Custom Agent Builder
Create a focused .agent.md agent from a job description and the user’s conversation context.
The skill asks the agent to inspect conversation history and generalize role, tool preferences, and scope into an agent configuration, and to save the file, but it provides no sensitive-data handling, data-flow disclosure, least-privilege guidance, explicit confirmation, external-effect controls, or rollback procedure. Basic trust is retained, while those missing safeguards are deducted. The README, LICENSE, and SECURITY.md establish Microsoft provenance, MIT licensing, and a security-reporting path, but this supports governance only and does not prove the skill itself is safe.
The intended flow is understandable: review the conversation, clarify when needed, draft and save an .agent.md, then iterate and summarize. However, the required agent-customization skill and agents.md are not included, and the skill gives no template, write-failure handling, permission-error handling, abnormal-input behavior, or diagnostic feedback. Because this is static review with no skill-specific tests, the score remains below the static ceiling of 10.
The name, description, and argument hint clearly target creation of a custom agent for a specific job, and the clarification questions cover job, selection conditions, and tool preferences. Non-fit cases, semantic trigger boundaries, input/output format, Chinese-language support, and mainland-China environment fit are not specified, so the score is limited.
The YAML metadata is concise, and the body is layered into extraction, clarification, and iteration with an explicit dependency reference. It lacks installation guidance, template examples, FAQs, known limitations, versioning, changelog, and a skill-specific maintenance/update path. Repository-level README, LICENSE, package metadata, and Microsoft attribution provide project governance context but do not fill the skill documentation gaps.
The workflow plausibly produces an .agent.md tailored to a concrete job and uses clarification plus example prompts to support follow-up use. However, the governing template principles are in an unavailable dependency, and there are no representative outputs, acceptance checks, or evidence that the result is directly usable without substantial review. Static evidence supports only the core-task claim, so the score stays below 7.
The selected SKILL.md is auditable primary material, while repository files provide project-level license, version, maintenance, and some CI evidence. The supplied CI does not cover this skill's key paths, and there are no skill-specific tests, sample outputs, or independent reproduction records. Under the static calibration rule, the score cannot exceed 5.
- Provide and review the referenced agent-customization guidance and agents.md template before use; the key generation rules are otherwise unverifiable.
- The skill reads and summarizes conversation history but does not specify privacy filtering, sensitive-data exclusion, pre-write confirmation, or rollback; manually review it for confidential projects.
- Invocation boundaries, output acceptance criteria, and Chinese localization behavior are unspecified, so the generated .agent.md requires human review.
What does this skill do, and when should you use it?
This skill guides users through creating a custom agent file for a specific job. It reviews the conversation to extract an assumed role, tool preferences, and domain scope, then asks clarifying questions when the specialization is unclear. It drafts and saves an .agent.md file, identifies its weakest or most ambiguous parts, and asks for another review. After finalization, it summarizes the agent, suggests example prompts, and proposes related customizations.
Reviews conversation history and generalizes the user’s specialized role or persona, preferred or avoided tools, and domain or job scope; asks what the agent should do, when it should be selected, and which tools it should use or avoid when needed; drafts and saves an .agent.md file; identifies ambiguous or weak sections and asks follow-up questions; then summarizes the agent, suggests prompts to try, and proposes related customizations. It also instructs the user to load and follow the agent-customization skill and agents.md for the template and principles.
- A developer repeatedly works in a specialized technical context and wants to turn established role and tool preferences into an .agent.md file.
- A user needs an agent limited to selected tools or focused on particular file types, but the desired behavior is not yet fully specified.
- A team wants a reusable configuration for a defined software-development job and wants to review ambiguous instructions before saving it.
- A user has developed a consistent workflow and wants a custom agent that is more suitable for that job than the default agent.
What are this skill's strengths and limitations?
- Uses conversation history to extract an existing role, tool preferences, and domain scope instead of relying only on a one-time description.
- Asks targeted questions about the job, when to choose the agent, and which tools to use or avoid when the requirements are unclear.
- Provides an iterative workflow covering drafting, weakness review, final summary, example prompts, and related customizations.
- Has a concrete output target: a custom .agent.md agent file.
- Depends on the agent-customization skill and agents.md, whose contents are not included in the supplied source.
- Sets disable-model-invocation: true, so automatic model invocation cannot be assumed and platform adaptation may be needed.
- The source does not document standalone installation, save locations, invocation commands, or a test suite.
- It creates a custom agent rather than directly providing general coding, debugging, or project-management capabilities.
How do you install this skill?
The source does not document a standalone installation command or a target directory for compatible clients. The skill is located at extensions/copilot/assets/prompts/skills/create-agent/SKILL.md in the microsoft/vscode repository. The repository as a whole is distributed under the MIT license; capabilities of its other bundled skills should not be attributed to this skill.
How do you use this skill?
Invoke it manually in a compatible environment because its frontmatter sets disable-model-invocation: true. For example, start with: “Create a custom agent for code review that focuses on TypeScript files and avoids unnecessary tools.” Answer its follow-up questions about the job, selection criteria, and tool preferences, then review the generated file. The source does not specify an invocation command, installation directory, or client-specific trigger syntax.
How does this skill compare with similar options?
Compared with the default agent, this skill focuses on creating a specialized .agent.md configuration from conversation context and user clarification rather than directly carrying out the task. If the user does not need a persistent role, tool preference, or job-scope configuration, using the default agent may be simpler.