The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from local session logs which contain previous user prompts. Malicious instructions embedded in those prompts could influence the agent during the retrospective analysis.
Ingestion points: scripts/transcript.py parses project session history from files in standard directories like ~/.claude/projects/ or ~/.cursor/projects/.
Boundary markers: The skill relies on simple digest markers ([turn] USER: ...) and model judgment without robust boundary delimiters or explicit 'ignore' instructions for the parsed content.
Capability inventory: The agent can clone repositories, push code branches, create GitHub Issues/PRs, and overwrite local skill files ('sync' functionality).
Sanitization: Prompts in the digest are clipped but not sanitized for instruction-like patterns.
[DYNAMIC_EXECUTION]: The skill supports a self-update flow where it loads and executes a newer version of its own instructions (SKILL.md) from a remote vendor-controlled repository (jinsyin/skills) if the local version is outdated.
[COMMAND_EXECUTION]: Local Python scripts (transcript.py) and external repository tools (e.g., ./skillsw install) are executed to perform session analysis and repository management tasks.

self-improve