Anthropic Skills — bottled, scoped, repeatable

Three ways to see it

1. 1

   An Anthropic Skill is a folder on disk. Inside it sits a SKILL.md file with a name, a description, and the instructions the AI should follow when the skill is invoked. Alongside SKILL.md can sit example files, reference data, helper scripts, and templates. When a user's request matches the skill's description, the AI loads the folder and proceeds with the captured workflow. The skill is bottled expertise. It does not depend on which user is talking to the AI; it does not drift between sessions. It is also versioned, reviewed, and approved like any other deliverable.

2. 2

   Why does this protect you? Three reasons. First, the skill is the only source of truth for that workflow. If five officers were each writing slightly different prompts to summarise tax returns, five different summary formats reached the boss. With a Skill, one format. Second, the skill carries its scope with it. A skill might say 'You may read files in /returns; you may not write outside /summaries; you may not call the network'. The AI obeys because the skill says so. Third, skills are reviewable. Your compliance officer can read SKILL.md, sign off, and the audit trail is the file itself.

3. 3

   Pakistani example: Easypaisa's customer-care team built a Skill called 'merchant-onboarding-checklist'. It contains the official 14-point checklist, the required document types, the rejection reasons, and three sample completed cases. When a new agent joins, instead of training her on the prompt, she invokes the Skill. Every onboarding ticket gets the same checklist applied in the same order. The compliance officer audits the Skill once a quarter. The agent does not write prompts; she writes ticket numbers. The error rate dropped 60% in the first month.

The why-tree

Try this with Claude

AI-edge prompt: 'I want to create an Anthropic Skill for [your workflow]. Generate the SKILL.md with frontmatter (name, description, when to invoke), three numbered steps, one input example in JSON, one output example in JSON, and a list of files the skill should never modify. Keep it under 250 words.' Then review it as if you were the compliance officer signing off.