The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes Python scripts (check.py, collect.py) and standard tools (pandoc, typst, git) to verify document structure and gather environment metadata. These executions are localized to the workspace and the skill's own directory.
[DATA_EXFILTRATION]: While the skill collects diagnostic data (environment facts, file hashes, and redacted proposal content), it explicitly maintains a local-only posture. The resulting bug-report/ bundle is written to the user's workspace, and the skill provides the user with local delivery options (email, manual GitHub issue) rather than performing network transmission itself.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied markdown files to detect structural errors.
Ingestion points: proposal.md and guidelines.md (processed in check.py and collect.py).
Boundary markers: None explicitly present for raw text processing.
Capability inventory: File writing (bug-report/), local shell execution of diagnostic scripts.
Sanitization: The diagnostic scripts use deterministic parsing and regex-based extraction rather than passing raw text directly to an LLM for structural validation, which mitigates typical injection risks during the mechanical check phase.
[EXTERNAL_DOWNLOADS]: The skill refers users to a well-known service (GitHub) to update their installation (github.com/hutzelmann/thesis-proposal-skills). This reference is documented neutrally as it targets a trusted repository relevant to the skill's purpose.

proposal-troubleshoot