The Agent Skills Directory

[SAFE]: The skill's primary focus is educational and operational documentation for developers. No malicious instructions, prompt injections, or persistence mechanisms were identified across the 15 analyzed files.
[COMMAND_EXECUTION]: The skill includes maintenance scripts and a test harness (tests/run.sh, scripts/check-model-table.py, scripts/context-budget.py) that perform local file operations and shell commands. These are legitimate utilities for verifying the integrity of the skill's documentation and providing cost-calculation helpers to users. The commands use safe invocation patterns (list-based subprocess.run) and are scoped to the skill's own directory.
[EXTERNAL_DOWNLOADS]: The scripts/check-model-table.py script queries the official Anthropic Models API (https://api.anthropic.com/v1/models) to verify that the model IDs listed in the documentation are current. This interaction targets a well-known, trusted service and is used solely for staleness verification.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: scripts/check-model-table.py parses local markdown files (SKILL.md and references/*.md) and the Anthropic API response to check for model ID drift.
Boundary markers: None present in the custom regex-based parsing logic.
Capability inventory: The script has the capability to execute curl via subprocess.run to query the Anthropic API.
Sanitization: The script performs structural validation on IDs but does not implement general sanitization of ingested markdown content. However, the attack surface is limited to the skill's own local files and the official provider API, representing a negligible risk.

claude-api-ops