AI Skills
AI Skills: Turn your repeatable workflows into portable skills your whole team can reuse.
You have figured out how to get Al to produce real work. A meeting brief that surfaces what matters. A report that pulls the right data. A task breakdown that catches edgecases you would have missed. The output is genuinely useful.
But next time you need the same thing, you start over. The approach that worked on Tuesday does not quite come together on Friday. The context is different, the setup takes just as long, and the result is not as good. A colleague asks how you did it, and the honest answer is: “I just kept rephrasing until it clicked.”
01
The gap
![](data:image/svg+xml,<svg display="block" id="4166636898" role="presentation" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M 18 11.55 L 18 9.09 C 18 8.231 18 7.803 17.829 7.416 C 17.658 7.031 17.333 6.726 16.682 6.119 L 11.354 1.147 C 10.792 0.623 10.512 0.361 10.163 0.206 C 10.091 0.173 10.017 0.145 9.942 0.12 C 9.578 0 9.181 0 8.388 0 C 4.737 0 2.912 0 1.675 0.93 C 1.425 1.118 1.198 1.33 0.997 1.563 C 0 2.719 0 4.422 0 7.829 L 0 11.55 M 10.125 0.525 L 10.125 1.05 C 10.125 4.019 10.125 5.505 11.114 6.427 C 12.102 7.35 13.693 7.35 16.875 7.35 L 17.437 7.35 M 17.719 14.7 L 14.906 14.7 C 14.285 14.7 13.781 15.17 13.781 15.75 L 13.781 17.85 M 13.781 17.85 L 13.781 21 M 13.781 17.85 L 17.156 17.85 M 0.281 21 L 0.281 18.375 M 0.281 18.375 L 0.281 14.7 L 2.25 14.7 C 3.337 14.7 4.219 15.523 4.219 16.537 C 4.219 17.552 3.337 18.375 2.25 18.375 Z M 7.031 14.7 L 8.719 14.7 C 9.961 14.7 10.969 15.64 10.969 16.8 L 10.969 18.9 C 10.969 20.06 9.961 21 8.719 21 L 7.031 21 Z" fill="transparent" height="21.00000027156895px" id="aYpKLXDcH" stroke-dasharray="" stroke-linecap="round" stroke-linejoin="round" stroke-width="1.61" stroke="rgb(0, 0, 0)" transform="translate(3.085 1)" width="18px"/></svg>)
The results are good because you know what good looks like. You bring the right context, ask the right follow-up questions, and shape the output until it matches your standard. But all of that lives in your head, applied fresh each time. The AI is not getting better at this. You are doing the hard work every session: steering it step by step, pulling in the next piece of context, guiding it back when it drifts. The expertise is yours. The effort of directing it is also yours, every time.
Everyone has the same models. The difference is the workflows built around them. The people getting consistent, shareable results stopped treating AI as something they prompt on the fly and started encoding their expertise into skills: structured workflows the AI executes the same way every time, with room for the judgment calls that make each situation different. The skill pulls in the next step on its own. You focus on the decisions that matter instead of the setup. This guide shows what a skill is, when to build one, and how to go from a pattern you already do to a workflow your whole team can run.
02
What a skill is
skill /skil/
A structured set of instructions that defines a repeatable workflow toward a specific goal.
Think of it like a playbook combined with a recipe. The playbook gives you the goal and the strategy. The recipe gives you the steps and the ingredients. Some skills produce a finished output. Others guide you through a process, surfacing the right questions and decisions in the right order. Either way, the AI follows the workflow and exercises judgment within it.
Three parts
A clear goal. What the skill produces. A meeting brief. A task breakdown. A weekly report. A reviewed draft. The goal defines what "done" looks like.
A structured workflow. The steps to get there. Not "prepare for the meeting" but "pull today's calendar, check for agenda items, surface open action items, flag what needs a decision." The structure is what makes it repeatable.
Supporting context. Your review criteria, templates, and domain knowledge that the AI could not produce quality output without. This is what separates a skill from a generic prompt.
GOAL
What it produces
A meeting brief. A weekly report. A reviewed draft.
WORKFLOW
The steps to get there
Pull calendar. Check agenda items. Surface open action items.
CONTEXT
Tuned to your needs
Review criteria. Templates. Domain knowledge.
Skills and code
A skill is not code, and building one does not require programming. The skill itself is a set of instructions and references. It is not a rigid script that tells the AI exactly what to write, or a prompt template with blanks to fill in. A good skill gives enough structure that the result is consistent, with enough room that the AI can adapt to the specific situation.
