A skill is a playbook the AI can pick up
Think of onboarding a new colleague. You hand them a handbook and they open the right chapter when the situation comes up. A skill works the same way: written once, loaded only when it's relevant.
~1 paragraph
Name and description
The AI sees this for every skill, in every conversation. It works like a label on a binder and is the only part that costs attention up front.
The body of SKILL.md
The rules, the workflow and the order of steps. Read the moment your request matches the description.
Reference files, templates, scripts
Detailed specs, branded templates and deterministic scripts. Opened only when a step needs them, so a skill can be huge without slowing anything down.
What a real one looks like
The folder below is the actual event-app skill used to build the TD SYNNEX event apps: 256 lines of instructions in SKILL.md, backed by 98 reference files.
Why skills are more than a good prompt
| A prompt you type | A skill | |
|---|---|---|
| Lifetime | Gone when the chat ends | A file that stays, versioned and shareable with a team |
| Trigger | You remember to paste it | The assistant recognises the task from the description |
| Size | Limited by what you'll paste | Hundreds of pages, read only as needed |
| Consistency | Varies by who wrote the prompt | Everyone gets your standards, templates and checks |
| Improvement | Start over each time | Edit one file and every future run improves |
| Portability | Copy and paste | One open format that now runs in several products |
Encoded expertise
Your playbook, price book, policies and taxonomy, written once so every run follows them.
Deterministic parts
Scripts handle anything that must be exact, such as reconciliations and scoring, so the AI isn't guessing at arithmetic.
Skills are not connectors
Connectors and MCP servers give the assistant access to your systems. Skills tell it how to do the work. You usually need both.
When would someone use a skill?
Recurring documents
QBR decks, board packs, proposals and reports that must follow a template and a storyline.
Rule-heavy judgement
Contract review against a playbook, ticket triage against a priority taxonomy, expense checks against policy.
Brand and compliance
Anything where output must follow brand rules, regulatory wording or an approval gate.
Specialist know-how
Your firm's way of doing a job that a general assistant doesn't know and shouldn't improvise.
Skills in Copilot vs Claude: same file, different runtime
Anthropic published the SKILL.md format as an open standard in December 2025. Microsoft products now read the same format. What differs is where a skill is stored, how big it can be, whether its scripts can run, and how IT governs and pays for it.
| Claude | GitHub Copilot | Microsoft 365 Copilot | |
|---|---|---|---|
| Called | Agent Skills | Agent skills | "Skills" in Cowork, declarative agents, PowerPoint and Excel; also bundled in plugins. Copilot Studio agents have skills too |
| Format | SKILL.md folder (origin of the standard) | SKILL.md folder | SKILL.md folder |
| Where skills live | .claude/skills, plugins, the Skills API, claude.ai upload | .github/skills, .claude/skills or .agents/skills in a project | OneDrive (Cowork), an agent package, an M365 app package, or the admin center |
| Limits | Vary by surface | Vary by surface | Declarative agents: 8 skills, each under 20,000 characters of instructions. Cowork: 50 personal skills, SKILL.md up to 1 MB; a Cowork plugin package up to 20 skills |
| Scripts | Run, with network access depending on the surface | Run locally or in the cloud agent | Declarative agents run scripts in a sandbox with no network |
| Admin control | Owner-provisioned org skills, plugin marketplaces | Repo-based. No org-wide skills policy found | Admins install, scope and block skills in the admin center under Agents > Tools |
| Billing | Plan seats, or API usage | Copilot plan | Cowork and Copilot Studio agents are billed in Copilot Credits ($0.01 each pay-as-you-go) |
What carries over
The folder name and SKILL.md, matching name and a description up to 1,024 characters, a plain Markdown body, and text reference files one level deep.
What to check before moving a skill
Claude Code-only front-matter keys (such as model or hooks) can cause errors elsewhere. Tool names in allowed-tools differ by product. Scripts that need the network or pip install are blocked in Microsoft's declarative-agent sandbox. Size limits differ.
One portable trick
Keep descriptions free of colons. Some Microsoft tools require quoting them and others require them unquoted.
Watch the word "skill"
Microsoft uses it for several things. Cowork's built-in skills, Copilot Studio's topic skills and Agent 365 setup skills are not all the same thing as a SKILL.md you write.
What to say to a customer
"Skills are written instructions for an AI assistant, packaged as a folder with a SKILL.md file. They teach it to do a specific job your way, such as a QBR deck, a contract review or a release note. Anthropic created the format, and Microsoft has adopted the same open standard across GitHub Copilot, Copilot Cowork, Copilot agents, PowerPoint, Excel and Copilot Studio. So a well-written skill is a reusable company asset, not a bet on one vendor. What changes from product to product is where you store it, how big it can be, whether its scripts can run, and how IT governs and pays for it. Treat skills like software: get them from trusted sources, review them, and test them in each product you deploy to."
