The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes research notes and abstracts which could contain malicious instructions designed to influence downstream LLM tasks. The risk is mitigated by a deterministic Python processing script that uses extensive sanitization logic to strip narrative content.
Ingestion points: Reads data from papers/paper_notes.jsonl, papers/evidence_bank.jsonl, and outline/subsection_briefs.jsonl.
Boundary markers: The skill does not implement structural delimiters for processed text but enforces strict text cleaning rules to isolate evidential claims from authorial narrative.
Capability inventory: Limited to local file read/write operations within the project workspace; no network access or dynamic code evaluation (eval/exec) capabilities are present.
Sanitization: Implements a _sanitize_source_text function in scripts/run.py that utilizes extensive regex patterns defined in assets/source_text_hygiene.json to remove authorial self-narration and promissory language.
[COMMAND_EXECUTION]: The skill utilizes uv run python to execute its internal data processing script. This is the intended primary purpose of the skill for joining and validating research artifacts.

evidence-draft