SOPs -> Skills

What a skill is

You walk out of this lesson with your three most-repeated procedures running as commands - starting with /meeting-capture, a skill that takes a raw transcript and files a structured note in the right folder of your real brain repo.

A skill is a procedure file Claude loads on demand. It lives at .claude/skills/<name>/SKILL.md in your brain repo. You run it by typing /name, or Claude picks it up itself when the task matches its description. Last lesson's rule applies: CLAUDE.md holds facts, skills hold procedures.

If you have ever written an SOP that nobody follows, this is the fix. The SOP becomes the skill. Running the procedure and reading the procedure are now the same act, so it cannot drift into a stale doc.

This is also rung two of the ladder the whole track climbs: chat once, find or build a skill, put the skill on a schedule, and - only where judgment is needed - grow it into an agent. You climbed rung one when you asked the brain its first question. Module three reaches the top.

Two mechanics worth knowing. First, custom slash commands have been merged into skills - a .claude/commands/x.md file and a skill both create /x, but skills are the recommended form. Second, skills follow the open Agent Skills standard (agentskills.io), so the SOP library you build here is portable across tools. You own it; it is not locked to one vendor.

SKILL.md anatomy

A skill is a folder with a SKILL.md inside: YAML frontmatter on top, the procedure in markdown below. Here is a real one - the meeting-capture skill you will build in a minute.

---
name: meeting-capture
description: Turn a raw meeting transcript into a structured note
  (decisions, action items, risks) filed in the right folder.
  Use when the user pastes or points at a transcript.
argument-hint: "[path to transcript or paste it after the command]"
allowed-tools: Read, Write, Grep, Glob
---

Today's date: !`date +%F`

# Meeting capture

Input: a transcript, either pasted as $ARGUMENTS or at the file
path given.

1. Read the transcript. Identify: attendees, decisions made,
   action items (each with an owner), and open risks.
2. Resolve owners against company/org-chart.md. If a name is
   ambiguous, ask - never guess an owner.
3. Write the note using the template at
   templates/meeting-note.md.
4. File it: client meetings go to clients/<client>/meetings/,
   internal ones to meetings/. Filename: <date>_<topic>.md
   using today's date from above.
5. End by printing the decisions and action items as a short
   list for the user to review. Do not create tasks in any
   external system - that is the meeting-followup skill's job.

• description - the field that drives auto-invocation. Claude reads every skill's description each session and loads the body when a task matches. Write it like a when-to-use sentence (there is also a when_to_use field; the two together are truncated at 1,536 characters in the skill listing).
• disable-model-invocation: true - makes a skill manual-only. Claude can never trigger it on its own; only you can, by typing the slash command. Mandatory for anything with side effects: sending, deploying, billing.
• allowed-tools / disallowed-tools - pre-approve or block tools while the skill runs. The grant clears on your next message.
• model: and effort: - pin the model per skill. model: haiku for a mechanical formatter, model: opus for a judgment call; leave it unset and the skill runs on the session model (Opus 5.5 on every plan as of September 2026). Pinning respects your org's model allowlist.
• context: fork - run the skill in a subagent so its work does not crowd your session. Forked skills run in the background by default since July 2026; set background: false when you want to watch it.
• paths: - a glob so the skill auto-loads only when you work in matching folders.
• Dynamic injection: !\command\ inside the body runs a shell command before Claude sees the prompt and inlines the output - the date line above, or !\ls clients/\ to show current clients.
• $ARGUMENTS, $0, and named arguments make skills parameterized: /proposal acme.

Where a skill lives, and how to draft one

Four places, and the choice is a distribution decision, not a filing one.

• Project: .claude/skills/ in the brain repo. Checked in, reviewed by PR, shared with everyone who clones. This is where company procedures live.
• Personal: ~/.claude/skills/. Follows you across every project; nobody else sees it. Your own shortcuts and drafts before they are ready to share.
• Team plugin: a private marketplace repo that packages skills, hooks, and MCP servers so a new machine installs the whole set with one /plugin install. The right home once the library is stable.
• Synced from claude.ai: skills you enable on your claude.ai account appear under ~/.claude/skills/synced/ (the anthropic-skills: namespace is reserved for them). Turn it off with syncClaudeAiSkills: false if you want the repo to be the only source.

To draft one, do not start from a blank file. Anthropic's official skill-creator skill (in the anthropics/skills repo) interviews you for the procedure, scaffolds the folder with a description written to trigger correctly, and can run evals against real inputs to measure whether the skill fires when it should. Our brain installs it into every project for exactly this reason. Run it, then edit the result down to your actual SOP.

