A Copy-Paste OCR Prompt for Handwritten Notes and Math

Scriveno Team · September 27, 2026

Modern AI chat models can read handwriting from a photo, and many people now use them instead of classic OCR. The results depend a lot on what you ask for. A bare "transcribe this" tends to get you tidied-up paraphrase, missing lines and equations as loose text.

Below is a prompt that asks for a faithful transcription, keeps math as LaTeX and marks anything the model couldn't read. After that come the variations for other formats, and the point at which a single chat prompt stops being the right tool.

The prompt

Attach a photo of one page and paste this:

Transcribe the handwritten page in this image exactly as written.

Rules:
- Copy the text word for word. Don't summarise, correct, reorder or finish sentences.
- Keep the original line breaks, headings, numbered steps and bullet points.
- Write inline math in $...$ and single displayed equations in \[...\]. Put each multi-line   derivation in one display block using \begin{aligned}...\end{aligned}, preserving the step order and aligning at relation signs where appropriate.” The aligned environment is intended for aligned equation groups. Use \frac, ^, _, \sqrt, \int, \sum and so on,
  never Unicode symbols like ² or √.
- Keep crossed-out text out of the transcription.
- If a word or symbol is unclear, write your best reading followed by [?].
  If it's unreadable, write [illegible].
- For a drawing, diagram or graph, don't redraw it. Write [diagram: short description]
  where it appears.
- Output only the transcription, with no introduction or comments.

Why each rule is there

"Exactly as written" and "don't correct". Chat models are trained to be helpful, so they quietly fix mistakes and smooth the wording. For notes you want what's on the page, including your own errors, so you can spot them.
LaTeX for all math. Plain text can represent math, but it can lose visual structure or become ambiguous. LaTeX marks that structure explicitly.” / “For portable math source, prefer LaTeX notation inside math delimiters—for example, $x^2$ and $\sqrt{x}$.
No Unicode symbols. Unicode symbols such as ² and √ may be treated or rendered differently by different editors; they do not always break.
[?] and [illegible]. Without them the model guesses silently and the guess reads just as confidently as everything else. Marked guesses tell you exactly where to look.
[diagram: ...]. Asked to "include" a diagram, models often invent an ASCII sketch or describe something that isn't there. A placeholder keeps its position without making anything up.
"Output only the transcription". It saves deleting "Sure! Here's your text:" from every page.

Variations

For Markdown (Obsidian, Notion): For Obsidian Markdown, use $...$ for inline math and $$...$$ for display math. For Notion, use $$...$$ to enter inline math and /math to insert a block equation; use only commands supported by KaTeX.” Notion’s equation-entry syntax differs from standard Markdown’s inline and display delimiters.

For chemistry: add "Write chemical formulas and reactions with the mhchem package's \ce{...} command, for example \ce{2H2 + O2 -> 2H2O}." For drawn structures, keep the [diagram: ...] rule; guessing a structure from a sketch is where models go wrong most confidently.

For a full LaTeX file: add "“Wrap the result in a compilable LaTeX document with \documentclass{article} and \usepackage{amsmath} in the preamble. If you also use the chemistry variation, include \usepackage{mhchem}." Then check it compiles with the LaTeX checker, which reports the first error in plain language.

How to check the result

Don't trust a transcription because it looks tidy. Put the photo and the output side by side and check:

Every [?], then the equations: exponents, subscripts, minus signs and limits on sums and integrals are where errors hide.
Symbols that look alike in handwriting: 1, l and |; x and ×; d and ∂; 5 and S; primes and commas.
Rendered math, not just source. Paste the LaTeX into any preview or the checker. A misplaced brace is obvious once rendered and easy to miss in source.

Where a single prompt stops being enough

One page in a chat window works well. It gets harder when:

You have a whole notebook. Some chat tools accept multi-page PDFs or multiple images, but their limits and support for scanned handwriting vary. For a long notebook, process pages in batches if needed, then check that none were shortened or skipped.” OpenAI Docs describe PDF inputs that include page images for vision-capable models, so splitting every PDF into one-page requests isn’t always necessary.
You need a document, not text. If the chat returns LaTeX source, compile it or use a compatible tool to make a typeset PDF. To turn a full LaTeX document into a Word file with editable equations, use a converter such as Pandoc.
Your notes have diagrams. The prompt above can only mark where a diagram was. Getting the actual drawing into your document means cropping it from the original page yourself.
You want to hear it. If your chat app offers read-aloud or voice output, use it. Otherwise, copy the text into a text-to-speech tool.” OpenAI Docs also describe generating spoken audio from text.

That's the gap Scriveno fills. Upload a photo or PDF, and each page is transcribed with math as LaTeX. You get a typeset PDF, a Word document with native, editable equations, Markdown and the LaTeX source, with a rendered preview to check against the page. Hand-drawn diagrams are detected and cropped into the document near where they appeared, and you can have the notes read aloud. Try it on a page of your notes, or see How to Convert Handwritten Notes to LaTeX for the full workflow.

FAQ

Can ChatGPT or other AI models read handwriting?

Yes, current models that accept images can read most legible handwriting. How well depends on the handwriting, the photo and the prompt. Asking for a faithful transcription and marking unclear words, as in the prompt above, makes errors much easier to find.

What's the best prompt for handwritten math?

Ask for LaTeX explicitly, forbid Unicode math symbols, and ask the model to flag unclear symbols rather than guess. Those three rules make the biggest difference for equations.

Why does the model change my wording?

Chat models default to cleaning up text. Say "exactly as written" and "don't summarise or correct" in the prompt, and check the output against the page.

Is it safe to upload my notes to a chat model?

That depends on the service's own data policy, so check it before uploading anything private. For how Scriveno handles uploads, see the Privacy Policy.