Microsoft shipped Skills in public preview as part of AI in SharePoint in April 2026. The feature quietly changes what AI in SharePoint can do on a tenant you already pay for. Instead of typing one-off prompts, anyone with Edit permission on a site can now capture a repeatable workflow as a file, save it, and hand it to the whole team.
The rest of this guide covers what a Skill actually is, where it lives, how permissions apply, what it cannot do, and the two or three places a practitioner team can start without waiting for a governance programme.
What is a SharePoint AI Skill?
A Skill is a Markdown file that describes a multi-step workflow AI in SharePoint should run on request. It captures the steps, the inputs, the outputs, and any organisation-specific rules the agent should apply. Once saved, the Skill is available to everyone on the site with View permission.
You create a Skill by describing it in natural language in the AI in SharePoint chat panel. The agent drafts the Markdown, you review it, make adjustments, and save. There is no code editor, no deployment pipeline, no install step. The file goes to a system library on the site and becomes usable immediately.
Skills are not prompts and they are not Copilot plugins. They sit between the two. A prompt is one-shot. A plugin requires engineering. A Skill is a tested instruction set for work your team already does manually, saved in a format anyone can read and a non-developer can author.
What changed with Skills in preview
Before Skills, the unit of work in AI in SharePoint was a single prompt. Users typed an instruction, received output, and moved on. Nothing was reusable. Even a good prompt had to be rewritten every time, and the quality of the output depended on whoever typed it.
Skills change the unit of work from the prompt to the workflow. Your team writes one Skill for contract review. It gets tested, tuned, and committed. From that point on, every contract review in your tenant runs the same tested version. The person running it does not need to remember the prompt pattern.
Skills make the workflow the unit of work, not the prompt.
How Skills work end to end
Three phases. None of them require a developer.
1. Create
Open AI in SharePoint chat on a site page or document library. Describe the workflow you want to capture. Be specific about inputs, outputs, and rules. The agent returns a draft Skill as a Markdown file. Review the draft, ask the agent to adjust the steps if needed, and save.
2. Store
The Skill is saved to the Agent Assets library on the site, in the path /Agent Assets/Skills/<skill-name>/SKILL.md. The library is created and managed by the product. Standard SharePoint governance applies: permissions, retention, sensitivity labels, version history, and auditing all work the way they do for any other file.
3. Run
Trigger the Skill from chat. You can name it explicitly ("Run contract review on the selected documents") or simply describe the task. AI in SharePoint will load the relevant Skill automatically when your prompt matches. A Skill indicator card in the chat UI confirms the Skill was loaded and ran.
# Contract review skill
## When to use
When a user asks to review or validate one
or more contract documents in this site.
## Inputs
- Selected document(s) in chat
- Or library path passed by the user
## Steps
1. Read the contract.
2. Find the lawyer ID. Format: xxx-xxx.
3. If missing or wrong, add the file
name to the Invalid Contracts list.
4. Create the list if it doesn't exist.
## Output
A short summary back in chat with
pass/fail count and any flagged files.
Where Skills live in your tenant
Skills live in the Agent Assets library at the site level. They are not tenant-wide. If you want the same Skill on three sites, you save it in three libraries. That sounds heavy, but it matches how SharePoint governance already works and keeps the permission model clean.
| Attribute | Detail |
|---|---|
| Storage path | /Agent Assets/Skills/<name>/SKILL.md |
| File format | Markdown |
| Scope | Site |
| Version history | Yes, via standard SharePoint versioning |
| Sensitivity labels | Supported |
| Retention | Supported |
Permissions and governance
Skills inherit from SharePoint site permissions by default. The short version:
- Edit permission on the site is the bar for creating a Skill.
- View permission on the site is the bar for running a Skill.
- A Skill can only do what the user running it already has permission to do. It cannot expand access.
- If you want tighter control, break permission inheritance on the Agent Assets library and apply your own rules.
There are no separate admin controls for Skills in the Microsoft 365 admin centre during preview. Governance is handled with the tools SharePoint already gives you.
What Skills cannot do
This is where practitioner framing matters most. Skills are intentionally bounded. They cannot:
- Call external APIs or connect to systems outside SharePoint.
- Run custom code.
- Escalate a user's permissions or bypass sharing policies.
- Replace Power Automate for cross-system automation.
If the work you want to automate reaches across tenants, pulls from a non-Microsoft SaaS, or needs bespoke logic, Skills are the wrong tool. That is a Copilot Studio agent, a Power Automate flow, or a custom Anthropic-side build, depending on the shape. Skills are for the inside-SharePoint layer.
Where to start this week
The three highest-leverage places to start are the ones where your team is already doing the same work by hand, on the same type of file, on the same site. Pick one of these and write your first Skill for it.
- Document review or validation. Contract field checks, mandatory column population, format audits.
- Metadata and tagging. Auto-tagging from body content, controlled vocabulary enforcement.
- Library housekeeping. Folder restructures, naming enforcement, ROT flagging.
Pick one of those. Write the Skill in chat. Test on three real files. Adjust. Share across the team. That is your first Skill shipped, and the mental model for every Skill after it is the same.
The teams shipping tested Skills this quarter will look very different in six months to the ones still typing one-off prompts.