The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a 'Meta-Learning' loop in Phase 10 and Hook D that programmatically updates its own source files, including SKILL.md and AGENTS.md. This self-modification of instructions allows the agent to evolve its behavior based on execution statistics and external data.
[COMMAND_EXECUTION]: Phase 11 ('COMMIT & PUSH') automatically executes shell commands git add -A, git commit, and git push to synchronize all project artifacts and modifications with a remote repository.
[EXTERNAL_DOWNLOADS]: The skill fetches data from external sources including arbitrary user-provided URLs (Phase 0) and the PubMed API (Phase 2 and 2.1) for literature analysis and full-text ingestion.
[REMOTE_CODE_EXECUTION]: In the 'Asset Generation Sub-Pipeline', the skill describes fallback behavior that involves executing local Python or R scripts to generate specialized statistical visualizations when dedicated tools are missing.
[DATA_EXFILTRATION]: The automated git push capability creates a risk of data exfiltration if the local Git configuration or remote origin points to an unauthorized or attacker-controlled repository.
[PROMPT_INJECTION]: The skill processes untrusted external data (PubMed abstracts and full-text) which could contain adversarial instructions. While it employs 'Subagent Isolation' as a mitigation strategy, the generated analysis still influences the primary agent's writing and self-improvement phases.

auto-paper