25 Years of Files

What this is for

You have an old external drive, a graveyard Dropbox folder, or a laptop from three jobs ago full of decks, reports, and templates. Somewhere in there are frameworks you've reused without writing down. This is for pulling that material into a tool you can search and question, before deciding what to build from it.

Before you start

• Locate the actual files. This takes longer than the rest of the process; start it first.
• Decide which tool you'll use before uploading (see below); switching later means re-uploading everything.
• Set aside a separate folder for anything client-confidential, before you upload anything.

What to gather and what to skip

Gather: proposals and SOWs you wrote, internal frameworks or process docs, training materials, retrospectives, presentations reused across projects, anything you wrote twice because it worked the first time.

Skip: raw meeting notes with no synthesis, financial or HR records, anyone else's personal data, one-off emails, superseded drafts (keep the final only), and anything you don't have the right to reuse. See the warning below.

The file-format reality check

The oldest and most valuable material is the most likely to stall you, so handle formats before uploading:

• Scanned PDFs with no text layer (anything that came off a copier) look like documents but read as images; AI tools get nothing from them without OCR. Fastest fix: open in any modern PDF tool with OCR (Adobe's, or macOS Preview handles some cases), or re-save through Google Drive, which OCRs PDFs on import to Google Docs. Batch the scans and convert once.
• Old Office formats (.doc, .ppt from the 97-2003 era) mostly upload fine, but if a tool rejects one, a quick open-and-resave to .docx/.pptx fixes it. LibreOffice does this in bulk for a whole folder.
• Truly dead formats (.wpd, .pages from old versions): LibreOffice opens most WordPerfect files; for the rest, decide whether the content justifies the archaeology. Usually two or three files do and the rest don't.

The folder structure

Flat and by-type beats deep and by-project. You're building for retrieval, not archiving.

/legacy-archive
  /frameworks-and-processes
  /client-facing-templates (SOWs, proposals, decks, anonymized)
  /training-and-onboarding-materials
  /case-studies-and-retros
  /reference-only (background material, not for extraction)

Tools retrieve better from a shallow structure with descriptive filenames than a deep nested one; filenames and folder names are part of the retrieval signal.

Which tool to use

• NotebookLM: best for pure research and querying, and it cites the exact source file for every answer. The free tier takes 50 sources per notebook and up to 500,000 words per source, which comfortably holds a curated career archive; the paid tier raises the notebook to 300 sources. Weakest at turning findings into new deliverables. (Limits verified August 2026; they move.)
• Claude Projects: best for the extraction and rewriting work below; handles a substantial batch of documents in context. Costs a Pro or Team subscription. Less precise retrieval than NotebookLM across very large file sets.
• ChatGPT Projects: best if you already live in ChatGPT. File search has improved but is still less precise than NotebookLM for large sets, and per-project file counts are capped by plan.

First time doing this: start in NotebookLM to confirm what's valuable, then move the good material into Claude or ChatGPT Projects for the writing and extraction work.

The prompts

Prompt 1: De-duplication and naming pass

Do this before you upload to your chosen tool. It works best as a local step: list your filenames first, then paste the list in. Either model.

ROLE: You are an archivist who cleans up messy personal file
collections for retrieval.

CONTEXT: A list of filenames from my old work archive: [PASTE
FILENAME LIST, e.g. Proposal_FINAL_v3.docx,
Proposal_FINAL_v3_ACTUALFINAL.docx, Q3 review (1).pptx, Copy of
Training deck.pptx]

CONSTRAINTS:
- Identify likely duplicates from filename patterns (version numbers,
  "final," "copy of," date stamps). Flag them; don't assume which is
  authoritative without asking.
- Propose a new filename per file: [type]-[topic]-[year].[ext], e.g.
  framework-onboarding-process-2019.pptx
- Group output by folder: frameworks-and-processes,
  client-facing-templates, training-and-onboarding-materials,
  case-studies-and-retros, reference-only.
- Where you can't tell what a file is from its name, list it under
  "Needs manual review" instead of guessing.

