Skills — Build Repeatable AI Workflows
Turn any successful conversation into a reusable, improvable standard operating procedure — without writing a line of code.
What you'll learn
What Skills Actually Are
Skills are plain markdown text files. That is the entire technical definition. They contain reusable instructions — a role definition, context, output format, constraints, and examples — and Co-Work reads them when you trigger the skill. Nothing more complex than that.
Felix Rieseberg at Anthropic confirmed this directly: "Skills are just markdown files." Multiple community sources — Paul, Jack Roberts, Brooke Wright — say the same thing. The official docs use the term "instruction sets," which refers to the same thing. There is no code, no configuration syntax, no templates to learn. You are writing instructions in plain text.
Where they live: in your Co-Work Skills library, accessible via "/" command or natural language trigger. When you activate a skill, Co-Work loads the file and uses it as pre-loaded context for the conversation. It arrives knowing its role, your format preferences, the constraints you care about, and examples of the output you expect.
What skills are NOT: scripts, code, rigid templates that must be followed word for word. They are instructions that Claude interprets. Co-Work adapts them to the specific task at hand — which is why they are far more flexible and durable than a hardcoded automation in a traditional tool.
Why Skills Change Everything
Think about the best task Co-Work has done for you so far. The morning briefing from Module 5, or the file organization run that worked exactly right. Now think about running that same task again next week — from scratch. New conversation, no context, no memory of what worked. You re-enter all the context, re-explain the format you want, re-specify the constraints. Every single time.
That is Co-Work without skills. With skills, Co-Work arrives pre-briefed. The context is already loaded. The format is already specified. The constraints are already defined. You trigger the skill, it runs, it produces the output you established in your best previous run. Same quality. Consistently.
Paul's prompt after any successful Co-Work session: "This was fantastic. Create a skill now." That immediate conversion — from successful conversation to codified skill — is the habit that makes a Co-Work setup compound in value over time.
Two Skill Creation Pathways
Pathway 1: From Conversation
This is the faster of the two pathways and the one you will use most often. Complete a task. Get output you are genuinely happy with. Then say: "This was fantastic. Create a skill from this conversation."
Co-Work extracts the task description, context, format preferences, and constraints from what just happened and drafts a skill markdown file. It will show you what it produced. Review it carefully:
- Does the role definition accurately describe what you needed?
- Are the output format instructions specific enough to reproduce what you just got?
- Are the examples drawn from this conversation's best output?
- Are the trigger keywords the words you would actually type to activate this skill?
Edit until yes on all four. Then save to your Skills library. Store the underlying markdown file in ~/CoWork/Skills/ so you have a backup outside Co-Work's interface.
Pathway 2: Skill Creator (Built-In)
The Skill Creator is the built-in interview tool: Settings → Skills → Create New → Skill Creator. It asks you a series of questions about the task you want to systematize — what it does, when you use it, what good output looks like, what constraints apply — and generates the skill markdown from your answers.
Brooke Wright's key insight: feed the Skill Creator examples of existing output you have approved. Paste in the best result from a previous Co-Work task and say "this is the output quality I am targeting." The generated skill will encode the format and style of that specific output rather than a generic approximation of it.
Pathway 2 is best for skills you want to build before you have run the task manually enough times. It trades some accuracy (because there is no existing good output to learn from) for speed of creation.
The Never Encode a Bad Skill Principle
Never Encode a Bad Skill is a practitioner principle articulated by Jack Roberts. It is not official Anthropic guidance. The underlying logic — that you should not codify a workflow until it consistently produces output you would be satisfied with — is sound, well-supported by community evidence, and directly applicable to every skill creation decision you will make.
The rule: run any task manually at least three times, refining the prompt each time, before converting it to a skill. The benchmark question after each run: "Would I be happy if Co-Work ran this skill without my review and sent the output directly to someone?"
If the honest answer is no — keep refining. If yes — encode.
The consequence of encoding too early is worse than not encoding at all: a bad skill produces consistently bad output. Automatically. At scale. Without you noticing until someone points out that the last six morning briefings missed the same important email thread.
