A build guide from AutomationTactics. It covers what a skill is, the seven parts every skill needs, and the click-by-click steps in Microsoft 365 Copilot, Claude, ChatGPT, and Google Gemini.
Verified against official documentation as of September 15, 2026.
A skill is a document. It holds the instructions for one task you repeat, written so an AI assistant can follow them without you re-explaining every time.
Think about a task you do every week. You open the same files and make the same handful of decisions, and the output looks the same every time. Right now, if you ask an AI to help, you retype all that context into the chat. A skill saves that context once. After that, the AI applies it automatically when the task comes up.
Documenting the work first matters because the skill can only be as good as the instructions inside it. If the steps live only in your head, the AI has to guess, and it guesses differently each time. When you write the steps down, you get two things at once. You get a record a human teammate can read, and you get the raw material for the skill.
This is the Document step in the AutomationTactics Document-Train-Automate method. Write down how the work runs, and the AI has something to follow. Anthropic describes building a skill the same way, as putting together an onboarding guide for a new hire.
Pick a task that you do at least weekly and that follows roughly the same shape each time. A task where every case is a special case is not ready to become a skill. Fix the process first, then document it.
A good skill has the same parts no matter which tool you load it into. Include these:
Copy this and fill it in. It works as a plain document, and it also works as the body of a SKILL.md file for Claude and Microsoft Copilot.
---
name: [Short descriptive name]
description: [One sentence. What the task is, and when to use this skill. Name the trigger words and file types. Be specific.]
---
# [Skill name]
## When to use this
[Describe the situations that should trigger this skill.]
## Inputs I will provide
- [Input 1, for example: a pasted list of line items]
- [Input 2, for example: the current month's budget file]
## Steps
1. [First step, written plainly]
2. [Second step]
3. [Third step]
4. [Continue as needed]
## Output format
[Describe the exact structure of the finished result. Headings, order, length, file type.]
## Example of good output
[Paste one real example here.]
## Things to avoid
- [Common mistake 1]
- [Common mistake 2]Do this once, in a plain document, before you open any AI tool.
When you finish, you have a document a person can follow. That same document becomes the skill.
If you cannot think of a task to start with, the idea prompt on the Skill Factory page asks your AI tool to list the tasks you repeat. Copilot users have a walkthrough of their own.
Microsoft splits this across a few places. For a non-developer, the simplest path is the agent builder inside the Microsoft 365 Copilot app, and that is where this section spends most of its time. Microsoft also shipped a feature named "Skills" in mid-2026 that lives inside Copilot Cowork, covered below. Copilot Studio is the advanced option and comes last.
Microsoft's agent builder produces a "declarative agent." That is Microsoft's term for a saved assistant made from your instructions, your knowledge sources, and starter prompts. It is the closest match to a skill for most users.
Licensing note. Microsoft states that Copilot Chat users, including free and pay-as-you-go, can create agents using the agent builder in Microsoft 365 Copilot Chat. When an agent runs inside Copilot Chat against your Microsoft 365 content, the usage sits on the user's Microsoft 365 Copilot seat.
On the Configure tab you can set these fields:
The minimum you need is a name, a clear description of when to use it, and the instructions. Knowledge sources and starter prompts are optional but worth adding.
Test it. After you create the agent, open the Try it tab and run it with a real request. Keep refining the instructions, knowledge, and prompts. You can adjust by natural language or by editing the Configure fields.
Share it. Select Create when you are ready. Then use the Share button to choose who gets access. Options include only you, specific users or groups, or everyone in your organization if org-wide sharing is enabled.
To make the agent broadly discoverable, open the More ellipses menu and select Submit to your org catalog. An admin reviews it in the Microsoft 365 admin center and, if approved, publishes it in the "Built by your org" section of the Agent Store.
Microsoft introduced a feature named "Skills" that uses the same SKILL.md format as Claude. Microsoft's documentation places it inside Copilot Cowork, the deeper, multi-step working mode you reach with the toggle in the Microsoft 365 Copilot app. The main Copilot Chat window does not have it.
Licensing note. Cowork requires a Microsoft 365 Copilot license, with Cowork enabled and usage-based billing turned on. Per Microsoft Learn, you can create up to 50 custom skills by placing SKILL.md files in your OneDrive /Documents/Cowork/skills/ folder, each up to 1 MB. That 50-skill limit is a June 2026 general-availability increase from the prior limit of 20.
