Skill Inspector: ESLint for Your AI Agent Skills
Discover Skill Inspector, a web-based validation studio and CLI for SKILL.md files to lint, format, and generate tests for AI agent skills.
Stock photo for illustration only, not from the actual event
- Skill Inspector is a web and CLI tool for agentskills.io packages
- Performs static validation, directory resolution, and token tracking
- Includes LLM-as-a-Judge scoring and generates 10-16 test scenarios
- Supports auto-fixes and includes a dedicated Claude Code linting mode
Agent skills are increasingly evolving into configuration files. A SKILL.md typically consists of frontmatter, a few instructions, and a folder containing scripts and references. Much like standard configuration files, they often break in subtle ways that are usually discovered only at runtime when the agent begins behaving strangely.
To catch these issues earlier in the development lifecycle, a developer has created a tool called Skill Inspector. It serves as a web-based validation studio and a command-line interface (CLI) for skill packages formatted according to the agentskills.io specification, functioning as a linter, formatter, and test generator for SKILL.md files.
Stock photo for illustration only, not from the actual event
Static validation covers every rule defined in the agentskills.io specification, including name formats and lengths, required fields, description limits, body size warnings, and unknown frontmatter fields. Each finding is tagged with a specific rule ID and a severity level of either FAIL or WARN.
Package resolution initiates when pointed at a skill directory using the --base-dir flag. It resolves all referenced files located under scripts/, references/, and assets/, subsequently flagging missing files, empty files, and any path traversal attempts such as ../.
Token economics provides a live breakdown of catalog footprints, activation bodies, and peak on-demand context usage if an agent reads every reference, allowing developers to understand skill costs prior to deployment. Additionally, the LLM-as-a-Judge feature scores four distinct dimensions—description quality, instruction clarity, scope and boundaries, and token efficiency—ranging from 0 to 100. It integrates with OpenAI, Anthropic, or local Ollama models, enabling usage without requiring an API key.
The introduction of specialized linting tools like Skill Inspector highlights the maturation of AI agent engineering, where configuration files require rigorous validation standards comparable to traditional software code. By shifting error detection, token cost estimation, and automated testing to the pre-deployment phase, developers can build significantly more reliable and predictable AI workflows.
For remediation purposes, the fix command displays a unified diff and applies deterministic fixes directly to the frontmatter and naming conventions. An optional LLM pass can further refactor the core instructions. Furthermore, the tool generates 10 to 16 synthetic test scenarios covering happy paths, adversarial inputs, tool failures, and scope boundaries, exporting them seamlessly into Pytest, JSON, or Markdown formats.
Stock photo for illustration only, not from the actual event
The tool also features a trajectory profiler that analyzes agent execution traces to flag redundant file reads, search loops, error ping-pong, and context bloat. Developers can install the CLI via script, run it locally via Docker Compose, or access the UI interface at http://localhost:5173. Licensed under the MIT license and currently at version v0.0.8, the project welcomes community feedback.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment