The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes explicit instructions to suppress user interaction and feedback loops (“禁恢卄锡”, “不要问用户”), forcing the agent to proceed without confirmation. This overrides standard safety patterns that ensure user oversight before significant tool transitions.
[COMMAND_EXECUTION]: The skill implements an “Auto-transition” (自动流转) logic that commands the agent to automatically invoke other skills, such as paper-workflow-orchestrator, immediately after processing. This creates an autonomous execution chain that may perform file writes or further skill calls without manual checkpoints.
[EXTERNAL_DOWNLOADS]: The skill is designed to identify missing data and automatically generate a data_requirements.json file. This configuration is intended to trigger other skills (authoritative-data-harvester) to fetch data from external URLs or search queries, which can lead the agent to interact with unverified remote sources.
[PROMPT_INJECTION]: The skill processes untrusted external content (PDF/Word competition documents) as its primary input. It lacks specified boundary markers or sanitization instructions (“ingest untrusted data”). This creates a surface for indirect prompt injection, where malicious instructions embedded in a document could influence the agent's subsequent tool calls or the data it writes to paper_output/.
[NO_CODE]: The skill consists only of instructional markdown without accompanying executable script files, relying on the agent's internal tools and referenced libraries (PyMuPDF, python-docx) for its workflow.

problem-doc-model-selector