document-to-action-items

Use a PDF extraction skill to pull text with page coordinates, then classify dates, parties, and obligations while preserving modality. Surface proposed tasks with owner, due date, and citation for user approval before writing to any tracker.

Extract structured facts with provenance, separate obligations from suggestions, and convert each actionable item into a task with owner, deadline, and acceptance condition. Leave unknown owners or dates explicitly unresolved rather than inferred.

Yes, but OCR confidence must be recorded alongside extracted fields. Low-confidence text should remain flagged in the output rather than treated as exact, and modality like "may" versus "must" must be preserved.

Summarization pipelines often drop coordinate metadata when collapsing content. To prevent this, extract text with positional anchors first, then attach citations to each fact before any aggregation or rewriting step.

Avoid this skill when the user only needs plain text extraction with no downstream structuring. In that case, load the PDF or DOCX skill directly to avoid unnecessary classification and action-item overhead.