To create a skill in Cowork:
Note on field limits. Microsoft's own documentation confirms the required name and description in YAML frontmatter, and its troubleshooting text refers to a description that "exceeds the character limit," but it does not publish the exact numeric caps. Some third-party guides cite a 64-character name and a 200-character description drawn from the Claude Skills spec. We could not verify those exact numbers in Microsoft's documentation, so keep the description short and specific and do not treat a precise cap as confirmed.
Test it. Start a session and give Cowork a request that matches the skill. Active skills appear in the side panel during a conversation.
Share it. Open the skill's detail page on the Customize page, select Share, and choose Only you or Specific users in your organization. When you change a skill you already shared, select Re-share to update everyone you shared it with.
Separate from Cowork, Copilot in PowerPoint and Copilot in Excel each have their own custom-skill feature, managed through a "Manage skills" dialog and a Skills folder in OneDrive. Those are per-app and are not the same as the Cowork Customize page.
Copilot Studio is the more customizable builder. It is aimed at makers who need connectors, governance controls, and publishing across channels. You can copy an agent you built in the Microsoft 365 Copilot agent builder into Copilot Studio when you need those extra capabilities.
Licensing note. Copilot Studio is priced on consumption. Since September 1, 2025, the currency for agents is Copilot Credits, replacing the older "messages" term. You buy credits through pay-as-you-go meters or prepaid packs.
A declarative agent that runs inside Copilot Chat against grounded Microsoft 365 content draws on the user's Microsoft 365 Copilot seat rather than Copilot Studio credits. Custom-engine and autonomous agents draw Copilot Studio credits.
For most non-developers documenting a repeatable task, the agent builder in path A is the right starting point. Move to Copilot Studio only when you need something the simpler builder cannot do.
Claude's feature is called Skills, and it is the clearest match to the concept in this guide. A skill is a folder with one required file, SKILL.md. That file starts with YAML frontmatter holding a name and a description, followed by markdown instructions.
Licensing note. Per the Claude Help Center, Skills are available on Free, Pro, Max, Team, and Enterprise plans, and the feature requires code execution to be enabled.
On individual Free, Pro, and Max plans you turn this on under Settings > Capabilities, confirming that Code execution and file creation is on. On Team and Enterprise plans, an owner enables both Code execution and file creation and Skills under Organization settings > Skills. Custom skills you create in claude.ai are private to your account on Free, Pro, and Max. On Team and Enterprise, you can share skills or have an admin provision them org-wide.
The description is the trigger. Claude reads it to decide when to load the skill. Anthropic notes that Claude tends to under-trigger, so make the description specific and slightly pushy about when to apply it. Instead of "How to build a dashboard," write something like "How to build a dashboard. Use this skill whenever the user mentions dashboards, data visualization, or internal metrics, even if they do not explicitly ask for a dashboard."
Claude has a built-in skill called skill-creator that builds a skill from a plain description. This is the easiest path for a non-developer, because you do not edit any files by hand.
Claude Cowork is the desktop working mode where Claude acts on your files. Skills you enable for your claude.ai account load automatically in Cowork. On Pro, Max, and Team plans, in Cowork on Claude for Mac, you can also record a skill. You send Claude a video of your screen, clicks, typing, and voice while you do the task once, and Claude proposes a skill for you to review before you save it. Recording is not available in chat, on Windows, or on Free and Enterprise plans.
For teammates who work in Claude Code, skills are files on disk. Place them in ~/.claude/skills/ for personal skills or .claude/skills/ for a project. Each skill is a folder with a SKILL.md file. This path is for developers and can be skipped by non-developers.
Test it. Start a new chat and give Claude a request that should trigger the skill. Try an explicit mention, "Use my weekly report skill," and an implicit one, "Draft this week's site report." Check that the skill loads.
Share it. On Team and Enterprise plans, open the skill and share it with specific colleagues, groups, or the whole organization, and admins can provision skills org-wide. On Free, Pro, and Max, skills are per-user. You can package the skill folder as a .zip and send it, and the recipient uploads it through their own Skills settings.
ChatGPT does not have a feature named "Skills" for everyday users. The closest equivalents are custom GPTs and Projects. Read the retirement note before you invest heavily in custom GPTs.
Retirement note. OpenAI has announced it plans to retire custom GPTs and move workflows to Plugins. New GPT creation is unavailable on personal ChatGPT accounts.
