If your agent’s instructions keep getting longer, you’ve already hit the problem skills solve.
It starts with a paragraph. Then you add how to handle email. Then the CRM rules. Then the report format, the follow-up cadence, the website checklist. Six weeks later the agent is reading a novel before every task and getting worse at all of them.
People don’t work that way. A good employee doesn’t recite the whole operations manual before answering the phone. They pull the right procedure when the situation calls for it.
Claude skills let your agent do the same.
What Claude skills are
As of October 2026, Agent Skills are folders of instructions an agent loads when a job calls for them. In Claude Code, skills are reusable instructions loaded on demand. Managed Agents lets you attach skills when you define an agent.
That word “folder” matters. A skill isn’t just a paragraph. It can hold the instructions plus the things the job needs: a template, a checklist, a reference document, an example of great output.
For the official details, read Anthropic’s Agent Skills overview. For how skills fit next to the other building blocks, see the Claude agent overview.
Skills are SOPs with a trigger
Here’s the reframe that made skills click for me: a skill is an SOP your agent knows when to use.
I already believed written process is the fuel for AI. I wrote about it on SOPs for AI. A skill takes that same SOP and adds one thing: a short description that tells the agent when this process applies. The agent sees the work, matches it to the description, and loads the full instructions.
That’s how my AI chief of staff on Claude Code handles so many different jobs without drowning. The daily briefing is a skill. Lead follow-up is a skill. Website and SEO work is a skill. My second brain, the knowledge base in Obsidian with Google Drive behind it, holds the source processes; the skills are the packaged, agent-ready versions.
How I turn an SOP into a skill
- Record the job once. Do the task for real and talk through it out loud. Transcribe it.
- Let AI draft the SOP from the transcript, then fix what it got wrong.
- Write the “when to use this” line. One or two sentences, specific enough that the agent won’t load it for the wrong job.
- Add what the job needs — templates, checklists, examples of good output — inside the skill folder.
- Run it while you watch. Give the agent real work and read every step.
- Fix the skill, not the output. When it gets something wrong, change the instructions so it won’t happen again.
The anatomy of a good skill
| Part | What it answers |
|---|---|
| Name | What is this job called? |
| When to use it | Which situations should load this skill, and which shouldn’t? |
| Trigger and inputs | What starts it, and what information does it need? |
| Steps | What happens, in order? |
| Decision rules | How should it handle the common forks? |
| Definition of done | What does finished, correct work look like? |
| Escalation | When does it stop and ask a person? |
| Never | What must it never do in this job? |
| Example | One example of great output |
The “when to use it” line does more work than people expect. Too vague and the agent loads the skill for the wrong job. Too narrow and it never loads at all.
Start with three skills
Don’t write twenty. Write three, make them reliable, then add more.
1. The daily briefing
What happened yesterday, what’s on the calendar today, what needs a decision, what’s overdue. It’s low risk, it runs every day, and it teaches you fast how specific your instructions need to be. The AI chief of staff page shows what a strong briefing covers.
2. Inbox triage
Sort, summarize and flag, read-only at first. The agent drafts replies for your review; it doesn’t send them. Once you trust the sorting, decide what, if anything, it may send on its own.
3. One recurring report or cleanup
A weekly scoreboard pull, CRM hygiene, or a follow-up list from stale leads. Pick the job that eats an hour of your week and has a clear definition of done.
Keep skills safe and current
- One owner per skill. Someone reviews its output and updates it when the process changes.
- Least access. A skill shouldn’t need tools its job doesn’t use.
- Human approval for anything irreversible — sending to customers, money, deleting records.
- Review on a rhythm. When the agent drifts, the skill is usually out of date.
- Outreach is regulated. If a skill drafts calls or texts, consent, the FCC’s February 2024 TCPA ruling on AI voices, A2P 10DLC and Do Not Call apply. Get counsel.
Your next step
Pick the one SOP you’d trust most to someone new, and rewrite it in the skill structure above, including the “when to use it” line. That’s your first skill. If you want help turning your processes into a working AI chief of staff, see how Chief works. And if you want the full month-long plan, follow the Claude agent training path.
Questions people ask
What are Claude Agent Skills?
As of October 2026, Agent Skills are folders of instructions an agent loads when a job calls for them. Instead of stuffing every process into one giant set of instructions, each repeatable job gets its own skill, and the agent pulls the right one when the work shows up.
How do I write a Claude skill?
Start from a written SOP. Give the skill a clear name and a short description of when to use it, then the trigger, inputs, steps, decision rules, definition of done, escalation rules and one good example. Test it on real work while you watch, and fix it every time it goes wrong.
Are Claude skills the same as prompts?
No. A prompt is something you type for one task. A skill is a saved, reusable set of instructions — often with templates or reference files — that an agent loads by itself whenever the job matches. Prompts are conversations; skills are processes.