Agent Skills
A skill is a folder with a SKILL.md file that teaches the AI how your company does one job, such as writing a quotation or preparing a month-end reconciliation. OrgRelay follows the open Agent Skills standard, so the same folder works in Claude and other tools.
Connectors give the AI data and actions; skills give it judgement. A WooCommerce connector lets the AI read orders. A skill tells it that your company never quotes a discount without checking the price table first.
How OrgRelay uses skills
The AI normally sees only each skill's name and description. When a task matches a description, it loads the body and any reference files. That keeps prompts short: twenty skills cost twenty lines until one is needed.
OrgRelay ships four built-in skills:
| Skill | Purpose |
|---|---|
diagram-design |
Drawing diagrams. From the open-source diagram-design project (MIT) |
competitor-analysis |
Steps and red lines for competitor research |
trade-quotation |
How to write export quotations and quotation emails |
brand-style |
Company colours and tone for charts, reports and PDFs |
Companies add their own.
The SKILL.md format
A YAML header between --- lines, then Markdown instructions.
---
name: month-end-close
description: Month-end reconciliation checklist for the finance team. Use when someone asks to close the month, reconcile receivables, or explain a difference between the ledger and the bank.
license: Proprietary
metadata:
orgrelay-tools: "finance.receivables, finance.summary, query_table, create_approval_request"
---
# Month-end close
## Order of work
1. Pull open receivables with `finance.receivables` (overdue_only: true).
2. Compare the total with the bank export in the Finance folder using `query_table`.
3. List every difference above 50 USD with the invoice number and the source of each figure.
4. Do not post adjusting entries. Draft them and submit with `create_approval_request`.
## What not to do
- Never estimate a missing figure. Write "to be confirmed".| Field | Rule |
|---|---|
name |
Required. Lower-case letters, digits and hyphens. Must match the folder name |
description |
Required. Say when to use the skill, with the phrases people actually use |
license |
Optional |
metadata |
Optional. OrgRelay reads orgrelay-tools as a hint for which tools the skill relies on |
allowed-tools, compatibility |
Optional, from the standard |
The description does most of the work. "Finance checklist" will rarely be chosen. "Use when someone asks to close the month or reconcile receivables" will.
Reference the tools by their real names. Connector tools follow the standard tool names; OrgRelay's own tools include search_documents, read_document, query_table and create_approval_request, listed in the MCP guide.
Uploading a skill
Admin → More settings → Skills, upload a zip. The zip may contain SKILL.md at the root or inside one folder.
The upload is scanned. Any of these rejects it:
| Rule | Limit |
|---|---|
| File types | Text (.md .txt .json .yaml .csv) and images (.png .jpg .webp .gif .svg) only |
| Scripts | None. A scripts/ folder with files, or any file starting with #!, is rejected. OrgRelay never runs code from a skill |
| Paths | No .., no absolute paths |
| Size | 2 MB per file, 10 MB and 200 files in total |
| Header | Must meet the standard: valid name, non-empty description |
SVG files are sanitised. Suspected prompt-injection text and external URLs are listed in the scan report for the reviewer to read.
Review and access
A skill that passes the scan is pending review and nobody can use it. The owner reads it, then clicks Approve and enable. Access is granted by department; built-in skills are available company-wide by default. Uploads, rejections, enabling, disabling, access changes and downloads are all recorded in the activity log.
This mirrors how write tools need approval: nothing that changes how the AI behaves reaches staff without a person looking at it first.
Skills over MCP
Staff who use OrgRelay from Claude Code, Claude Desktop or Codex get three skill tools:
| Tool | Does |
|---|---|
list_skills |
Skills available to me |
load_skill |
The body and file list of one skill |
read_skill_file |
One reference file |
They can also download a skill as a zip from My account → Skills and install it in their own client. Either way, they only get skills granted to their department.
Writing good skills
- One job per skill. "Quotations" and "complaint replies" are two skills.
- Point to sources. "Use only prices from the price table" is better than a price list inside the skill, which goes stale.
- Say what to do when data is missing.
- Route anything that changes data through approval, as the example does.
- Keep secrets out. Skills are text the AI reads and staff can download.
The business view of how skills shape department agents is on orgrelay.com's agents page.
Common questions
Can a skill run scripts? No, never.
Skill or connector? Connectors are data and actions. Skills are know-how.
In Claude or Codex? Yes, through MCP or by zip download, limited to your department's skills.
Next
- Build the connector the skill will call.
- Learn the enterprise MCP patterns that apply to skills shared across a company.
- Read the security requirements for anything a skill can trigger.
- Pair skills with connectors that already follow the standard names, such as Odoo, QuickBooks Online or Salesforce.
- Package know-how that depends on an approval step together with approvals and write tools.
- Developers can check how skills are listed over MCP in Cursor as well.
- Skills reach staff in Claude through the same connection described in Claude Desktop and claude.ai and Codex.
- Skills that call connector tools assume those tools exist; the connector spec and the connector directory show what is available.
Related pages
- Authentication: tokens, OAuth client credentials and rotation.
- MCP transport: serve a connector over Streamable HTTP.
- Self-test tool: the self-test tool and what each check means.
- Code templates: download the developer kit and four templates.
- Human-in-the-loop: why approvals live outside the AI client.