Paper notebook, OBTF, and LLMs: connecting analog and digital notes

Updated: October 27, 2026

A paper notebook is good for capturing thoughts, OBTF is good for a durable searchable archive, and an LLM is good for extracting meaning. Together they close the loop: capture → store → find → understand.

Why three tools

Each layer has a job. Paper does not distract with notifications and works well for handwriting, diagrams, arrows, and slow thinking. OBTF keeps everything in one plain text file that can be searched, versioned, encrypted, and passed to models. An LLM helps read a large archive, find connections, and draft conclusions.

The key rule: the LLM is not the source of truth. The paper notebook and OBTF are. A model can be wrong, confuse dates, and invent facts, so its answers must be checked against the original entries.

• Paper — capture, thinking, sketches, quick lists.
• OBTF — memory, structure, search, backups.
• LLM — reading, summarization, questions, draft conclusions.

How to keep a paper notebook

The simpler the rules, the more likely they survive. A few conventions are enough:

• Number the pages and put the date in YYYY-MM-DD format.
• Write in one direction — for example, top to bottom within a day.
• Use markers: TODO:, DONE:, Idea:, ?, !, Quote:.
• Leave margins for tags and notes made during digitization.
• Reference other pages: see p. 37.
• Keep a simple topic index at the front or back.

Do not rewrite the notebook neatly. The goal is to preserve the thought and context, not to produce a perfect document. Digitization and cleanup happen later, in OBTF.

Digitizing: OCR or manual

Scan or photograph whole pages: this preserves the source even if OCR fails. Use your phone camera, Apple Notes with Live Text, Google Lens, Microsoft Lens, or Adobe Scan. For batch processing, use PaddleOCR or Tesseract; for handwriting, try MyScript, Nebo, or Apple's built-in recognition.

Handwriting OCR makes more mistakes than printed text, especially with diagrams, arrows, and corrections. So:

• transcribe important entries manually;
• OCR bulk lists, but review the result;
• mark quality: OCR: high, OCR: medium, transcribed;
• keep scans in attachments/notebook-2026/ and link to them from OBTF.

Use predictable file names: notebook-2026-p042.jpg, notebook-2026-weekly.pdf. This makes search and backups easier.

The OBTF entry format

Add new blocks at the top, like any OBTF entry. One block per page or per day. Example:

──────────────────────────────
2026-10-27 Mon · paper p.42
──────────────────────────────
Source: notebook-2026, p.42
Scan: attachments/notebook-2026/p042.jpg
Tags: #paper #idea #product
OCR: high

TODO: test the onboarding hypothesis
Idea: reduce the first screen to one action
Quote: "complexity is a tax on attention"

Note:
OCR text after manual cleanup. Diagrams and arrows are described
in words where their meaning would otherwise be lost.

Link: see p. 37, note on retention

The key elements: date, source, scan link, tags, OCR quality, and task markers. If OCR produced garbage, do not paste all of it: a short summary plus a scan link is better than noise in the main file.

Do not mix verified text and raw OCR without labels. A month later it will be impossible to tell what is a fact, what is recognition output, and what is a guess.

LLMs: local and external

First decide what matters more: privacy or model quality. External services such as ChatGPT, Claude, and Gemini are usually stronger and easier, but they require uploading text to someone else's server. Local models through Ollama, LM Studio, llama.cpp, or Jan run on your computer and keep data local, but require hardware and produce more modest quality.

For small files, search and paste the context manually. First find the relevant entries:

rg -n "TODO|Idea|paper p\." everything.txt
grep -n "2026-10" everything.txt | head -40

Then give the model only the relevant fragments. For large archives, RAG makes sense: split OBTF by date and heading, build embeddings, store them in Chroma, LanceDB, or sqlite-vec, and return top-k fragments with date, page, and scan link. Good embedding models include nomic-embed-text and bge-m3.

A local start with Ollama:

ollama pull llama3.1
ollama run llama3.1
ollama pull nomic-embed-text

For an external model, upload only to a service you trust, disable training on your data if the setting exists, and never send passwords, keys, or third-party personal data. For sensitive notes, use a local model or encryption.

Prompts that work

Ask the model to cite dates and pages, separate facts from conclusions, and admit when data is missing. Examples:

• "Find open TODOs in paper notes from the last 30 days. Return the date, page, and a short quote."
• "Summarize the themes of the week. Separate decisions, open questions, ideas, and tasks."
• "Compare paper notes with digital entries. Find contradictions and repeated ideas. Cite dates."
• "Build a plan for next week only from open tasks. Do not add anything that is not in the text."
• "Extract quotes worth keeping, with page and date."

A good system prompt: "Answer only from the provided text. If the data is missing, say so. Cite the date and page for every claim. Do not invent facts or links."

The working loop

1. Day. Write in the paper notebook without worrying about structure.
2. Evening. Five minutes: photograph the page, move the important parts into OBTF, add Source and Scan.
3. Week. OCR longer pages, clean the text, close finished tasks, update tags.
4. Review. Ask the LLM about themes, decisions, and open questions. Check its answers against OBTF and the scans.
5. Month. Archive scans, make a backup, rebuild embeddings if needed.

Keep the loop short. If digitization takes longer than writing, simplify the format: date, page, three markers, and a scan link.

Privacy and security

Paper does not protect against theft or loss, and scans may contain more than you remember: names, addresses, numbers, diagrams. Keep scans in an encrypted location, sync only ciphertext, and do not send sensitive pages to external models.

Use encryption and backups for OBTF. If the file syncs, set up secure sync and two-factor authentication. Detailed commands are in the guides on age, GPG, and VeraCrypt and syncing one text file.

Limitations

• OCR fails on handwriting, diagrams, and corrections.
• An LLM can invent a fact, date, or quote.
• The context window is limited: a large OBTF does not fit in one prompt.
• Embeddings go stale after major edits.
• A paper notebook can be lost, and scans can be damaged.
• An overly complex digitization process kills the habit.

So: keep the scans, verify answers against the source, make backups, and do not let an LLM write to OBTF without human review.

Checklist

• A notebook with page numbers and dates.
• Clear markers: TODO, DONE, Idea, question.
• Page scans stored in a predictable folder.
• OBTF entries include Source, Scan, and OCR quality.
• Raw OCR is not mixed with verified text.
• A local or external LLM is chosen based on privacy.
• Prompts require citations and dates.
• Model answers are checked against OBTF and scans.
• Backups exist and, when needed, encryption is enabled.
• A short weekly review and task cleanup.

To go deeper on the AI layer, start with OBTF as a knowledge base for AI. If you are still deciding whether OBTF is right for you, read the comparison with Notion and Obsidian.