills. Custom skills uploaded to claude.ai belong to the individual and are not centrally managed.
In Claude Code: put a folder with SKILL.md in ~/.claude/skills/ (all your projects) or .claude/skills/ (one project). To use Anthropic's examples, run /plugin marketplace add anthropics/skills, then /plugin install example-skills@anthropic-agent-skills (or document-skills@anthropic-agent-skills). The built-in document skills are not available in Claude Code.
Skills do not sync between claude.ai, Claude Code and the API. Upload separately to each place you use them.
For the gaps (your brand, your invoicing rules, your tone), use the skill-creator skill or write a SKILL.md yourself. Fill in one card per department before you write anything:
| Field | Your answer |
|---|---|
| Trigger | The exact phrases you type when you need this |
| Inputs | What Claude must ask you for |
| Output | Format, length, file type |
| Rules | Your standing rules (house style, VAT wording, never-say list) |
| Test | One real task and what "good" looks like |
Keep the description short. The help centre says 200 characters maximum; the API docs allow 1,024, so stay under 200 and it works either way. Use lowercase letters, numbers and hyphens for the name, and avoid "claude" or "anthropic" in it.
Prompt to build a skill from your card:
You are helping me write a SKILL.md for my business.
Business context: [WHAT_YOUR_BUSINESS_DOES_AND_WHO_YOU_SERVE]
Department: [FINANCE_OR_SALES_OR_OTHER]
Trigger phrases: [PHRASES_YOU_TYPE]
Required inputs: [WHAT_CLAUDE_MUST_ASK_FOR]
Output format: [FORMAT_AND_LENGTH]
Standing rules: [YOUR_RULES]
Goal: a SKILL.md with a lowercase-hyphen name, a description under 200 characters that says what it does and when to use it, then ordered steps, the output format, and a short list of things it must never do.
Rules: if any field above is empty, ask me for it before writing. Do not invent business facts, prices or legal wording. Before you answer, check the description length and that every step is something I can test.
Prompt to test it:
Use the [SKILL_NAME] skill on this real task: [PASTE_ONE_REAL_TASK].
Then tell me which instructions from the skill you followed, which you ignored, and what you had to guess. Do not fill gaps silently; list questions instead.
Start with the department where you lose the most hours to repeat formatting, usually finance sheets or client documents. Do not start with the most impressive skill. Add the second only after the first has produced one result you would send to a client.
SKILL.md and any scripts first.brand-guidelines as is. It carries someone else's brand.brand-guidelines and internal-comms; the table describes their purpose from their names, so confirm before use.