product
How to turn an SOP into an AI skill employees can run in Claude, ChatGPT or Copilot
A step-by-step guide to rewriting a standard operating procedure as an AI skill: a SKILL.md template, a worked example, MCP tools, tests, versions and an owning team.
- #skills
- #mcp
- #roles
- #claude
- #copilot
Most companies already wrote down how their work is done. Employees ask Claude, ChatGPT or Copilot for help, and the assistant never sees that method.
To turn a standard operating procedure (SOP) into something employees can run in their AI assistant, rewrite it as an Agent Skill: a folder with a SKILL.md file that holds a name, a trigger description and step-by-step instructions. Then connect the tools each step needs and publish the skill where every assistant can load it.
What does it mean to turn an SOP into an AI skill?
An AI skill is your SOP rewritten for an assistant. The Agent Skills specification defines it as a folder whose one required file, SKILL.md, carries a name, a description and instructions. According to the vendors’ documentation in November 2026 (linked below), Claude, ChatGPT and Microsoft Copilot Studio all read this format, so one skill can serve all three.
The skill holds the method; live data and actions come from connectors built on the Model Context Protocol (MCP), the open standard explained in what an MCP server is and how it works.
Which SOP should you convert first?
Start with a workflow, not a whole process. You do not agentize “manage a construction site”; you agentize “prepare the daily site report”. A good first candidate runs often, follows fixed steps, reads from a few systems and ends in a standard document.
The example below is fictitious sample data, not any real company’s procedure. Sample SOP-SITE-014, daily site report (version 3, owner: Site Operations), completed by the site foreman before 17:00 each working day:
- Record the crews on site: company, trade, headcount, hours.
- Record progress per work package; flag any task more than 2 days late.
- List deliveries; note missing or damaged material.
- Record safety incidents. Any injury is escalated to the health and safety (HSE) manager within 1 hour.
- Note weather and any weather-related stoppage.
- Attach at least 3 dated photos.
- Fill in the report template (sections A to F) and send it to the project manager and the client.
- File the report as
YYYY-MM-DD_<site-code>_daily-report.pdf.
Steps 1 to 5 gather data, step 7 has a fixed format, step 4 holds a hard rule.
What goes in a SKILL.md file?
A SKILL.md file has YAML frontmatter on top and Markdown instructions below. Only name and description are required. Optional references/ and assets/ folders hold long documents and templates. A template to copy:
---
name: your-skill-name
description: What the skill does and when to use it. Name the words employees type.
metadata:
owner: team-name
version: "1.0"
source-sop: SOP-ID
---
# Skill title
## When to use
## Steps
## Rules and escalations
## Checks before finishing
| Field | Required | Limit (per specification) | What to write |
|---|---|---|---|
name | Yes | 1 to 64 characters: lowercase, digits, hyphens; same as the folder | A stable identifier, such as site-daily-report |
description | Yes | 1 to 1,024 characters | What the skill does and when to use it, in the words employees type |
metadata | No | String keys and values | Owner, version, source SOP |
| Body | Yes | Under 500 lines | Steps, edge cases, output format |
Assistants keep only the name and description in context; the body loads when a request matches, and supporting files when a step points to them. A precise description therefore matters more than a long body.
How do you map each SOP step to the skill?
Convert one step at a time: what the assistant must do, which tool gives it the data, and what the step must produce. Judgement becomes instructions, live data becomes tool calls, formats become a template. The tool names below are examples only; use the names your own MCP servers expose.
| SOP step | Skill instruction | Example tool | Output |
|---|---|---|---|
| 1. Crews | Group today’s diary entries by company and trade | site_diary.get_entries | Table A |
| 2. Progress | Compare actual and planned progress, flag delays | schedule.get_tasks | Table B |
| 3. Deliveries | List deliveries, problems separately | deliveries.list | Section C |
| 4. Incidents | On any injury, stop and tell the user to alert HSE | incidents.list | Section D |
| 5. Weather | Add weather and stoppages | weather.get | Section E |
| 6. Photos | Ask for 3 dated photos if missing | None (user upload) | Section F |
| 7. Send | Show a draft, send only after confirmation | mail.send | |
| 8. File | Apply the naming rule | files.save | Named PDF |
Write the description with the phrases foremen type (“daily report”, “site report”) and name what the skill does not cover. Move rules with consequences into a dedicated section of “always” and “never” statements the assistant can verify.
