Agent Skills for Claude

TL;DR: You will author an open SKILL.md Agent Skill that packages Microsoft 365 governance expertise into a portable, version-controlled bundle Claude loads on demand, so agents follow your policies and controls instead of improvising.

What you will learn

How the open SKILL.md format structures a skill: YAML frontmatter, instructions, and bundled resources.
How to write progressive-disclosure instructions so Claude loads only what a task needs.
How to package supporting files (reference docs, scripts, templates) into a self-contained skill folder.
How to load and invoke a skill in Claude or Cowork, the same format this library publishes.
How to apply governance controls to skill authorship: least-privilege knowledge, data minimisation, and audit trails.

Prerequisites

Familiarity with Microsoft 365 governance concepts: sensitivity labels, DLP, Purview audit, and Microsoft Entra ID roles.
A Claude account with Agent Skills enabled, or access to Cowork, for loading and testing the skill.
A text editor and Git, since skills are plain folders that belong in version control.
Comfort writing Markdown and YAML frontmatter.
Recommended: complete the earlier track tutorials on building and governing agents before this advanced one.

Understand the SKILL.md anatomy

A skill is just a folder. At its root sits a single SKILL.md file, optionally accompanied by reference documents, scripts, and templates the skill can draw on. Claude reads the frontmatter to decide when a skill is relevant, then loads the body only when a matching task arrives. This is progressive disclosure: minimal tokens until the skill is actually needed.

Step 1: Lay out the folder

Create a dedicated folder named after the skill, for example m365-label-governance.
Place a SKILL.md file at its root.
Add subfolders such as references/, scripts/, and templates/ only for material the skill genuinely needs.

A typical layout looks like this:


m365-label-governance/
  SKILL.md
  references/
    sensitivity-labels.md
    dlp-policy-catalogue.md
  scripts/
    export-label-coverage.ps1
  templates/
    label-review.md

Step 2: Write the frontmatter

The YAML frontmatter is the part Claude scans to match a skill to a task, so the description must be specific and outcome-focused. Keep it tight.


---
name: m365-label-governance
description: Review and remediate Microsoft 365 sensitivity label and DLP coverage. Use when auditing label scope, classifying data, or checking policy gaps.
---

Write the description in the third person and lead with what the skill does, then when to use it. Vague descriptions cause Claude to skip a relevant skill or load an irrelevant one.

Write the instruction body

Below the frontmatter, the Markdown body is the operating manual Claude follows. Treat it as a procedure a careful practitioner would hand to a capable colleague, concrete steps, named tools, and explicit guardrails.

Step 3: Structure the procedure

Open with a one-line statement of the skill’s purpose and scope.
List the concrete steps the agent should take, in order.
Point to bundled resources by relative path so Claude opens them only when required.
State the boundaries: what the skill must not do, and when to stop and ask a human.

A trimmed body might read like this:


## Purpose
 
Audit Microsoft 365 sensitivity label and DLP coverage and propose
least-privilege remediations. Scope is read-and-recommend only.
 
## Procedure
 
1. Read references/sensitivity-labels.md for the approved label taxonomy.
2. Run scripts/export-label-coverage.ps1 to gather current coverage.
3. Compare coverage against the taxonomy and flag unlabelled or
   over-permissioned content.
4. Draft findings using templates/label-review.md.
 
## Boundaries
 
- Never change labels or policies. Recommend only.
- Do not export or quote document contents in findings.
- Escalate any Confidential-or-higher gap to a human reviewer.

Step 4: Bundle resources for progressive disclosure

Move long reference material out of SKILL.md and into the references/ folder, so the main file stays scannable.
Reference each file by relative path in the procedure, never paste its full contents inline.
Keep scripts small, single-purpose, and free of embedded secrets or tenant identifiers.
Use templates to standardise the agent’s output, which makes results reviewable and auditable.

Load and run the skill

Step 5: Install the skill in Claude or Cowork

Place the skill folder where your Claude or Cowork environment discovers skills, following your tenant’s deployment guidance.
Confirm the skill appears in the available-skills list.
Start a task whose intent matches the description, for example asking Claude to audit label coverage.
Verify Claude announces it is using the skill, then watch that it opens only the bundled files the task needs.

Step 6: Iterate and version

Treat the skill folder as code: commit it to Git and review changes through pull requests.
Tighten the description whenever Claude loads the skill at the wrong time, or misses it when it should fire.
Add a CHANGELOG entry for material changes so reviewers can track what the agent’s behaviour now permits.

Governance call-outs

Data minimisation: a skill is knowledge that travels. Never embed tenant data, document contents, secrets, or connection strings in SKILL.md or bundled files. Reference data sources; do not copy them in.
Least-privilege knowledge: scope each skill narrowly and prefer read-and-recommend procedures over autonomous change. Make destructive actions an explicit human escalation, not a default.
Identity and access: skills inherit the permissions of the agent that runs them. Confirm the agent’s Microsoft Entra ID identity holds only the roles its skills require, and review those grants under your P2 access-review cadence.
Compliance gates: route skill authorship through the same review pipeline as code, with pull requests, named approvers, and a CHANGELOG, so a skill cannot silently expand what an agent is allowed to do.
Audit trails: log when skills load and what bundled files and scripts they touch, and reconcile that against Microsoft Purview audit so you can answer who-did-what when an agent acts.

Hands-on lab

Write a minimal SKILL.md for one read-only audit task, load it into Claude or Cowork, run it, then compare it to a published Educ4te skill.

Before you start

A Claude account with Agent Skills enabled, or access to Cowork.
A text editor and Git, with a folder where your environment discovers skills.
A non-production tenant or a sample dataset, so the audit reads nothing sensitive.
An identity for the agent that holds read-only roles only.

Steps

Create a folder named m365-label-audit and add an empty SKILL.md file at its root.
Add YAML frontmatter with a name and an outcome-focused, third-person description that says the skill audits Microsoft 365 sensitivity label coverage and when to use it.
Below the frontmatter, write a ## Purpose line stating the scope is read-and-report only.
Write a ## Procedure with three or four concrete, ordered steps: gather current label coverage, compare it against the approved taxonomy, and flag unlabelled or over-permissioned content.
Write a ## Boundaries section that forbids changing any label or policy and forbids quoting document contents.
Place the folder where your Claude or Cowork environment discovers skills, then confirm it appears in the available-skills list.
Start a task whose intent matches the description, for example asking Claude to audit label coverage, and confirm Claude announces it is using the skill.
Open the skills library and compare your SKILL.md against a published Educ4te skill, noting differences in description specificity, procedure clarity, and stated boundaries.

A minimal starting point looks like this:


---
name: m365-label-audit
description: Audit Microsoft 365 sensitivity label coverage and report gaps. Use when checking label scope or finding unlabelled content.
---

Check your work

The frontmatter description is specific, third person, and states both what the skill does and when to use it.
The procedure has concrete, ordered steps that read data and report only.
The boundaries explicitly forbid any change and forbid quoting document contents.
Claude or Cowork discovers the skill and announces it when the task intent matches.
You can name at least one concrete difference between your draft and a published Educ4te skill.
Governance reminder: keep the skill read-only and run it under a least-privilege identity, so the audit can observe but never modify labels, policies, or content.

Next step

You have completed the track — browse the skills library to put it into practice.

Sources

Licensed under CC BY 4.0 by Educ4te .