The Agent Skills Directory

[PROMPT_INJECTION] (HIGH): Identified high-risk Indirect Prompt Injection surface (Category 8). The skill ingests untrusted data from external spreadsheets and possesses side-effect capabilities that can be abused. Evidence: 1. Ingestion points: The read command in scripts/gsheets.ts pulls cell data into the agent context. 2. Boundary markers: No delimiters or instructions to ignore embedded commands are present in the skill's guidance. 3. Capability inventory: Commands write, append, create, and add-sheet allow for significant state modification. 4. Sanitization: No sanitization of ingested cell data is specified.
[COMMAND_EXECUTION] (LOW): The skill executes local TypeScript files using npx tsx. While these are local scripts, they represent the execution of logic outside the model's direct control.
[DATA_EXPOSURE] (MEDIUM): The skill accesses sensitive local files including OAuth credentials at ~/.config/google-skill/credentials.json and tokens in .claude/google-skill.local.json.
[METADATA_POISONING] (MEDIUM): The setup instructions reference a gmail.ts script for a Google Sheets skill, which is misleading and suggests potential hidden cross-service functionality or poor maintenance.

gsheets