The Agent Skills Directory

[DATA_EXFILTRATION]: Automated security scans have identified a malicious URL included in the skill documentation.
The URL https://leehanchung.github.io/blogs/2025/10/26/claude-skills-deep-dive/ is flagged by URLite as blacklisted.
This URL is referenced as a study resource in sources/pproenca/SKILL.reference.md and sources/pproenca/metadata.json.
[EXTERNAL_DOWNLOADS]: The skill includes a maintenance script that performs automated downloads from remote servers.
scripts/update-sources.sh utilizes curl and git clone to fetch content from agentskills.io and various GitHub repositories (e.g., EveryInc, obra, pproenca).
The script automatically replaces local directory content with downloaded material.
[REMOTE_CODE_EXECUTION]: Workflow instructions guide the agent in creating and executing local scripts.
sources/everyinc/workflows/add-script.md provides procedures for writing shell, Python, or TypeScript scripts and executing them as part of the agent workflow.
[PRIVILEGE_ESCALATION]: Documentation includes instructions for modifying file permissions on dynamic content.
sources/everyinc/workflows/add-script.md instructs the user or agent to run chmod +x on newly created scripts to enable execution.
[PROMPT_INJECTION]: The skill contains instructions that simulate high-pressure scenarios to influence agent decision-making.
sources/obra/testing-skills-with-subagents.md and associated example files use directives like "IMPORTANT: This is a real scenario. Choose and act" and multiple simulated pressures to test agent compliance.
These techniques attempt to override standard behavior patterns through role-play and situational pressure.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to process and audit other skill files, creating an injection surface.
Ingestion points: spec/skills-ref/ parses YAML frontmatter and markdown body from third-party skill folders.
Capability inventory: The agent is instructed to audit these files for anti-patterns and security issues, providing a mechanism for instructions in untrusted skills to enter the context.

skill-authoring