The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell commands (ls, cat, date) within the allowed-tools constraints to gather environment information and project context. This is limited to local directory listing and timestamping for document headers.
[DYNAMIC_CONTEXT_INJECTION]: Uses dynamic context markers (!) to embed the current date and a listing of the specs/ directory directly into the session context. These operations are benign, do not involve user-controlled input, and serve to keep the specification metadata accurate.
[INDIRECT_PROMPT_INJECTION]: The skill reads external project files such as README.md and existing specifications to maintain consistency. While this introduces a potential ingestion point for untrusted data, the skill's capabilities are limited to markdown document generation.
Ingestion points: Phase 1 instructions require reading CLAUDE.md, README.md, and files within the specs/ directory.
Boundary markers: No specific delimiters are used to isolate the contents of read files from the agent's primary instructions.
Capability inventory: The skill uses Read and Glob for file access, and Write to save the final markdown specification and a configuration file.
Sanitization: The skill relies on the underlying LLM's instructions to filter and summarize relevant context for the specification rather than executing instructions from the ingested data.

spec