The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from project source code, configuration files, and internal memory logs to diagnose token efficiency. This activity creates an attack surface for indirect prompt injection if malicious content is embedded in the scanned files. Ingestion points include CLAUDE.md, MEMORY.md, and source code. No explicit boundary markers are present, though the skill requires user confirmation before saving any generated files. Capabilities include Bash execution and file writing.
[COMMAND_EXECUTION]: Shell commands are used to perform project audits, such as searching for .claudeignore, counting file lines, and inspecting internal memory directory sizes. These commands are diagnostic and limited to local metadata analysis.
[EXTERNAL_DOWNLOADS]: The skill recommends external tools like rtk and headroom for advanced token management. These are presented as user-facing recommendations for installation from GitHub rather than automatic remote code downloads by the skill itself.

context-doctor