Feature request: Local pre-processing step before Cowork ingests large local file sets (esp. for legal/document-heavy workflows)

Context I'm a practicing lawyer using Claude Desktop (Chat + Cowork) for case work, primarily processing large batches of local PDFs — court filings, contracts, scanned exhibits. I've been testing claude-for-legal alongside a custom local skill I built (pdf-smart-extract, using pdfplumber/pypdf) that reduces PDF-to-text token usage by roughly 80–90% compared to direct ingestion. Problem When Cowork reads a local folder with a large number of documents (20+ PDFs, especially scanned or image-based ones), it processes each file individually in its native format. This is noticeably slower and more token-intensive than the same task run through Chat with manually pre-processed/uploaded text. For document-heavy professions (legal, but likely also finance, compliance, research), this creates a real friction point: batches large enough to justify Cowork's automation are exactly the batches where Cowork's current file-reading approach is least efficient. What I've validated Running a lightweight local extraction pass (via Claude Code + a custom skill) before handing files to Cowork/Chat produces a meaningful speed and token improvement. This isn't theoretical — I've been using this two-step workflow manually (Claude Code pre-processes → Chat Project consumes the summary → Cowork executes the final task) and it's noticeably faster than letting Cowork read raw files directly. Proposed improvement Cowork plugins already support Hooks (pre/post-tool triggers). I'd suggest formalizing this pattern as a built-in, optional capability:

When Cowork is about to ingest a local folder, detect file count/size/type (e.g., PDF-heavy, image-heavy, or exceeding a token-cost threshold). If the threshold is met, offer (or auto-trigger, if configured) a lightweight local pre-processing pass — extracting text/structure before the files enter the main task pipeline. Only the processed/lightweight output flows into the main Cowork task, rather than every raw file being read individually.

This would essentially productize the "Claude Code as pre-processing gateway" pattern that's currently only achievable by manually chaining Claude Code → Chat → Cowork. Why this matters for claude-for-legal specifically Legal workflows are inherently document-heavy and often involve exactly the kind of large, scanned, or poorly-OCR'd file sets that make this inefficiency most visible. A built-in pre-processing step would make the plugin's core value proposition (batch document review, timeline building, claim chart generation) meaningfully faster and cheaper to run at realistic case-file volumes. Happy to share more detail on the local skill I built if it's useful as a reference implementation.