The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses standard local developer tools to analyze project context and history during its discovery phases.
Evidence: The skill instructs the agent to use rg (ripgrep) for searching codebase patterns and git log to identify recent changes (found in SKILL.md and references/spike-protocol.md).
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because it reads and processes untrusted data from the project's local files.
Ingestion points: The agent scans project root files (e.g., package.json, Cargo.toml), source code directories (src/, lib/), and git logs to understand the project stack and architecture (SKILL.md Phase 0 and Spike Protocol).
Boundary markers: The skill relies on its internal structured phases to manage data flow but does not specify explicit boundary markers or delimiters for untrusted file content.
Capability inventory: The agent has the capability to perform file searches, git log retrieval, web searches for technical documentation, and write documentation files to the local file system.
Sanitization: Content retrieved from the codebase is processed for insights without explicit sanitization or filtering of potential malicious instructions embedded in comments or code.

pocket-pitching