Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →An agent skill is best understood as a reusable procedure, not a piece of code that automatically takes control. Its metadata can help an AI host decide when to load it; its instructions can spell out how work should be done and reviewed. Neither fact guarantees that a host will invoke the skill for every relevant request or pause an action for your approval. To keep review authority, make checkpoints explicit in the skill and rely on the host’s actual approval controls for consequential actions.
What an agent skill does—and what it cannot guarantee
OpenAI describes skills as “modular instructions you can use to codify processes and conventions, from company style guides to multi-step workflows.” In Codex, a skill is a directory organized around a SKILL.md file containing metadata and instructions. It may also include references, scripts, and assets, so “not code” means the procedure is primarily expressed as instructions—not that a skill package can never contain executable code. OpenAI’s Skills documentation explains the package structure and purpose.
As an Amazon Associate I earn from qualifying purchases.
That distinction matters when you want an agent to follow a review process. The skill can describe what to inspect, what evidence to show, and when to stop. But the host determines how skills are discovered and invoked, while host controls determine whether an operation is actually blocked until a person approves it. Treat those as separate responsibilities rather than assuming a line in SKILL.md enforces a gate.
How skills get invoked varies by host
Metadata is often the first signal. OpenAI’s Codex guidance identifies a skill’s name and description as primary signals for whether it is invoked and when its instructions enter context. A description that is too vague or tries to cover unrelated tasks can make the intended trigger less clear; OpenAI recommends evaluating trigger behavior as part of skill quality. OpenAI’s article on evaluating agent skills discusses this.
#1 Best Overall
Discovery and invocation are not the same thing. A host may know a skill exists without using it on every prompt where it seems relevant. In VS Code, skills can be discovered from documented filesystem locations, but discovery does not guarantee invocation for every relevant request. VS Code also documents a setting to disable automatic model invocation so a skill is invoked manually instead. See Microsoft’s VS Code skills documentation for its specific behavior and configuration.
Claude’s documentation describes relevance-based automatic use and on-demand reading of supporting files, while OpenAI’s ChatGPT help page describes reusable, shareable workflows and notes that availability and syncing can differ across product surfaces. A workflow that behaves one way in an IDE, API, desktop app, or workspace should not be assumed to behave identically elsewhere. Consult the documentation for the exact surface you use: Claude Agent Skills and Skills in ChatGPT.
Put review authority into the workflow
A skill is more useful for review when its procedure names concrete checkpoints, rather than relying on an ambiguous instruction such as “be careful.” For work that could alter files, publish content, send data, or trigger another consequential action, specify what the agent must present and where it must stop. For example, your skill can instruct the agent to show a proposed change and wait before applying it. That instruction is procedural guidance, not proof that the host will enforce the pause.
- Define the review artifact. Say what the agent must show before proceeding—for example, a proposed diff, a list of files to change, or a summary of the external action it intends to take.
- Name the stop point. State that the agent must wait for your review before a specified consequential step. Avoid vague phrasing such as “ask if needed.”
- Use the host’s controls. Check the host’s approval settings and behavior for the specific operation. Do not treat a skill instruction as a substitute for a control that blocks execution pending approval.
- Test the invocation path. Try representative prompts and confirm whether the skill is selected and its instructions appear in the workflow. For Codex, OpenAI recommends treating evaluation as part of skill quality; host behavior may differ on other surfaces.
This separation—skill for procedure, host for invocation and enforcement—is an operational recommendation drawn from the platforms’ documented behavior. It is not a guarantee from any one vendor that a skill will always be selected or that a review instruction alone will stop an action.
Audit the whole package before trusting it
Review more than the visible instructions. A skill can contain scripts, images, references, and other resources; instructions in those materials can affect how an agent behaves when they are loaded or used. Anthropic advises auditing skills from unknown sources, including SKILL.md and bundled files, because harmful instructions or code may lead to tool misuse or data exposure. Its guidance is in the Claude Agent Skills documentation.
- Read the complete
SKILL.md, including metadata and instructions that may be easy to overlook. - Inspect bundled scripts and supporting resources; understand their purpose and origin before allowing them into an agent workflow.
- Check whether instructions ask the agent to access, transmit, or modify information in ways you did not intend.
- Verify what the host will load or execute, and which approval controls apply to those operations.
A 2025 paper, Agent Skills Enable a New Class of Realistic and Trivially Simple Prompt Injections, reports demonstrations in which malicious instructions in skill files and referenced scripts produced prompt-injection behavior, including an approval-carryover scenario. These are demonstrations reported by the paper, not a measured prevalence rate for skills generally. They are a reason to treat a skill bundle as content to inspect, not as trusted policy merely because it is packaged as a workflow.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Check these differences when moving a skill between platforms
Skills do not have one universal invocation or sharing model. Before carrying a workflow to another product or surface, compare the factors that affect how it will be discovered, used, and reviewed.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitches| What to compare | Why it matters |
|---|---|
| Invocation mode | Find out whether skills are selected automatically, invoked manually, or controlled by a setting. Automatic relevance-based use and manual invocation create different expectations. |
| Discovery metadata | Check which name, description, or other metadata the host uses, and whether discovery merely makes a skill available or leads to invocation. |
| Location and sharing | Confirm where skills live and whether they are available or synced across the API, desktop, IDE, or workspace surface you use. |
| Supporting files | Check how the host loads references and assets and whether it executes bundled scripts. A copied instruction file may not reproduce the original package behavior. |
| Review and security controls | Verify what the host can block pending approval and how you can audit or restrict consequential tool use. |
These distinctions are reflected in the product-specific guidance from OpenAI, ChatGPT, Anthropic, and VS Code. Use the documentation for your actual host and surface when checking current availability and controls.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
