Point this at a messy folder — notes, screenshots, half-written drafts, PDFs — and get back a clear picture of what you were actually trying to do, grouped into coherent "threads," each with a plain-language brief and a next-action list.
Say something like:
- "Reconstruct this folder for me"
- "What was I working on in here?"
- "I have a pile of screenshots and notes about something — help me figure out what it was"
- "Pull these scattered files into a project brief"
Claude will scan the folder, read across formats (OCR on screenshots, text extraction from PDFs, direct reads of notes), group the fragments by what they seem to be about rather than their file type or date, and hand back a brief per group: what it looks like you were doing, which files belong to it, and what to do next.
The skill bundles a script (scripts/thread.py) that does the heavy lifting —
cross-format extraction plus topic clustering via TF-IDF and agglomerative
clustering — and writes a markdown report. Claude then reads that report
alongside the original fragments and synthesizes the final brief in plain
language, sharpening the script's first-pass guesses with its own judgment.
The script needs a few Python packages, which the skill installs on first run:
scikit-learn, pytesseract, pdfplumber, pillow, numpy
Image OCR additionally needs the tesseract-ocr system binary. If it's not
present, OCR on screenshots returns empty text but everything else (PDFs, text
notes) still works fine.
This skill only reads files in the folder you point it at — it never modifies, moves, or deletes anything.