Build the meeting-capture skill

This is the first skill we have every ops team write, because the input is messy, the output is structured, and you will use it weekly.

1. Create the folder .claude/skills/meeting-capture/ in your brain repo and add the SKILL.md from above.
2. Create templates/meeting-note.md with your note format: a header block (date, attendees, client), then Decisions, Action items (owner + due), and Risks sections.
3. Find a real transcript from last week - export one from your meeting recorder, or use raw typed notes.
4. Run /meeting-capture with the transcript path. Read the output note carefully: are owners right, did it file to the correct folder, is the date format correct?
5. Fix the weakest spot by editing SKILL.md - usually the owner-resolution step or the filing rule - and run it again on a second transcript.
6. Commit the skill and the template. It is now part of the brain: anyone who clones the repo has /meeting-capture.

Notice what the skill deliberately does not do: it does not push tasks to your project tool or update the CRM. Keep the first version filesystem-only. You will chain it into the full meeting pipeline in module two, after connections and governance are in place.

Check the library first

Before you build the next nine, learn the other half of the habit: there are two ways to get a skill - borrow a ready one or build your own. Skills are portable SKILL.md files in an open standard, so large free libraries already exist. Start with Anthropic's own: the anthropics/skills repo holds the document skills (docx, pdf, pptx, xlsx), skill-creator, and the spec itself, and installs as a marketplace. Then the big index: SkillsMP (skillsmp.com) is a free independent catalog of over 3,000,000 collected SKILL.md files from public GitHub repos (as of late September 2026), browsable by category with a free REST API and MCP; Business, Documentation, and DevOps are the shelves an ops team uses most. agentskills.io maintains the standard and its own directory. Many vendors also ship official downloadable skills - Airtable and Notion both do - and when a vendor publishes one for a tool you use, start there.

# Anthropic's official marketplace is pre-added:
/plugin install <name>@claude-plugins-official

# Anthropic's skills repo (document-skills, example-skills):
/plugin marketplace add anthropics/skills

# The community marketplace: the install suffix is NOT the repo name
/plugin marketplace add anthropics/claude-plugins-community
/plugin install <name>@claude-community

# A skill found on GitHub: read it first, then copy
# the folder into your brain repo
cp -r path/to/invoice-reconciler .claude/skills/

The dividing line for borrow versus build: install the generic shapes - an invoice reconciler, a document generator, a meeting-note formatter, a candidate-screen rubric scaffold - and hand-build only what encodes YOUR company: your actual SOPs, your org-chart routing, your policies.

The ten-skill SOP library

Here is the core library we deploy for an ops company brain. This is the part you build yourself, because your SOPs are yours - the library has the generic shapes, but nobody else has your procedures. You will not build all ten this week - you will build three. But knowing the full map tells you where you are headed.

• new-client - onboarding checklist: create the folder, scaffold its CLAUDE.md, kickoff tasks, CRM record.
• meeting-capture - transcript to structured note, filed and synced (you just built v1).
• weekly-report - pull KPIs from source systems, assemble against a template, flag anomalies.
• proposal / sow - fill templates/sow.md from the client folder plus pricing.md, render to HTML or PDF.
• invoice-check - the Stripe or QuickBooks reconciliation procedure.
• brain-capture - end of project: write the case file, update the index.
• inbox-triage - label taxonomy, routing rules, and the draft, don't send policy.
• screen-candidates - rubric as markdown, evidence-quoting required, human-review output only.
• data-clean - dedup and normalization rules between CRM, spreadsheet, and email tool.
• onboard-employee - the day-1/week-1 checklist run as a skill.

To pick your three: take the procedures you personally explain most often. The ones where you have said "check the doc" and then watched someone do it wrong anyway. Highest repetition, clearest steps, lowest blast radius if v1 is imperfect.

Keeping docs and skills in sync

The honest answer: do not maintain two copies. The skill IS the SOP. If a human needs to read the procedure, they read the SKILL.md - it is plain markdown and reads fine. Your playbooks/ folder holds the judgment-heavy material (how we think about QBRs); the skill holds the executable steps.

Do this now

Sources and further reading

• Claude Code skills docs
• Anthropic's official skills repo (skill-creator, document skills, the spec)
• Anthropic plugin marketplaces (official, community)
• SkillsMP - open-source skills catalog
• Agent Skills open standard