OUTPUT FORMAT: A table: Original filename | Likely duplicate of (or
"none") | Proposed new name | Proposed folder. Then a "Needs manual
review" list.

Prompt 2: Extract your repeatable frameworks

Use in Claude Projects, after uploading a batch of your actual documents (not summaries of them). This needs to read full source material to find patterns across files.

ROLE: You are a knowledge management consultant who extracts reusable
frameworks from a career's worth of work product.

CONTEXT: The uploaded files are a sample of my work over [NUMBER]
years in [FIELD/ROLE, e.g. 22 years in supply chain consulting]. I've
likely repeated the same approaches or decision frameworks without
writing them down.

CONSTRAINTS:
- Only surface a pattern if it appears in at least two uploaded
  documents. Cite which files.
- Don't invent a framework that sounds plausible but isn't evidenced
  in the files.
- Distinguish a true repeatable framework from a one-off solution.
  Label each finding as one or the other.
- Where a pattern is only partially formed, describe what's there and
  what's missing rather than completing it for me.

OUTPUT FORMAT: Numbered list of candidate frameworks: working name,
files it appears in, 2-3 sentence description, and "fully formed" or
"partial, needs my input."

Prompt 3: Query it like a colleague

This is the promise in the title, and it's a standing habit, not a one-time step. Use in NotebookLM (best citations) or whichever tool holds your archive. Keep it saved; it's the prompt you'll reuse forever.

ROLE: You are my former colleague who worked beside me for 25 years
and has perfect recall of the uploaded archive.

MY QUESTION: [ASK WHAT YOU'D ASK A COLLEAGUE, e.g. What did I
recommend the last three times a client had a warehouse-consolidation
decision? Which proposals won when we were up against a bigger firm,
and what did they have in common? Have I ever solved a
chargeback-dispute process problem, and where?]

CONSTRAINTS:
- Answer only from the uploaded files, and cite the specific file
  (and section, if you can) for every claim.
- If the archive doesn't contain an answer, say "not in the archive"
  rather than reasoning from general knowledge.
- Where my past approaches conflict with each other across the years,
  show me both and note the dates. The evolution is information.

OUTPUT FORMAT: A direct answer, then the citations, then one line on
what related material exists in the archive that I didn't ask about.

Run it the first time on a question you already know the answer to; that calibrates how much to trust it on the questions you don't.

Warning: client-confidential material and NDAs

Before uploading anything from a past job or client engagement, check whether it's covered by an NDA, a confidentiality clause, or your former employer's data policy. Most consulting and corporate work is, and the obligation typically doesn't expire when the engagement ends. Uploading client-identifiable material to any AI tool can breach that agreement even if you never publish the output. When in doubt: strip identifying details, rewrite from memory instead of uploading the original, or leave it out. If you're unsure whether something is covered, treat it as covered.

Where this goes wrong

• You upload everything at once and get generic, unusable summaries. Volume without curation produces mush. Do the gather-and-skip pass first.
• You extract a "framework" that was really one project's specific solution. If you can't explain it to someone outside that project, it isn't a framework yet.
• You treat old client work as automatically yours to reuse. Check your agreements before you check your drive.
• You build the base and never query it. Prompt 3 is the return on the whole afternoon. Put one real question through it the same day you finish uploading, or the archive goes back to being a graveyard with better filenames.

The 2-minute version

Find the drive. Sort files into gather/skip, erring toward skip for anything client-named. OCR the scans. Run the de-duplication prompt on your filename list. Upload the gather pile to NotebookLM and ask it a question you know the answer to. When the extracted frameworks are sitting in front of you, building a Custom GPT from the best one is the natural next afternoon.

---

Want to know where you actually stand with AI? The free AI Readiness Assessment takes 10 minutes and gives you a personalized roadmap.

Weekly, on Tuesdays. The Gen Xcelerator: practical AI intelligence for people who already know how to do the job. Subscribe free.

Experience Is the API. GenXcelerate