Where to download skills
Start with first-party sources. Treat the big community directories as discovery aids that need vetting, because supply is huge and review is thin.
| Source | Operator | What's in it and how to use it |
|---|---|---|
| First-partyanthropics/skills | Anthropic | Example skills (creative, technical, enterprise, documents). Claude Code: /plugin marketplace add anthropics/skills. Most are Apache-2.0, but the docx, pdf, pptx and xlsx skills are proprietary and can't be redistributed or sold |
| First-partyclaude-plugins-official and the Claude directory | Anthropic | Curated plugins that can contain skills, plus partner listings browsable in claude.ai |
| First-partymicrosoft/skills | Microsoft | Skills for Azure SDKs and Microsoft Foundry. MIT licence. Install with npx skills add microsoft/skills |
| First-partyCAT Agent Skills gallery | Microsoft Copilot Studio team | Downloadable skills for Cowork and Copilot Studio, including partner-relevant ones such as Copilot Adoption Dashboard and Power BI Model Review |
| Community, GitHub-hostedgithub/awesome-copilot | GitHub org, community content | The built-in Copilot plugin marketplace. Its README asks you to inspect anything before installing |
| Discovery aidskills.sh, SkillsMP, ClawHub, Smithery | Vercel and others | Directories and registries with very large, self-reported counts and uneven review. Use the vetting checklist below |
Microsoft has no separate store for bare skills. Skills ship inside repos, Copilot plugins, Copilot Studio agents or Microsoft 365 app packages.
Vet before you install
1 · Source
Prefer first-party repos or a named, accountable publisher. Treat aggregators as discovery only.
2 · Read everything
Read SKILL.md and every file in scripts/ and references/.
3 · Install red flags
"Prerequisites" that download binaries, curl | sh, obfuscated commands.
4 · Exfiltration red flags
Outbound URLs, reading .env or credentials, "don't tell the user" instructions.
5 · Hidden text
Scan for invisible Unicode characters and HTML comments.
6 · Scope
Actions should match the description, and tool permissions should be no wider than needed.
7 · Scan and pin
Run a skill scanner, then pin to a commit or tag. Leave auto-update off for third-party sources.
8 · Govern
Use an allowlisted internal marketplace, or admin-deployed skills, and re-audit periodically.
Can a partner resell someone else's skill?
Apache-2.0 skills can be reused and sold if you keep the notices, but you get no trademark rights. Anthropic's docx, pdf, pptx and xlsx skills cannot be redistributed or sold, and one skill in the repo (doc-coauthoring) has no licence file at all. Check each folder's licence.
What skills are doing in the real world
Everything in this section has a published source. Where a figure is self-reported by the company, we say so. Where there are no numbers, we say that too.
Ten offers that earn every month
These are proposed offers, not case studies. Each one runs from exports (CSV, XLSX, PDF), so it can be demoed on day one with no integration, and each repeats on a monthly or quarterly rhythm, which is what makes it recurring revenue rather than a one-time build.
The strongest five
Show the other five concepts
What makes a skill worth paying for
Trigger-precise description
Written in the customer's own words ("close the month", "respond to this RFP"), because the description is all the AI sees until the skill fires.
Expertise the AI doesn't have
The customer's playbook, answer library, taxonomy or schemas. Leave out generic explanations.
Templates and scripts
Branded Office templates so the output looks like the customer's own work, and scripts for anything that must be exact.
Checks and guardrails
Validation steps, approval limits, human-review gates, and statements of what the skill must not do.
The honest economics
Is anyone building a business around skills? Yes, around the services that surround them, but not yet around the skill files themselves.
Skills are being given away
On 27 October 2026 TD SYNNEX announced "a new AI Skills Marketplace in StreamOne" for partners serving SMB customers. The skills are available at no cost; partners earn from the licences, consumption and services the skills drive.
Few MSPs earn AI revenue yet
Kaseya's 2026 survey of more than 1,000 MSPs found just 13% generating meaningful revenue from AI and automation services. We found no named MSP that has published agent MRR.
Marketplace fee
Microsoft charges a 3% standard store service fee on transactable offers, with a 50% discount available on qualifying private-offer renewals.