That said, skills can invoke scripts, call into your other software through APIs, MCP, or a CLI, and connect to the tools your workflow depends on. Scripts are valuable for the parts of a workflow that need precision without AI involvement: pulling data from a system, formatting output in an exact structure, or running a calculation. They add predictability, save cost, and keep the AI focused on the steps where judgment actually matters.
The AI tool is rented.
The skills are yours. Skills follow an open standard adopted by 26+ AI platforms including Claude, ChatGPT, GitHub Copilot, Google Gemini, and Cursor. A skill you build today works across tools. Your expertise, your workflows, and your domain knowledge stay encoded in files you own and control.
Portable by design
Skills follow an open standard adopted by 26+ AI platforms including Claude, ChatGPT, GitHub Copilot, Google Gemini, and Cursor. A skill you build today works across tools. Your expertise, your workflows, your domain knowledge, all encoded in files you own and control. Switch the AI tool without losing the investment.
03
When to build one
Not everything should be a skill. The question is whether the investment pays back.
Three signals
You do it repeatedly. Not twice. Regularly. If you have done something three or more times and expect to keep doing it, there is a pattern worth capturing.
The setup matters. The quality of the result depends on getting the approach right, not just asking a question.
Consistency matters. You want the same workflow every time, not a different approach depending on the day. Anything where drift between runs would be a problem: brand voice, reporting formats, review checklists, onboarding steps.
When not to build one
One-off tasks do not need skills. If you are only going to do it once, just prompt well. Skills also go wrong when they are too prescriptive or too numerous. Over-specifying instructions actually makes the AI's output worse. Curate a focused set that each earn their place through regular use. A dozen well-maintained skills will outperform fifty that nobody updates.
A practical test
Before you build a skill, answer two questions:
Can I describe the steps I follow every time? If you cannot walk someone through the process, the workflow is still in your head.
Can I tell whether the result is good or bad? If you cannot evaluate the output, you cannot improve the skill. Without judgment, you just automate inconsistency.
Yes to both? You have a skill candidate. The appendix has a prompt to get started.
07
How we can help
AI results do not have to be accidental. The best results come from encoding your expertise into structured workflows that the AI executes consistently. Start personal. Share with your team. Build a library that improves through use.
Building your first skill is something you can do this week. Building a skill system that works across a team, stays consistent, and improves over time is where the investment compounds, and where it helps to work with a partner who has done it before.
04
A skill in action
Consider something you do every week: preparing for meetings. You check the calendar, look for notes from the last conversation, remember what was discussed, pull up action items, and format it all into something you can scan before walking in. Some days it takes ten minutes. Some days you skip it because you are running late. The quality varies every time.
Prompting from scratch vs. running a skill
Prompting from scratch
Check the calendar manually
Search for notes from last meeting
Remember what was discussed
Look up open action items
Prompt AI with scattered context
Steer it through multiple rounds
Running a skill
Calendar pulled automatically
Open action items surfaced
Agenda items checked
Output formatted consistently
You review and adjust
Reformatting is already handled
When this becomes a skill, the steps are defined, the context is pulled in automatically, and the output format is consistent. You focus on judgment calls: is this the right issue to raise? Does this agenda need to change? Should I bring up the delayed timeline?
Skills augment, they do not automate
The skill did not replace the person's judgment. It replaced the setup, the steering, and the structure. That is the difference between a skill and automation. Automation removes the person. A skill removes the busywork so the person can focus on the decisions that matter.
Some skills produce a finished output you review. Others guide you through a process: an end-of-day recap that walks you through what happened, what to reschedule, and what is ahead tomorrow. The AI presents and you decide. The skill keeps you on track either way.
05
From personal to team
Skills naturally move through three scopes, and each one multiplies the value.
The sharing ladder
Personal
Your Workflow
Your preferences, your shortcuts. A skill that makes your day better is reason enough. A
Team
Proven approach, shared
New team members get your approach on day one instead of developing their own.
Company
Active standard
Not a document people forget to check. A workflow that runs the same way every time.
You are building something, not buying something. A skill starts as a set of instructions in a file. That is all it takes to get running. The best skills grow over time: scripted elements for operations that need precision, connections to tools the workflow depends on, reference material refined from real use. Off-the-shelf software gives everyone the same features. A skill captures the way your team does something well and can grow in sophistication as the workflow demands it.
The learning loop
A skill improves through use, but only if you decide to invest in it. Pay attention to what the skill gets right and wrong. Keep a learnings file alongside it: “the model tends to skip the context step when the source material is short.” “The output format works for weekly reports but not monthly.” Periodically fold those lessons back into the skill itself.
