The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes a local Python script provided in the skill folder to manage logging and scoring.
Evidence: SKILL.md contains multiple instructions to run python3 scripts/reflect_log.py for appending lessons, recording scores, and pruning stale data.
The script scripts/reflect_log.py uses only the Python standard library and does not perform any network operations or shell command interpolation.
[DATA_EXPOSURE]: The skill reads and writes to a local .paper-memory/ directory within the user's working directory.
Evidence: It accesses profile.yml, lessons.md, and scores.ndjson to personalize feedback and track improvements.
The skill explicitly advises the user to add this directory to .gitignore to prevent accidental versioning or exposure of local research context.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection by ingesting data from the lessons.md file to guide its critique process.
Ingestion points: Historical lesson data is read from .paper-memory/lessons.md in SKILL.md.
Boundary markers: No explicit delimiters are used for the ingested lesson text.
Capability inventory: The skill is restricted to local file I/O and text generation, with no ability to execute external code or access sensitive system paths.
Sanitization: The supporting Python script uses strict regex and JSON parsing to read the data files, ensuring the content is treated as data rather than executable logic.
[SAFE]: The overall design of the skill follows security best practices for AI agents, including the implementation of a deterministic 'regression guard' that prevents the agent from adopting changes that negatively impact measurable metrics.

reflect-and-improve