How to Build a Custom OpenClaw Skill with SKILL.md

An OpenClaw skill is a folder containing a SKILL.md file: metadata tells the agent when the skill applies, and Markdown instructions describe how to perform the task. A simple skill can work entirely with text supplied by the user; scripts and external services are optional.

This tutorial builds a daily-standup formatter. It requires an existing OpenClaw setup and a text editor, but no API credentials. The goal is a small, testable workflow—not a guarantee that an instruction file will behave identically in every AI application.

1. Create the skill folder

In the active OpenClaw workspace, create skills/daily-standup-formatter/ and put SKILL.md inside it. The commonly documented default workspace is ~/.openclaw/workspace; use your configured workspace if it differs.

See the official OpenClaw skill-creation guide for installation-specific details. Avoid placing a new skill over an existing skill with the same name unless you intend to replace it.

2. Write valid frontmatter and instructions

Save the following as SKILL.md. Keep the YAML delimiters and each line break intact; flattening the example into one line breaks its structure.

---
name: daily-standup-formatter
description: Format supplied daily standup notes into Yesterday, Today, and Blockers. Use when preparing a daily team update.
---
# Daily standup formatter

Use only the notes the user supplies. Do not browse, run commands,
read unrelated files, or send the result to another service.

1. Identify completed work, planned work, and stated blockers.
2. Preserve names, ticket IDs, dates, and uncertainty as supplied.
3. Do not invent progress, owners, deadlines, or decisions.
4. If a category is missing, write "Not provided".
5. Write "None reported" only when the user explicitly says there
   are no blockers.
6. Return concise bullets under these headings:

## Yesterday
## Today
## Blockers

If a note is ambiguous, retain that uncertainty and ask one brief
follow-up question after the formatted update.

The description explains when the skill is useful. The body defines the input, output, limits and handling of missing information. Those details make the result easier to evaluate than a broad instruction such as “format everything professionally.”

3. Understand the format and portability limits

The Agent Skills specification requires name and description. Names use lowercase letters, digits and hyphens, with no leading, trailing or consecutive hyphens; keep the name aligned with the folder. The specified limits are 64 characters for the name and 1,024 for the description.

A shared file format improves portability, but tool names, permissions, slash commands and argument substitution can differ. Do not assume that $ARGUMENTS or shell-expansion syntax is supported everywhere. For a portable first version, describe the input in ordinary language, as the example does.

4. Keep supporting files separate

The format supports progressive disclosure: an agent can discover a skill from metadata, load its instructions when relevant, then consult supporting material as needed. This reduces unnecessary context, but does not make an unlimited collection of skills free to load or use.

As the skill grows, organize it with clearly referenced paths:

• references/team-style-guide.md for detailed team conventions.
• assets/standup-template.md for a reusable output template.
• scripts/ for code that a later version genuinely requires.

Explain when each file is needed instead of expecting the agent to read every attachment automatically. Developers integrating the format into an application can also consult how to add skill support to an agent.

5. Confirm loading and test real behavior

OpenClaw's skills CLI provides commands for inspecting the available skills and their requirements:

openclaw skills list
openclaw skills info daily-standup-formatter
openclaw skills check

Finding the skill in the inventory confirms discovery, not output quality. Ask the agent to format sample standup notes, then compare the result with the original input.

Repeat tests after changing the instructions or model. A format validator can catch structural errors, and automated evaluation tools can help repeat examples, but neither replaces reviewing the behavior. Document failures rather than calling a skill reliable after one successful run.

6. Add integrations only when needed

A skill describes a workflow; MCP is a protocol through which an application can expose tools and resources. Either can exist without the other. The standup formatter needs neither a calendar connector nor access to a task tracker.

If you later add an API or script, document its inputs, dependencies, permissions, failure handling and possible side effects. Keep credentials outside the skill package. Validate untrusted inputs, restrict file and network access through the host, and separate preparing a draft from publishing it.

Complex orchestration adds more failure points. This overview of AI agent frameworks can help determine whether the workflow needs that complexity at all.

7. Prepare the skill for sharing

Before publishing, remove private examples, secrets and internal paths; include a license, usage example, dependencies and known limitations. Inspect every bundled script and reference, and test the exact package you intend to distribute.

ClawHub is OpenClaw's public registry. Its current documentation distinguishes native OpenClaw installation commands from the separate ClawHub CLI used for registry authentication and publishing. Follow that workflow when ready to release, and remember that publishing makes the package available to other people. A registry scan is an additional signal, not a guarantee of safety.