After six months, the skill knows things you forgot you learned. This is a practice you adopt, not a feature that happens automatically. Version 1 is never the final version. The starter prompt in the appendix builds a verification step into every skill from the start, giving you a structure to notice what works and what needs adjustment.
Skills your team could build
Task breakdown and triage. A skill that structures thinking before starting work: scope, edge cases, approach, estimate. It can also triage incoming requests, sorting them by urgency, effort, and who should own them. It encodes how your best people plan and prioritize, shareable with the whole team.
Weekly status report. A skill that pulls project data, open items, and recent activity into a consistent format for leadership. Same structure every week regardless of who writes it. The report always covers what shipped, what is at risk, and what needs a decision.
Proposal drafting. A human-in-the-loop skill, one that walks you through each decision rather than producing a finished output. It guides you through assembling a proposal: gathering requirements, scoping the work, applying your pricing approach, and formatting to your template. You make the judgment calls on scope and pricing. The skill keeps the structure consistent and the voice on brand.
Client onboarding. A skill that guides a project manager through each step when a new engagement begins: contracts confirmed, access provisioned, kickoff agenda built, stakeholders introduced. Nothing falls through the cracks, and every client gets the same thorough start.
Brand and voice review. A skill that checks whether content matches your standards. It encodes your quality bar, not a generic grammar check. Every piece of outgoing communication runs through the same review, whether the author is a senior leader or a new hire.
From skills to a workspace.
When skills, context, and connections come together in one place, you have an AI Workspace. See The AI Workspace guide for how to build one.
06
Getting started
Start from something you already do well. Do not invent a skill from scratch. Pick a workflow you have already refined through repetition. The one where you know exactly what “good” looks like. Write down the steps. Write down the context you always gather. Write down how you judge whether the result is right.
Run it three times. The first version is a hypothesis. Run it on real work. Where did it skip a step? Where did it add something you did not want? Each run gives you signal.
Refine from what you learn. After three runs, you know what the skill needs: a step reordered, a reference file added, a constraint loosened. The best skills are not written perfectly the first time. They improve through use.
01
Notice
Spot a pattern you keep repeating
02
Write
Define the goal, steps, and context
03
Run
Use it on real work three times
04
Refine
Fix what broke and tighten
Repeat. The appendix includes a starter prompt you can paste into any AI tool to build your first skill. It walks you through the Write step in a guided conversation. The Run and Refine loop is where the skill earns its place.
A
Appendix: Build your first skill
This prompt uses the CRAFT pattern from our Collaborating with AI guide. Paste it into any AI tool (Claude, ChatGPT, or similar) and follow the conversation. It will walk you through defining your skill and produce a file you can save and reuse.
I want to build a skill for [describe the workflow].
Ask me one question at a time to work through each of these areas. Do not batch
questions. Wait for my answer before moving to the next one.
**Goal:** What does this skill produce? What is the specific output?
**Workflow:** Walk me through the steps I already follow when I do this manually.
Ask me one question at a time until we have the full sequence. Focus on the process, not the implementation details.
**Context:** What reference material, templates, checklists, or domain knowledge does this workflow depend on that the model would not have on its own?
**Degrees of freedom:** For each step, identify whether it needs high freedom (judgment call), medium freedom (preferred pattern), or low freedom (exact sequence matters). Default to high unless there is a reason to constrain.
**Verification:** How will I know if the output is good? What are the 3-5 things I check every time? Build these into the skill as a review step.
Once we have worked through all of this, generate a SKILL.md file with:
- The workflow as structured steps
- References to any supporting files we identified
- A verification/review step at the end
Keep the instructions focused on what the model does not already know or would not already do. If a step is obvious to the AI, leave it out. The goal is signal, not coverage.
RESOURCES
Further reading
Agent Skills open standard. The cross-platform specification adopted by 26+ AI tools. Your skills work everywhere.
Anthropic's knowledge-work plugins. Ready-made skills for common white-collar workflows: research, writing, analysis, and more. Start from what they built, then add your own context and standards. Generic skills produce generic results.
Skill marketplaces. Community directories where you can browse and install skills others have built. Skills Marketplace is one example. The same principle applies: start from something that works, then customize it with your own context and standards.
Skill Creator plugin. For Claude Code users who want a more structured build and evaluation workflow. Install with /install skill-creator.
Collaborating with AI. The companion guide that covers the CRAFT and OWN frameworks for effective AI communication. If you have not read it, start there.