For affected Enterprise workspaces, the OpenAI Help Center lists three planned milestones. September 17, 2026 is the target for the migration experience and user banner. September 25, 2026 is planned as the end of creation of new custom GPTs. December 11, 2026 is the scheduled retirement, when custom GPTs stop running.
Existing GPTs stay usable until retirement. For a team that wants a durable skill today, Projects with project instructions is the safer choice.
Licensing note. Creating a custom GPT requires a paid ChatGPT plan, Plus or higher. Once created, a GPT can be shared and used by others.
Test it. Use the preview panel on the right side of the builder. Run a real request and adjust the instructions until the output is right.
Share it. Click Share and choose the access level. Options are only me, anyone with the link, and the GPT Store. Public listings require a category and a privacy policy.
A Project is a workspace that holds related chats, files, and one set of instructions that apply to every chat inside it. Projects are available on Free, Plus, Go, Pro, and Team plans, with smaller file caps on Free.
Test it. Start a new chat inside the project and give it a task. The instructions and files apply automatically. Confirm the output follows your rules.
Share it. Share the project with teammates by email. Personal memories are disabled in shared projects.
Google's feature is called Gems. A Gem is a saved custom version of Gemini with its own name, instructions, and optional knowledge files. Gems are free on every Gemini plan.
The minimum you need is a name and clear instructions. Knowledge files are optional.
Test it. Use the preview panel on the right. Enter a prompt, review the response, and edit the instructions. The preview does not save automatically, so click Save after you make changes.
Share it. Gems are private to your Google account by default. Inside Gemini for Workspace, click Share and choose specific email addresses, everyone in your organization, or a public link. There is no public Gem marketplace.
This one is for teams that keep work in a code repository. It is a developer tool, so most non-developers can skip it. It is included because the instruction file is simple and readable.
GitHub Copilot reads a plain-Markdown file named .github/copilot-instructions.md at the root of a repository. There is no YAML frontmatter and no special syntax. You write natural-language rules, Copilot detects the file automatically, and the rules apply to Copilot Chat and code review in that repository. Because the file lives in the repo, it is version-controlled and shared with everyone who pulls the branch, with no per-person setup.
Cursor uses a similar idea with its own rules files. Both are built for coding work. A general business skill belongs in Claude, Microsoft Copilot, ChatGPT, or Gemini.
The skill does not trigger. The description is almost always the cause. Rewrite it to be specific and to name the trigger words and file types. In Claude, make it slightly pushy about when to apply.
As a fallback, name the skill directly in your request, for example "Use my weekly report skill". In Claude and Microsoft Cowork, confirm the skill is toggled on and that code execution or the required setting is enabled.
The output is inconsistent. Add specificity to the instructions and include validation steps the AI can self-check against. Paste a real example of good output so the AI has a target. Test triggering and output separately, so you know which one is failing.
The skill or instructions are too long. Trim to the steps that matter. Microsoft's Copilot instructions field caps at 8,000 characters, and the custom GPT instructions field caps at 8,000 characters. If you hit a limit, move long reference material into a knowledge file and keep the instructions to the procedure.
It uses the wrong knowledge. Check that the right files or SharePoint sources are attached and that permissions let the AI read them. Ask the AI to list the files it can see before you rely on it.
It does too much or acts on things you did not intend. Only load skills from sources you trust. A skill is instructions the AI follows, so a skill from an unknown source can direct the AI in ways you did not expect. Read the full text of any skill before you install it.
What would change these recommendations. If Microsoft moves the "Skills" Customize feature into the main Copilot Chat app, path B becomes as easy as path A and should be promoted. If OpenAI ships its Plugins replacement with a clear no-code builder before the December 2026 retirement, revisit whether ChatGPT Projects is still the safer default.
A skill is a written, reusable set of instructions for one task you do often, saved so an AI assistant follows the same steps every time. Claude and Microsoft 365 Copilot both call it a Skill and both use a SKILL.md file. ChatGPT uses custom GPTs and Projects. Google Gemini uses Gems.
The tool you already pay for. On Microsoft 365 with Copilot, use the agent builder. On Claude, use skill-creator. On ChatGPT, use a Project. In Google Workspace, use a Gem.
The description is almost always the cause. Rewrite it to be specific and to name the trigger words and file types. As a fallback, name the skill directly in your request.
For anything you want to keep past 2026, prefer a Project with project instructions. OpenAI plans to end creation of new custom GPTs on September 25, 2026 and to retire them on December 11, 2026, and describes those dates as targets subject to change.