The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs shell command execution to compile Typst documents and run verification checks. This is handled through a custom Python probe (workshop-deck.py) and a bash script (check-overflow.sh). These operations are intrinsic to the skill's purpose—automating deck generation and validation—and utilize argument lists in subprocess.run to prevent shell injection. The platform's mechanicalChecks also invoke these scripts.
[EXTERNAL_DOWNLOADS]: The skill utilizes the uv tool to manage and run Python dependencies at runtime. It specifically references well-known packages such as pypdf (for PDF text extraction) and pytest (for verification suite testing). These are pulled from the standard PyPI registry, which is a trusted source.
[INDIRECT_PROMPT_INJECTION]: The skill processes external research papers provided by the user, which presents a surface for indirect prompt injection. The skill mitigates this by: 1) using a structured Phase 1 and 2 (Clarify and Plan) where success criteria are defined; 2) employing a dedicated verification probe that checks the final artifacts against the plan; and 3) using a read-only Explore agent for qualitative verification lenses to prevent the model from executing malicious instructions that might be embedded in the source text.
[DYNAMIC_EXECUTION]: The skill dynamically generates Typst source fragments and combines them with templates at runtime. These generated files are then compiled using the typst compiler. The compilation is restricted to the project root directory via the --root flag, ensuring that dynamic inclusions cannot access sensitive files outside the project scope.

workshop