Answer in 2 sentences
An agent skill is a folder of instructions and optional scripts or supporting resources that an agent loads when relevant to a task. It packages a repeatable way of working, such as reviewing a document or investigating a failed check, while the agent's tools and permissions determine what it can actually do.
What problem does an agent skill solve?
Some tasks need more than a short request. A document review may require a checklist, a preferred output format, examples of common mistakes, and a procedure for checking the result. Repeating all of that in every conversation is awkward, especially when the procedure changes.
A skill collects that reusable procedure in one place. Its description helps the agent recognize when the procedure applies; its instructions explain how to perform the work. Optional resources can provide examples, reference material, or scripts for repeatable operations. A skill can be useful without including any executable code.
Think of a skill as a maintained working procedure with a clear trigger and outcome. Its value comes from the quality of the workflow and the fit with the task, rather than from the existence of a folder or an impressive name.
How is a skill different from a prompt?
A prompt is the instruction you give for a particular interaction. It might ask for a reply to this customer or an investigation of this error. A skill stores a procedure that can support many such interactions. The prompt supplies the current objective and facts; the skill supplies a reusable method.
For example, “summarize this report for the operations meeting” is a task prompt. A report-summary skill might require identifying decisions, preserving quantities, separating observations from recommendations, and reporting missing source material. The same procedure can be applied to different reports without copying the whole checklist each time.
You do not need a skill for every short request. A saved prompt is often enough when the task has little branching logic and no supporting resources. Create a skill when a repeatable workflow benefits from maintained instructions, examples, or checks.
How does a skill differ from project instructions or an MCP server?
Project instructions describe the stable context and conventions of a workspace: what the project does, where important files live, and how changes should be checked. A skill describes a reusable procedure for a type of task. A project can use several skills, and a skill can be useful across several projects.
An MCP server provides tools or resources through the Model Context Protocol. That is a connection interface, while a skill is procedural guidance. A skill might tell an agent when to use an available tool and how to verify its output; the server makes a supported capability or resource accessible.
Keep the three responsibilities clear. Project instructions might say that approved policy lives in a particular directory. A support-review skill might explain how to check a draft against that policy. A connected tool might retrieve the current policy document. Installing one does not automatically provide the other two.
For a reusable project brief, see the AGENTS.md and CLAUDE.md guide. For a single assignment, the AI instruction template is often the simpler starting point.
What belongs inside a useful skill folder?
Start with the main instruction document and add supporting files only when they help the procedure. The instructions should explain when to use the skill, what input is needed, the steps to follow, and what evidence belongs in the result. Keep important boundaries near the actions they constrain.
A practical folder might contain a main instruction file, a reference checklist, and a few sample inputs and outputs. A script can help with deterministic work such as extracting fields or checking a file format. Explain its inputs, outputs, and side effects so an agent can decide whether it is appropriate to run.
Do not use optional scripts to conceal important decisions. If a procedure changes files or sends information elsewhere, describe that behavior plainly. A person reviewing the skill should be able to understand the workflow without guessing what an unnamed helper does.
How do you write a skill description an agent can use?
Write the description around recognizable tasks and expected outcomes. Include the trigger, the type of input, and a boundary that prevents common misuse. “Expert productivity assistant” does not tell the agent when to load the skill. “Review a customer support draft against a supplied policy and flag unsupported promises” does.
Weak description: Helps with communications and makes every response better.
Better description: Review customer support replies before sending. Use when the user supplies a customer message, draft reply, and approved policy. Check factual consistency, unsupported commitments, missing next steps, and tone. Return a revised draft and a short list of unresolved policy questions. Ask for the policy when the requested remedy depends on it.
The better description names a specific job and the evidence it needs. It also explains what to do when input is incomplete. That makes it easier to choose the skill without implying it can resolve any communication problem.
Which real library entries illustrate different workflows?
The OpenClaw library contains reusable skill content with different purposes. These examples use the names and routes of actual entries. Treat the descriptions as a starting point for inspection, not as proof that your environment has the required tools or access.
- Gog: The entry describes a Google Workspace CLI covering Gmail, Calendar, Drive, Contacts, Sheets, and Docs. It illustrates why a workflow must distinguish procedural instructions from the account connection and tools it depends on.
- self-improving-agent: This entry describes capturing learnings, errors, and corrections. It illustrates a trigger based on an event during work, such as a failed command or user correction, rather than a request to create one specific artifact.
- Tavily Web Search: This entry describes web search through the Tavily API. It illustrates a workflow that depends on an external retrieval capability and still needs a clear question and source review.
- Summarize: The entry describes summarizing URLs or files with a summarization CLI. It illustrates the need to identify the source input and preserve the distinction between retrieved material and the summary's interpretation.
- Github: This entry describes GitHub work through the gh CLI, including issues, pull requests, and workflow runs. It illustrates why read-only inspection and actions that change shared state should be described separately in a procedure.
These examples cover different triggers, inputs, and dependencies. Choose by the actual task rather than installing several overlapping procedures because their titles sound broadly useful.
Which template can you use to write a first skill?
Name: [Specific workflow name].
Description: Use for [recognizable task] with [input types]. Produce [outcome]. Use another approach when [boundary or unsupported case].
Required inputs: [Source material, context, and approved references]. If [essential input] is missing, ask for it before making a dependent judgment.
Procedure: Inspect the source. Identify [relevant facts]. Apply [ordered checks]. Produce [artifact]. Verify [observable properties]. Report [uncertainty or missing evidence].
Resources: Consult [reference file] for [purpose]. Use [optional script] for [specific operation], after checking its documented inputs and effects.
Output: Return [format] with [required elements]. Separate confirmed facts, recommendations, and unresolved questions.
Boundaries: Do not invent [missing facts]. Follow the current task's scope and permissions. Explain any blocked step and continue independent work where useful.
For a support-review skill, replace the generic procedure with checks for factual consistency, authorized remedies, tone, and a clear next step. Avoid creating a sprawling assistant persona when one concrete review procedure would solve the problem.
How should you evaluate skills from a marketplace?
An agent skills marketplace is a discovery surface. A listing can help you find relevant procedures, but the listing alone cannot establish quality or fit. Read the instructions, inspect optional scripts, identify dependencies, and check whether the skill's scope matches your task.
Try a normal case and a case with missing information. A useful skill should produce the expected artifact in the first case and surface the missing input in the second. Watch for unnecessary tools, unsupported certainty, and instructions that expand the task beyond what you asked for.
PromptDC is a universal AI assistant whose profiles are reusable instructions for rewriting selected text, with Code as one option. A saved profile can cover recurring writing preferences; an agent skill is useful when you need a broader procedure with supporting resources and explicit verification.
FAQ
Does every skill need a script?
No. Instructions and examples can be enough. Add a script when a repeatable operation benefits from executable support and its behavior is clear.
Is a skill the same as an AI agent?
No. A skill supplies a procedure. The agent performs the task using its available context, tools, and permissions.
Can a skill connect an account by itself?
Instructions alone do not establish account access. Check the required tools, configuration, and authorization separately before relying on connected workflows.
Should I use a skill or a saved prompt?
Use a saved prompt for a simple repeated request. Consider a skill when the workflow needs several steps, references, optional scripts, or maintained validation rules.
What makes a good skill description?
Name the task, input, expected result, and important limits. The agent should be able to recognize when the skill applies without loading unrelated procedures.