A skill encodes your workflow exactly as you define it. Encode a poor workflow and you get consistent poor output at scale. The refinement step in the Automation Pipeline — do manually → refine → create skill → schedule — is not optional. It is the step that makes the difference between an automation that saves you time and one that creates new problems at 6 AM.
Skill Metadata: Teaching Co-Work When to Activate
Three metadata fields determine whether your skills work as intended — or whether Co-Work picks the wrong one, ignores the right one, or requires you to remember the exact "/" command every time.
- Name: Short, verb-led, unambiguous. "Draft Client Proposal," "Triage Inbox," "Morning Briefing," "Organize Downloads." If the name describes what the skill does in one action, it is good.
- Description: One sentence: when should Co-Work use this skill? "Use this when asked to prepare any client-facing proposal or estimate." If the description answers the "when" question, it is good.
- Trigger keywords: Words in a prompt that should activate this skill automatically — without requiring a "/" command. "morning brief," "daily briefing," "morning report" should all activate the Morning Briefing skill. List synonyms. List natural-language variations. Brock, Paul, and Brooke all emphasize that natural-language activation is more reliable than requiring exact "/" commands in real use.
Conflict prevention: if you have two skills with overlapping trigger keywords, Co-Work will pick the wrong one some of the time. Disable skills you are not actively using. Audit trigger keywords across your full skills library every few weeks.
Skills Get Better Over Time
The most underused practice in skill management is the reflection prompt. After every skill run, say:
"Reflect on our conversation and update the skill so we get to the correct answer faster next time."
Brooke Wright uses this after every skill run as a habit. Co-Work reviews what happened in the session, identifies gaps between the skill instructions and what actually produced the best output, and rewrites the relevant sections. Context gets refined. Examples get updated. Edge case handling gets added.
This is why skills in a mature Co-Work setup produce better output than skills in a new setup: they have been refined by dozens of real runs, not just one carefully considered first draft.
The Complete Automation Pipeline
Module 5 gave you three stages: do manually, refine, schedule. This module adds the fourth, which is the one that turns automation from a one-time setup into a compounding asset:
Do manually → Refine → Create Skill → Schedule.
The skill is the quality gate. It is the codified, tested, named version of a workflow — the version you are willing to put your name on and have run automatically. Module 8 shows how to attach a skill to a scheduled task with full control over model selection and output routing. Module 7 shows how to bundle skills into plugins for department-level distribution.
This is the Tier A capstone exercise. Create three skills — two from the pathways above, one improved through the reflection prompt. By the end you should have three active skills in your Skills library and a clear understanding of how each one behaves.
- Skill from conversation (Pathway 1): Take your morning briefing from Module 5. After running it once more and confirming it is still producing good output, say: "This was fantastic. Create a skill now." Review the generated markdown. Confirm trigger keywords include "morning brief," "daily briefing," and "morning report." Save to
~/CoWork/Skills/morning-briefing.md. - Skill from scratch (Pathway 2): Use Skill Creator to build a skill for the automation you identified in Module 1 — the task you most want to automate. Go through the interview. Provide an example of the output you want. Confirm trigger keywords. Run it three times manually before finalizing.
- Skill self-improvement: Run Skill 1 (morning briefing) one more time. Then: "Reflect on this run and update the skill so it produces better output next time." Review the diff — what did Co-Work change? Save the updated version.
Success criteria: Three skills visible in your Skills library with distinct names, descriptions, and trigger keywords. Each skill produces the correct output when triggered by natural language (not just "/"). The morning briefing skill has been updated at least once through the reflection prompt.
Tier A Complete
You have reached the end of Tier A. You now have a Co-Work setup that actually works: global instructions configured, folder architecture built, connectors live, quick wins running on a schedule, and three skills codifying your most useful workflows. Alex's transformation is complete.
If you are stopping here: your Co-Work setup will continue to improve on its own. Keep using the reflection prompt after every good session. Add a skill whenever Co-Work does something you would want to repeat. Review and update your Business Brain quarterly.
If you are continuing to Tier B: Module 7 introduces Plugins — the bundling layer that turns a collection of skills into a department-level automation package. That is where Morgan's journey begins.