The finished SKILL.md
---
name: site-daily-report
description: Prepares and sends the daily construction site report (SOP-SITE-014).
Use when a foreman asks for a daily report, site report or end-of-day report for
a named site. Do not use for weekly client reports.
compatibility: Needs MCP access to site diary, schedule, deliveries, incidents,
weather, mail and project files.
metadata:
owner: site-operations
version: "1.0"
---
# Daily site report
## Steps
1. If the site or date is missing, ask for it first.
2. Build Table A (crews) and Table B (progress); mark tasks over 2 days late.
3. List deliveries; list missing or damaged items separately.
4. Read the incident log and follow references/escalation-rules.md.
5. Add the day's weather and any stoppage from the diary.
6. If fewer than 3 dated photos are attached, ask for them.
7. Fill assets/report-template.md and show the full draft.
8. Send only after the user replies "send", then save the PDF.
## Rules
- Any injury: stop and tell the user to alert the HSE manager within 1 hour.
- Never invent headcounts or progress. If a source is empty, write "No data".
## Checks before finishing
- Sections A to F present, late tasks flagged, file name correct.
How do you make the same skill run in Claude, ChatGPT and Copilot?
The SKILL.md file is portable; the tools are the part to plan. Expose your systems once through remote MCP servers and connect each assistant to them, so the same tool names work everywhere. The table reflects each vendor’s documentation in November 2026.
| Assistant | Native skill support | Tools | Distribution limit |
|---|---|---|---|
| Claude | ZIP upload in settings, Skills API, Claude Code folders | Custom connectors (remote MCP) | claude.ai uploads are per user; surfaces do not sync |
| ChatGPT | Build, install and share skills in a workspace | Custom MCP server | Workspace owners control sharing |
| Copilot Studio | Upload SKILL.md to an agent (preview) | MCP tools and resources | One agent at a time; replace the file to update |
The catch is distribution: keeping one skill current across three assistants means three uploads per change, unless one central source serves the skill, and the tools it calls, to every assistant over MCP.
How do you test, version and own an SOP skill?
Treat a skill like software: test it, publish it from one source, version every change and name an owner. Anthropic’s documentation on skills for the enterprise, consulted in November 2026, recommends three to five test queries per skill and authors who do not review their own work.
- Triggering: three requests that should load the skill (“daily report for site LT-02”), two that should not.
- Output: run it on a past day and compare with the report actually sent.
- Rules: a day with an injury must stop the skill before sending.
- Empty data: a day without deliveries must show “No data”.
- Each assistant: repeat the triggering tests in Claude, ChatGPT and Copilot.
Keep it under version control, keep the previous version for rollback and rerun the tests on every change. The team that owns the SOP owns the skill, here Site Operations: when the SOP changes, the skill changes in the same release. The skill then captures the team’s know-how, not just its documents, as explained in knowledge vs. know-how for AI agents.
How do you run SOP skills across several assistants and teams?
Native tools leave distribution to you, and a skill covers only the method. In production it also needs the MCP tools it calls, the rights of the person asking and, sometimes, a form around the workflow.
skilder makes a team’s know-how available in everyone’s own AI assistant. The AI stays the assistant employees already use. The container is a role: it bundles a team’s skills, its scoped MCP tools from an internal, curated MCP registry, its rights and, where useful, an app interface. The industry calls such bundles plugins (how agent plugins and skilder roles differ). Roles are configured from a blueprint and run by the business team, while IT governs access, permissions, hosting and audit. As the skilder documentation describes:
| Need | How skilder handles it |
|---|---|
| One source for every assistant | Claude, ChatGPT, Copilot Studio, VS Code or any MCP-compatible agent connects over MCP with OAuth |
| Existing skills | Import a SKILL.md, export back to the Agent Skills format |
| Permissions | The assistant acts with the requester’s own permissions, applying least privilege to AI agents |
| Versions and trace | Versioning and rollback; execution trace per user, exportable audit log |
| Hosting | Switzerland (Infomaniak) and the European Union; on-premise on request |
skilder includes a governed MCP layer but is not a pure MCP gateway, and it can run alongside one. A team using only Claude Code on one repository may need nothing more than Git and Claude Code plugins; roles pay off once several assistants and teams share skills. To choose the first SOP and its role, a free 30-minute exchange with an expert helps structure AI adoption in the team.
What are the most common mistakes?
