Most people treat Claude Cowork as a smart file organizer.
Save this :)
Ask it to sort some files. Convert a spreadsheet. Maybe rename a folder.
That is the junior version of Cowork.
The senior version is building a plugin — a complete AI employee that knows your industry, follows your exact process, produces your exact output format, and runs your workflows autonomously while you do something else.
An AI employee that shows up every day, never calls in sick, never needs training twice, and gets better every single week.
A Cowork plugin is a structured bundle that contains everything Claude needs to perform a specific role: the skills, the commands, the reference materials, the rules, and the quality checks. Install it once and you have an employee.
This is the full course on building one.
What a Plugin Actually Is
A plugin is a folder. That is it. A folder with a specific structure that tells Cowork exactly what to do and how to do it.
Inside that folder:
plugin.json tells Cowork this folder is a plugin, what role it plays, and when to activate.
SKILL.md is the brain — the step-by-step process your AI employee follows for every task. This is the single most important file. Get this right and everything else works.
Commands are slash-command triggers. Type /prefix:command and the entire workflow fires.
References are the templates, benchmarks, industry data, and examples your employee needs to do the job properly.
Global instructions are the standing orders — personality, communication style, quality standards, and default assumptions.
Folder instructions are project-specific context — what is in this folder, what the current priorities are, and how to handle the specific data here.
Phase 1: Research the Role (30 Minutes)
Before you write a single file, you need to know exactly what your AI employee does.
Open Claude and use this prompt:
"Research the complete professional workflow for a [ROLE] in [INDUSTRY]. Include: the step-by-step process they follow, tools and data sources they use, key metrics and decision criteria, common output formats, and expert-level edge cases and pitfalls. Be comprehensive — I am building an automated workflow from this."
Read the output carefully. This is the raw material your skill file will be built from.
Now interview yourself. What does YOUR version of this process look like? What shortcuts do you take? What quality checks do you always run? What mistakes do you always watch for? What does "good" look like versus "bad"?
The best AI employees are not built from generic best practices. They are built from YOUR specific expertise.
Phase 2: Write the Skill File (60 Minutes)
The SKILL.md is your AI employee's brain. Everything it knows about how to do its job lives here.
Here is the structure:
name: [skill-name] description: [When should this activate? Be aggressive with trigger phrases. "Use this skill when the user says: [list 5-7 phrases]. Do NOT use for: [list things that sound similar but are different]."]
# Overview [One paragraph: what this skill does and what it produces]
Process
[Numbered steps. Every step is specific, testable, and unambiguous. Not "analyze the data" but "compare current period to previous period and calculate the percentage change for each metric."]
1. [Step with specific instruction]
2. [Step with specific instruction]
3. [Step with specific instruction]
...
Output Format
[Exactly what the deliverable looks like]
- Title format
- Section headers in order
- Length constraints
- Formatting requirements
Rules
[Your non-negotiable quality standards]
- [Rule 1]
- [Rule 2]
- [Rule 3]
Edge Cases
[What to do when things are not straightforward]
- If [situation]: [specific action]
- If [situation]: [specific action]
Quality Checklist
[Run this before delivering any output]
- [ ] [Check 1]
- [ ] [Check 2]
- [ ] [Check 3]
The description field in the YAML frontmatter is the most critical part. If it is too vague, the skill never activates. If it is too broad, it hijacks unrelated conversations. List 5-7 specific trigger phrases AND explicit negative boundaries.
Phase 3: Build the Supporting Files (30 Minutes)
The plugin.json:
{ "name": "my-ai-employee", "description": "A [ROLE] that [WHAT IT DOES] for [WHO]", "version": "1.0" }
The slash command:
Create a markdown file in /commands/ that triggers your workflow:
# /employee:run
Execute the [primary-task] skill on the data in the current folder.
Steps:
1. Read all relevant files in the working directory
2. Execute the skill following every step in SKILL.md
3. Run the quality checklist before delivering
4. Save the output as [format] to the current folder
5. Provide a brief summary of what was produced
Global instructions:
You are a [ROLE] with [YEARS] of experience in [INDUSTRY].
Standing Orders:
- Lead with the recommendation, explain after
- Always use specific numbers, never vague descriptions
- If data is missing or ambiguous, flag it — never guess
- Default output format: [YOUR PREFERENCE]
- Communication style: [DIRECT/CONVERSATIONAL/FORMAL]
- When in doubt, ask rather than assume
Reference files:
Add any templates, benchmark data, industry standards, or examples your employee needs. The more specific your reference materials, the more expert-level the output.
Phase 4: Install, Test, and Refine
Install the plugin folder to your Claude Cowork environment. Use this prompt in Cowork:
"I have a plugin folder at [PATH]. Verify the structure is valid — check plugin.json, SKILL.md frontmatter, and command files. Install it and run a quick test with the simplest slash command."
Now test it on real work. Not sample data. Real data from your actual workflow.
Run it 5 times with different inputs. After each run, evaluate:
- Did it follow every step in the SKILL.md?
- Did it follow the rules?
- Did the output match the format specification?
- Would you use this output as-is or does it need significant editing?
Every time something misses the mark, update the SKILL.md. Add a rule. Tighten a step. Add an example showing what good versus bad looks like.
This refinement loop is what turns a mediocre AI employee into an exceptional one. By run 10, the output quality will be dramatically higher than run 1.
Phase 5: Scale Your AI Employee
Once your primary skill is running reliably, expand the employee's capabilities.
Add a second skill. Your research analyst can now also do competitive monitoring. Your content strategist can now also repurpose content. Each new skill is a new SKILL.md in the skills folder.
Add automated workflows. Chain multiple skills into multi-step processes triggered by a single slash command. Research → analysis → report → distribution. One command, four skills, zero manual steps.
Add scheduled tasks. Your AI employee runs the weekly report every Friday at 4pm. Processes the daily inbox every morning at 7am. Scans competitors every Monday. True autonomous operation.
Add sub-agents. For complex workflows, your AI employee can spin up multiple sub-agents that work in parallel. Five files processed simultaneously instead of sequentially. The speed improvement compounds with every additional sub-agent.
The Performance Review System
Here is what separates people who build a decent AI employee from people who build a great one.
Every week, review the outputs. Note what worked perfectly, what needed corrections, and what you had to redo manually. Then update the SKILL.md.
This takes 15 minutes per week. The compound effect over two months is massive.
By week 1, your employee is functional. By week 4, it is good. By week 8, it is producing work that would take a human junior hire months of training to match.
The tool does not get better on its own. Your instructions get better. And your instructions are entirely within your control.
Where to Start Right Now
Pick the task you spend the most time on every week. The one you dread. The one that follows the same process every time.
Spend 2 hours building the plugin following this course.
By tonight you will have an AI employee that handles your most time-consuming task. By next month you will wonder how you ever did it manually.
Most people will keep doing everything themselves because building an AI employee "sounds complicated."
The ones who spend 2 hours today will have an autonomous AI employee working for them every day for the rest of the year.
*Follow me *@eng_khairallah1 *for more automation architectures, workflow designs, and business AI playbooks.*
hope this was useful for you, Khairallah ❤️