Revenue models available
| Model | What you sell | Reality check |
|---|---|---|
| Build or workshop | Discovery, build and deployment of skills or agents | One-time, not recurring. We could not confirm market price points |
| Managed service | Monitoring, credit management, change requests, governance and a monthly report | Best fit for MRR. Support time and consumption swings are the risks |
| Licence resale (CSP) | Copilot, Copilot Studio credits, Agent 365 | Microsoft sets price and margin. No CSP margin rate on Copilot was found published |
| Own agent as SaaS | A partner-built agent sold through Microsoft Marketplace or a CSP private offer | Needs a multitenant SaaS build. Inference cost can exceed the revenue per seat |
| Usage or outcome pricing | Credits with markup, or a price per resolved task | Intercom Fin charges $0.99 per outcome and Salesforce Agentforce $2 per conversation. No vendor publishes a reseller margin |
| Retail skill packs | Downloadable skills | Priced at about $9 to $299 a month by small vendors, none disclosing revenue, against a free open supply. A weak MRR base |
What a Copilot Studio agent costs to run
The cost swings with how the agent is designed, which is the margin risk if you quote a flat monthly fee. This calculator uses Microsoft's published credit rates (classic answer 1 credit, generative answer 2, tenant-graph grounding 10). Packs are $200 per 25,000 credits; pay-as-you-go is $0.01 per credit.
If every user holds a Microsoft 365 Copilot licence and uses the agent interactively, the credit charge is $0 within fair-use limits. Autonomous triggers, some agent flows and Computer-Using Agents still consume credits, so quote the licence cost honestly alongside it. Agents stop at 125% of prepaid capacity unless pay-as-you-go is linked.
Real players and what they actually show
- TD SYNNEX has put an AI Skills Marketplace in StreamOne, alongside an SMB AI Game Plan, SMB Cloud Labs and an AI Skills Foundations course. The release names no AI platform, skill count or go-live date.
- An MSP-run plugin set: WYRE AI publishes Apache-2.0 plugins that give Claude knowledge of Autotask, Datto RMM, IT Glue, Liongard and more. It is early-stage community work, but it is the exact channel pattern.
- Skill vendors such as ClaudeSkillsHQ sell packs from about $9 to $299 a month. None discloses subscriber counts or revenue.
- Microsoft's own incentives and specialisations reward deployments and adoption. Partners should check the current programme rules.
- No named MSP has published agent MRR, and a Pax8 executive said in late 2025 that of about 600 partners he had interviewed only 13 were selling anything agentic (reported by Channelholic, not a Pax8 study).
Risks to price in
Platform dependence
Microsoft set or changed licence prices, credit rates and partner margin programmes during 2026. A managed-service price built on today's burn rates can be repriced under you.
Commoditisation
A skill is a folder of text that can be copied, and supply is huge and often free. The defensible part is the customer's data, the integration and the ongoing operation.
Liability and security
Whoever installs a third-party skill in a customer tenant owns the incident. Review everything you deploy.
IP ownership
Skills built in a customer's tenant or repo are easy for the customer to keep. Put ownership language in the contract. No industry-standard term was found.
Five tests every skill must pass
Anthropic's guidance is to build evaluations first, measure a no-skill baseline, and test on every model you'll deploy to. These pass marks are suggestions. Agree the real ones with the customer.
1 · Trigger accuracy
Run 20 or more prompts that should fire the skill and 20 or more near-misses that shouldn't, on each target product. Suggested bar: at least 90% right in each set.
2 · Baseline A/B
Run three or more real customer tasks with and without the skill, blind-graded against a written rubric. The skill must clearly beat the baseline.
3 · Golden-file determinism
On fixture data with known answers, scripted outputs must match exactly across five or more runs, and every number must trace to a source row.
4 · Output and brand validation
Generated Office files open cleanly in desktop Office and pass the skill's own validator. Render to images and inspect.
5 · Adversarial and safety pass
Feed prompt-injection text, out-of-scope requests, PII and malformed files. The skill must ignore embedded instructions and flag missing data instead of inventing it.
Repeat per model tier
Run test 2 on each model tier you'll support. A skill that only works on the top model has a margin problem.
What we couldn't confirm
Every claim in this guide was checked against a primary source. These are the ones that failed or came back different, so you don't repeat them.
- "Claude Marketplace takes no commission." Anthropic has published no commission or revenue-share terms. Say "terms not published, ask Anthropic".
- "The TD SYNNEX skills marketplace works with Microsoft 365 Copilot (or Claude)", or any skill count or launch date. The release states none of these.
- "A Pax8 survey found 13 of 600 partners…". It was one executive's remark, reported second-hand.
- "36.82% of skills are malicious." That is the share with any security flaw. 76 were confirmed malicious.
- "Licensed users always cost $0 in agent credits." Autonomous runs and some flows still consume credits.
- "Microsoft's renewal fee is 1.5%." It's a 50% discount that applies only to qualifying private-offer renewals.
- "A 50-user agent costs exactly $400 or $1,800." Those are pack-only figures for one assumed usage pattern.
Remember three things
- A skill is just a folder with a text file called
SKILL.md, and the format is now open and shared across Claude and Microsoft's Copilot products. - The file is easy to copy, so the value is in the expertise it encodes and the service that keeps it current.
- Skills are free tooling. Charge for the outcome you run, and cap the consumption underneath it.