Failed SOP skills usually share one of six causes: the whole SOP pasted in, a vague description, data out of reach, rules buried in prose, sending without confirmation, and no owner. Each has a simple fix.
| Mistake | What happens | Fix |
|---|---|---|
| Pasting the whole SOP | History confuses the assistant | Keep actions, rules, output; move background to references/ |
| Vague description | The skill loads for the wrong request, or never | Name the task, trigger phrases and exclusions |
| Data out of reach | The assistant guesses numbers | One MCP tool per source, “No data” when empty |
| Rules buried in prose | Escalations get skipped | A rules section, repeated in the final checks |
| Sending without confirmation | Unreviewed emails reach clients | An explicit “send” before any write action |
| No owner, no version | Copies drift | One source, one owning team, versioned releases |
A skill from an unknown author is also a security risk: it can make the assistant run tools or scripts. Review every file first. Unreviewed SKILL.md files spreading through a company are shadow AI (AI used outside any approved framework), as SKILL.md, the king of shadow AI shows.
Takeaway: start with one frequent SOP, write its description as carefully as its steps, and give the skill an owner before you give it to a second assistant.
To map your first SOP to a skill and decide where it should run, book a free 30-minute exchange with an expert.
Frequently asked questions
How do I make our SOPs usable in ChatGPT or Copilot?
Rewrite each SOP as an Agent Skill: a SKILL.md file with a name, a description that names when to use it, and numbered instructions. Connect the systems each step reads through MCP connectors. Then upload the skill to ChatGPT or Copilot Studio, or serve it from one central source, packaged with its tools, so every assistant gets the same version.
Does the same SKILL.md work in Claude, ChatGPT and Copilot?
Yes. The file format comes from the open Agent Skills specification, and in November 2026 the documentation of Claude, ChatGPT and Microsoft Copilot Studio describes support for it. Test it in each assistant anyway: tool names, script execution and network access depend on the host, and each model decides differently when to load a skill.
Do I need MCP, or is a SKILL.md file enough?
A SKILL.md file alone is enough for a skill that only formats or checks what the user provides. As soon as a step reads a live system, such as an ERP, a site diary or a CRM, or sends something, the assistant needs a connector. MCP is the connector format that Claude, ChatGPT and Copilot Studio all accept.
How long should an AI skill be?
Short. The Agent Skills specification recommends keeping the SKILL.md body under 500 lines. A typical SOP skill fits in 50 to 150 lines of steps, rules and checks. Templates, long policies and background go into separate files that the assistant opens only when a step points to them, which keeps the context small and the instructions clear.
Who should own an SOP skill, the business team or IT?
The business team that owns the SOP. It holds the method, approves each version and keeps the skill in line with the procedure. IT governs what surrounds it: connectors, access, permissions, hosting and audit. Splitting ownership this way avoids the common failure where IT maintains a skill it cannot judge and the experts stop updating it.
How do we stop employees from using outdated versions of a skill?
Serve the skill from one source and pin versions in production. With native uploads, every copy per user or per agent must be replaced by hand at each change. With one central source that every assistant reaches over MCP, publishing a new version updates all connected assistants, and the previous version stays available for rollback.
Is it safe to install a skill someone else wrote?
Only after review. A skill can instruct the assistant to call tools, run scripts or send data, so an unreviewed skill behaves like unreviewed code. Anthropic's documentation advises reviewing every file in a skill before deploying it. Inside a company, keep an approved list of skills with a named owner for each one.
What is the difference between an SOP, a prompt and a skill?
An SOP is written for people. A prompt is a one-off instruction typed into an assistant and usually lost after the conversation. A skill is the SOP rewritten for an assistant, stored as a file with a trigger description, steps, rules and checks, so the assistant loads the same method every time the matching request comes up.
Related articles
-
November 17, 2026
How to share Claude skills with non-technical colleagues
A practical guide to sharing Claude skills with colleagues who never open a terminal: built-in sharing options, a publishing flow, common fixes, and when a central role fits better.
-
November 15, 2026
AI agent glossary: role, skill, MCP server and the terms around them
Plain definitions of the terms that come up when a company governs AI agents: role, blueprint, skill, SKILL.md, MCP, MCP server, connector, registry, context lake, execution trace and more.
-
November 11, 2026
MCP gateway vs skills platform vs Copilot Studio: which one to start with
An MCP gateway governs tool access, Copilot Studio builds Microsoft 365 agents, and what buyers call a skills platform is really agent roles. What each one does, and which one to start with.