The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run within its test suite (tests/make_fixtures.py and tests/run_tests.py) to automate data generation and validation.
Evidence includes calls to pdftoppm (part of the standard poppler-utils) to rasterize PDF pages for testing scanned-document scenarios.
The test runner invokes the skill's own scripts (extract_statement.py and prove.py) using the current Python interpreter.
These operations are limited to the testing environment and do not pose a risk during normal skill operation.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external PDF files, which represents a potential surface for indirect prompt injection if the agent were to blindly follow instructions embedded in the PDF text.
Ingestion points: Untrusted data enters the context via pdfplumber.open() in scripts/extract_statement.py.
Boundary markers: The skill uses strict CSV/JSON output formats and emphasizes deterministic parsing, which serves to isolate the data from the agent's instructions.
Capability inventory: The skill is limited to local file system writes (CSV, JSON, XLSX) and has no network access or arbitrary shell execution capabilities.
Sanitization: Data is validated through regex for dates and decimal.Decimal for currency, with a dedicated prove.py script that verifies arithmetic invariants (opening + sum = closing).
The skill's design, which mandates script-based extraction over visual reading, significantly mitigates the risk of an agent being misled by malicious content within a statement.

statement-extract-and-prove