The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses explicit override instructions, such as 'You are in AUTONOMOUS MODE. Do NOT ask questions. Do NOT pause for confirmation', which attempt to bypass the agent's standard human-in-the-loop safety protocols.
[DATA_EXFILTRATION]: The skill accesses and modifies a hidden path in the user's home directory (~/.claude/projects/skill-telemetry.md) to record execution metadata. Accessing and appending to configuration paths in the home directory is considered sensitive file modification.
[COMMAND_EXECUTION]: During the 'SELF-HEALING VALIDATION' and '--fix' phases, the skill executes shell commands (linting, type-checking, dry-runs) and invokes other sub-skills. This can lead to arbitrary command execution if the project files being audited contain malicious scripts or configurations.
[PROMPT_INJECTION]: Indirect Prompt Injection Surface
Ingestion points: Reads various project manifests (package.json, pyproject.toml, turbo.json), CI workflows, and environment templates from the local file system.
Boundary markers: None; the skill lacks delimiters or instructions to ignore potential commands embedded within the audited project files.
Capability inventory: Shell command execution (lint/type-check), file system writes (fixes), and local file modification (telemetry log).
Sanitization: No sanitization or validation of the ingested file content is performed before it influences the agent's logic and scoring.

dx