Back to skill
Skillv4.0.0
ClawScan security
AI.MD · ClawHub's context-aware review of the artifact, metadata, and declared behavior.
Scanner verdict
BenignMar 11, 2026, 11:16 AM
- Verdict
- benign
- Confidence
- high
- Model
- gpt-5-mini
- Summary
- The skill's declared purpose (convert CLAUDE.md to a structured, AI-native format) matches what it contains and it requests no credentials or installs; it reads/writes CLAUDE.md and creates backups as part of normal operation.
- Guidance
- AI.MD appears internally consistent and aligns with its stated purpose, but before installing or running it: 1) Preview mode first — inspect the proposed converted output and only apply changes you review. 2) Expect the skill to read your CLAUDE.md and related ~/.claude/* reference files and to create backups; verify backup locations and contents. 3) The README offers a curl-from-GitHub install snippet — only run that if you trust the source (review the SKILL.md on the GitHub repo first). 4) The skill may run test calls to models via the hosting platform (it does not request API keys itself); ensure you are comfortable with the platform using your model access for those tests. 5) If you keep sensitive secrets in CLAUDE.md or the referenced files, remove them or move them before using the skill. Overall: reasonable and coherent for its stated goal, but follow the preview/backup steps and inspect outputs before accepting changes.
Review Dimensions
- Purpose & Capability
- okName/description, SKILL.md, README, and examples all describe a conversion/distillation tool for CLAUDE.md files. The skill requests no binaries, env vars, or external credentials — which is proportionate for a doc-conversion/helper. Nothing requested appears unrelated to that purpose.
- Instruction Scope
- noteSKILL.md describes reading the user's CLAUDE.md and producing a converted version, running a preview, creating backups, and testing with multiple LLMs. It also includes example references to files under ~/.claude/ref and example connection strings. These behaviors are consistent with a conversion/testing workflow, but they do involve reading and writing user config/docs (and running test calls to LLM endpoints via the platform). There are no instructions to exfiltrate data to third-party URLs, but the README shows an optional curl install-from-GitHub command (user-run) — review before executing.
- Install Mechanism
- okThis is instruction-only with no install spec in the registry; that is low-risk. The README suggests a user-run curl to GitHub raw content as an install convenience; that is common but should be treated as a normal code-download action (user decision). There are no opaque download URLs or extract operations in the skill metadata.
- Credentials
- okThe skill declares no required environment variables, no primary credential, and no config paths. The conversion/testing steps reference local files (CLAUDE.md, ~/.claude/ref/*) which is expected. Although SKILL.md says it will 'test with multiple models', it does not request any external API keys — it relies on the agent/platform's existing model access, which is typical for a skill running inside an agent.
- Persistence & Privilege
- notealways:false and disable-model-invocation:false (normal). The workflow explicitly reads and writes the user's CLAUDE.md, creates backups, and may update local reference files as part of conversion/testing; this is expected for a conversion tool but you should be aware it will modify your local prompt/config files (with backups). It does not request permanent platform-level privileges or modify other skills.