The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements a robust sanitization function, _safe_name, using regular expressions to strip illegal filesystem characters (such as < > : " / \ | ? *) and control characters from user-provided data. This effectively prevents path traversal and potential command injection through crafted filenames.
[DATA_EXFILTRATION]: The skill and its associated scripts are designed for local-only execution. It utilizes standard Python libraries for file processing (csv, docx, openpyxl) and explicitly avoids network-capable modules. The documentation correctly identifies that personal data stays on the local machine and nothing is sent to external services.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from spreadsheet and text files, creating a potential attack surface for indirect prompt injection. This risk is effectively mitigated by the skill's mandatory workflow, which requires the agent to propose a token mapping and wait for explicit human confirmation before performing any bulk file generation.
Ingestion points: User-provided .docx, .xlsx, and .csv files processed via scripts/fill_template.py.
Boundary markers: The workflow enforces a review gate where the user must sign off on the tokenized template (ConfirmationLetter_tokenised.docx).
Capability inventory: Limited to local file reading (Document, load_workbook) and local file writing (Path.mkdir, save). No network access or shell execution capabilities are present.
Sanitization: Filenames are sanitized using the _safe_name helper to prevent filesystem abuse.
[PROMPT_INJECTION]: The skill instructions contain strong guardrails against model hallucination, specifically the "Never Invent" principle, which ensures that missing data is represented by a visible «MISSING: Token» flag rather than the model attempting to guess values.

fill-template