file-intel

File Intel — Gemini File Processor

Runs scripts/process_files_with_gemini.py on a folder of files and produces Obsidian-ready summaries.

Step 1: Get the folder

Use AskUserQuestion:

Question: "Which folder should I process?"
Options:
1. "This vault's inbox/" — process the inbox folder
2. "Custom path" — user specifies a folder

If the user selects option 2, they'll type the path in the "Other" input.

Step 2: Run the script

Run via Bash from the vault root:

python scripts/process_files_with_gemini.py <folder_path>

• If inbox/: python scripts/process_files_with_gemini.py inbox/
• If custom path: pass it as the argument

Show the terminal output as it runs so the user can see files being processed live.

Step 3: Open the output

After the script completes, open the output folder:

open "outputs/file_summaries/YYYY-MM-DD/"

Replace YYYY-MM-DD with today's date from the script output.

Step 4: Report back

Tell the user:

• How many files were processed
• Where the summaries landed
• Point them to MASTER_SUMMARY.md as the single-file digest of everything
• Suggest: "Open Claude Code and say: Sort everything in inbox/ into the right folders"

Notes

• Supported formats: PDF, PPTX, XLSX, DOCX, CSV, JSON, XML, MD, TXT, PY, JS, HTML, CSS
• Output: outputs/file_summaries/YYYY-MM-DD/
• Each file gets its own *_summary.md
• MASTER_SUMMARY.md combines all summaries into one digest
• Summaries are context-aware: deliverables (invoices, reports) vs reference files (code, config) get different formats

---

Gotchas

• Encoding detection is best-effort, not deterministic: Files saved as Windows-1252 or Latin-1 may be processed as garbled UTF-8 instead of failing loudly. Spot-check the first summary of any batch from unknown sources — if accented characters render as mojibake, the source encoding was misdetected.
• Password-protected and encrypted PDFs return blank summaries: Gemini cannot extract text from locked PDFs but the script does not flag them as failures. Check the file size of each *_summary.md — anything under ~200 bytes is suspect.
• Scanned-image PDFs depend on OCR confidence: Low-DPI scans, handwriting, or rotated pages produce summaries with hallucinated content rather than honest "could not read." Verify scanned documents against the original before trusting downstream decisions.
• XLSX files with multiple sheets summarize only the active sheet: The processor reads what the workbook opens to by default; other sheets are skipped silently. For multi-sheet financials, split into separate files or expect partial coverage.
• MASTER_SUMMARY.md grows linearly and exceeds context on large folders: A 200-file inbox produces a digest too large to feed back into another LLM call without truncation. For batches over ~50 files, work from the per-file summaries instead of the master.
• Re-running on the same folder writes to a new YYYY-MM-DD/ subdirectory: Two runs on the same day overwrite each other; runs on different days produce duplicates without cross-reference. Clear or archive prior output before re-processing.