Create a Claude skill for a repeated task in 3 hours

What a skill actually is

A skill is a directory containing a SKILL.md file that opens with YAML frontmatter holding a name and a description. Anthropic's engineering write-up says the agent preloads the name and description of every installed skill into its system prompt at startup, then reads the full file only when it judges the skill relevant to the task.

That is the whole mechanism. The description is a router, not documentation. A skill nobody triggers costs almost nothing, and a skill with a vague description never fires without telling you. Anthropic's own skills repository ships a template skill to copy.

How to create a Claude skill from a repeated task

You write one markdown file and put it where Claude looks. No code is required, though a skill can ship scripts if you want them. The times below are my estimates, not measured results.

1. Do the job once and watch yourself (about 40 minutes). Anthropic's authoring guidance is to work the task with ordinary prompting and notice the context you keep supplying: the field names, the house format, the always-exclude rules. That repeated context is the skill.

2. Write SKILL.md (about 45 minutes).

---
name: quote-intake
description: Collects and checks the fields needed for a quote, then formats it in the house layout. Use when a new enquiry arrives or someone asks for a quote.
---

## When to use this
## Required fields
## Output format
## Never do this

The name is capped at 64 characters, lowercase letters, numbers and hyphens only, and cannot contain the reserved words anthropic or claude. The description is capped at 1,024 characters and should be third person, covering what the skill does and when to use it. Keep the body under 500 lines and split the rest into separate files the skill references.

3. Install it (about 20 minutes). In Claude Code the folder goes in ~/.claude/skills/ for you or .claude/skills/ in the project for the team, with no upload step. On claude.ai you upload a zip, on Pro, Max, Ultra, Team or Enterprise with code execution enabled.

4. Test on three real jobs, two of them awkward (about 45 minutes). The failure to look for is the skill not firing. Fix that in the description, not the body.

5. Hand it to one other person (about 30 minutes). If their output differs from yours, your instructions describe a decision you never actually made. The same loop works for a repeat client fix and for a LinkedIn writing workflow.

Total: about three hours.

Which repeated tasks are worth packaging

Package a task when the right output has one shape. Leave it alone when the shape is the judgement.

Run the impact on your own numbers. A few people rebuilding the same prompt every week adds up to real hours across a year. I have not measured a before and after on a team, so the arithmetic is the claim, not the result.

Limitations

• A skill is instructions, not a gate. It cannot stop someone ignoring it and it does not record what happened, so pair it with a way to audit what your agent actually did.
• Discovery is a judgement call on the description, so two overlapping skills can both look right to the model.
• In managed agents a mounted repository's root .claude/skills is scanned at session start with no review step, so anyone who can commit can change what your agent is told. Anthropic's advice is to install skills only from trusted sources and audit them first.
• I have not tested the claude.ai zip upload on a Team plan myself, and the three hour budget is my estimate.
• My guess: most first skills are too long. The ones that hold up read like a checklist.

My take

The value is not the file, it is the hour you spend deciding what the correct output is. Most teams have never written that down, which is why four people produce four versions. Pick one task, write the file, and let the second person test whether your instructions survive contact with someone who was not in your head.