AI skills for the back office: teach AI your way of working

An AI skill is a reusable set of instructions that teaches an AI app one task, such as checking an invoice or writing a reminder in your tone. Unlike a prompt you paste every time, the AI loads the skill by itself when the task fits. ChatGPT offers skills in Business, Enterprise and Edu workspaces, Claude on every plan (as of 08/10/2026), Grok since 18/05/2026 (as of 20/10/2026).

What is a skill, and how is it different from a prompt?

A skill packages how you do a task: the steps, the output format, your tone, the checks at the end. It is a folder with one required text file, SKILL.md, written in plain Markdown. It can also hold templates and examples.

The difference from a prompt is about when the instructions are used:

• Prompt: you write or paste it in every chat. Good for one-off questions.
• Custom instructions or project: standing background that is always loaded. Good for "who I am and who my clients are".
• Skill: a task-specific procedure. The AI first sees only its name and short description. When your request matches, it loads the full text and follows it. That keeps chats light, even with many skills.

Like handing a new colleague the reminder process sheet once, so you stop explaining it every Monday.

Skills are an open format. Anthropic created it and released it as an open standard; many tools now read it (agentskills.io, as of 08/10/2026). A skill you write once works in other apps too.

Where do skills work today?

Status as of 08/10/2026 (Grok: 20/10/2026), from the vendors' own pages. Menus and plans change, so the vendor page wins.

OpenAI's pages differ on plans and beta status, and Free, Plus and Pro are not confirmed, so check your own settings. Also: skills from strangers are instructions your AI will follow. Install only skills you have read or that come from a source you trust.

If your tool has no skills, the text of a skill still works as instructions you paste into a chat, a project or a custom GPT.

Which back-office routines make good skills?

A routine is a good skill if you do it again and again, with the same rules, and can describe a good result. Typical candidates for freelancers, studios, agencies and small businesses in Germany and Austria:

1. Invoice check. Does the invoice have all mandatory details? Rules differ: see Invoice requirements Germany and Austria. Ready-made skill: Invoice checker DE/AT.
2. Reminders in your tone. Friendly first nudge, firm second, formal third, each with the right deadline. Background: Payment reminders. Ready-made skill: Payment reminders in your tone.
3. Hours to invoice lines. Turn raw time entries into lines a client understands. Ready-made skill: Invoice writer DE/AT.
4. Quote from a brief. Scope, items, assumptions, terms.
5. Receipt sorter. Name, date and file receipts; list what is missing for the month. Ready-made skill: Monthly receipt routine.
6. Month-end checklist. The same twenty minutes every month. Ready-made skill: Month-end checklist.
7. Handover note for your tax adviser. What happened, what is open, what you need to ask.

Poor candidates: one-off tasks, anything where the rules are unclear to you, and decisions that should stay with you, such as whether to write off an invoice.

Every legal skill should say which country it covers (Germany, Austria or both), carry a date and keep a human in the loop. AI prepares, you approve.

These five are ready in our free skill library, to download and use with or without Backbone.

How do you write a skill that works?

Keep it short:

1. Name and one-line description. The description decides when the AI picks the skill, so name the task and the trigger: "Use when checking an invoice before sending."
2. Inputs. What the AI needs: the invoice text, the country, the client type.
3. Steps. Numbered and concrete.
4. Output format. A table, a short email, a checklist.
5. Final checks. "List anything you could not verify. Never invent a VAT ID."

Then test it on three real cases, including one where the answer should be "something is missing". Fix the wording, test again.

How do you keep a skill correct when rules change?

A skill is a copy of your rules at one point in time. Treat it like a document with an owner:

• Date it. Put "checked on" and the country in the skill itself.
• Keep facts out of the steps. Put rates, thresholds and deadlines in one clearly marked block, so an update touches one place. Late-payment interest and base rates are reset on fixed dates (the current ones are valid until 31 Dec 2026), see Late-payment interest.
• Re-check on triggers. A change in law (for example the phased e-invoice rules in Germany, see E-invoicing Germany), a new year, a new client type, or a wrong answer.
• Add a rule of doubt. "If the date of this skill is older than twelve months, say so."
• Use one source. Keep the master copy in one place and re-upload it everywhere, instead of editing copies.

Our library skills carry a "checked on" date and are re-checked together with the guides here.

What changes when a skill can reach your real data?

A skill on its own works on what you paste in. Connect Backbone to your AI app and the same skill works on your real invoices, hours and receipts. The skill knows how you work, the connector has the data. "Check this month's invoices" without copying ten PDFs into a chat. Asking ChatGPT or Claude to prepare something creates a draft you review, and the AI asks you before it sends. How the connection works is in What is MCP?.