How to Build a Claude Cowork Plugin

Markdown files + JSON that teach AI your specific methodology. No compiled code, no infrastructure, no deployment pipeline.

11 open-source plugins from Anthropic (Jan 30, 2026) showed a few KB of structured thinking can replace million-dollar workflows.

The hardest part: deciding what "good" looks like and writing it with enough precision for an AI to follow. Forces intuition → structure.

Craft skills are learnable in months. Codified judgment — structured methodologies, quality criteria, repeatable systems — compounds over a career.

Name, version, description (critical for activation accuracy), author, license. The description determines WHEN Claude uses your plugin.

Each skill is a directory inside skills/ containing a single SKILL.md file. YAML frontmatter (name + trigger description) → Markdown body (the expertise). Keep under 500 lines; offload detail to references/.

The YAML description is the most important line in your plugin. It determines activation accuracy. Test: "Would I want this to trigger?"

Without the --- YAML block containing name and description at the top of each SKILL.md, skills register with empty descriptions and will never auto-trigger. The plugin system does NOT read your markdown headings — it only reads frontmatter.

One .md file per command in commands/. Contains: usage syntax, parameters, step sequence, expected output. Filename = command name. (Note: commands stay as flat .md files, unlike skills which use directory/SKILL.md format.)

Use namespace prefix: /xd:command. Prevents collision with other plugins and makes discovery intuitive.

Single .mcp.json at plugin root. Web search uses "type": "builtin". All other connectors use "type": "http" with a real endpoint URL (e.g. "url": "https://mcp.slack.com/sse").

Never use "type": "mcp" with placeholder notes — the plugin loader will fail trying to initialise servers without valid endpoints. Only include connectors you can actually configure with real URLs.

Brand colours (hex), typography, dual-branding (your 70% / client 30%), slide sequence template, speaker notes requirements, quality checklist.

Identify 5–7 core workflows. For each: what separates senior from junior output? What knowledge is trapped in heads? Where does quality vary by person?

Map each expertise area to a skill directory containing SKILL.md. Write YAML trigger descriptions. Encode judgment, not just process. 80–200 lines each.

Map to deliverables. Each command = one complete workflow → one clear output. Use namespace prefix. Composability over monoliths.

Industry playbooks (start 2–3), methodology docs, PPT generation guide, output templates. Be specific: "12–15 stakeholder interviews" not "we do research."

Start with web search. Add Drive/Notion in phase 2. Slack/CRM only when workflows demand real-time data.

Run past projects where you know "good." Compare AI vs actual team output. Gaps reveal where skills need more specificity.

Every slide: key message (1 sentence), 3–5 talking points, transition cue, likely objections. Deck without notes = half the deliverable.

Auto-extract from client website/guidelines: primary + accent colours (hex), headline/body typefaces, visual style, photography direction.

Zip plugin folder + README.md (setup, commands, customisation) + settings.local.json template + connector documentation.

Not "what can I automate?" but "what expertise in my practice is too valuable to exist only as tribal knowledge?"