Denys Shkola
EmailCV (PDF)

Project 10AI engineering

Math-notes agent skills

Two Agent Skills for Claude that turn mathematical notes into study material: an offline dependency map of every numbered result in a LaTeX document, and one-to-one LaTeX transcription of handwritten lecture PDFs, with deterministic scripts and fail-closed gates around the model steps.

Code on GitHubLive demo
Status
Public repo · MIT · live demo
Role
Solo: design, skills, scripts and tests
Period
Aug 2026
Stack
JavaScript (Node) · Python · MathJax · LaTeX · Agent Skills

The problem

A language model that has read a section can write a precise statement, the key steps of a proof and why one result needs another. It is unreliable at the mechanical parts: matching the numbering of the PDF, covering every item, producing LaTeX that renders, and admitting when a proof is not in the source. Those failures don’t crash anything; they produce a document that looks finished and is wrong.

What I built

  • latex-mindmap turns LaTeX notes, a textbook or a thesis into one offline HTML file: every numbered definition, theorem, lemma and example as a node in a dependency map, each edge with a written reason, plus statements, proof outlines and a searchable cheat sheet.
  • handwritten-notes-to-latex transcribes handwritten lecture PDFs into a LaTeX document that matches them one to one: same numbering, titles, statements and proofs, in a fixed house style.

Each skill gives the mechanical work to deterministic scripts and puts fail-closed gates between the model’s steps.

How it works

  • Validation gate. Between LLM extraction and the build, a merge step refuses missing or extra items, duplicate ids, dangling edges, dependency cycles, unbalanced $, braces or \left–\right, a bare < and LaTeX in labels.
  • Render gate. MathJax’s noerrors and noundefined packages are removed, so an undefined macro aborts the build and names the macro instead of shipping red text.
  • Provenance. Reconstructed proofs and outlines are labeled “reconstructed (not in the notes)”.
  • Handwriting audits. Page tiling for a vision model, verbatim lecturer numbering, and audits for duplicates, backwards numbering, gaps and coverage.

Results

count
18numbered results in the demo map, with 27 dependency edges that each carry a written reason
evidence
14refusal cases tested for the validation gate
count
69automated tests (46 pytest + 23 node:test)

The live demo maps my BSc thesis: 18 numbered results and 27 dependency edges in one 0.39 MB offline HTML file. CI rebuilds it byte for byte from the committed sources, and a headless audit passes 17 of 17 checks.