Inspect every bundled file
Review scripts, references, templates, network calls, and generated outputs before installation.
Official source verified Jul 10, 2026
Agent Skills turn good workflows into reusable capabilities
A skill is a folder containing a SKILL.md playbook and, when useful, scripts, references, or assets. The format is now supported across major agent products, making it a practical portability layer for repeatable work.
Monitored revision 38a2ff82958a
Put context in the right layer
The formats are complementary. Keeping each one focused reduces repeated context and makes agent behavior easier to maintain.
LayerBest forTypical contents
AGENTS.mdRepository-wide operating contextCommands, architecture, conventions, boundaries, definition of doneRulesStatic or file-scoped guidanceAlways-on constraints, path-specific patterns, concise project factsAgent SkillsProcedural, on-demand workflowsInstructions, scripts, references, templates, and validation stepsMCPConnections to tools and dataRemote or local capabilities, authentication, schemas, and actions
Start with one bounded workflow
A useful description helps the agent decide when the skill applies. The body should make the required inputs, steps, output, and final checks unambiguous.
---
name: release-check
description: Verify a release candidate before production deployment.
---
# Release check
## Required inputs
- Target environment and release identifier
- Approved acceptance criteria
## Workflow
1. Run the documented build and test commands.
2. Check migrations, environment changes, and rollback steps.
3. Verify the critical user journey end to end.
4. Report failures with evidence; do not deploy on a failed check.
## Output
- Decision: ready or blocked
- Evidence reviewed
- Failed checks and owners
- Known limitations
Review a skill like executable supply-chain input
Skills may contain scripts and may direct an agent to use powerful tools. Portability does not imply safety or quality.
Check tool boundaries
Confirm the skill asks only for the permissions and systems required by its stated workflow.
Test trigger behavior
Use positive and negative prompts to verify when the skill loads and when it stays out of context.
Pin important dependencies
Record required packages, CLIs, versions, environment variables, and operating-system assumptions.
Evaluate the outcome
Run representative cases and compare accuracy, correction time, and failures against the workflow without the skill.
Assign an owner
Track source, license, maintainer, version, review date, and a process for updates or removal.
Primary resources
Agent Skills specificationCanonical SKILL.md structure, metadata, and directory format.OpenAI Skills catalogOpenAI-maintained skills for Codex, including system and curated examples.GitHub Copilot documentationSupported skill locations and surfaces across Copilot products.Cursor skills releaseCursor's distinction between dynamic skills and static rules.