Developer Shares AI-Powered Private Error Notebook Tool and Workflow

On November 2, 2026, a developer posted on V2EX sharing and open-sourcing an AI-based private mistake notebook tool along with its supporting workflow. The motivation was personal: their child is in middle school and produces a steady stream of math workbooks and exam papers, while traditional mistake management means copying problems by hand, pasting up torn sheets, and hunting around at review time. Off-the-shelf photo-search apps, the developer noted, only hand back an answer without addressing why the mistake happened or whether it will recur. What they wanted was a private question bank built on their own data: photograph a page to digitize it, personally review which problems deserve entry and label the cause of each error, drill by topic, generate papers from knowledge points, and auto-suggest archiving after three consecutive correct answers so mastered material stops consuming attention.

The technical foundation came from a prior project, the quiz-demo question generator, which shipped a battle-tested quality pipeline: geometric figure numerical checking via geocheck, independent solve-based verification via verify_answers, and vector paper export via export_paper. The mistake notebook reuses this pipeline wholesale, requiring every AI-generated question to pass all three stages with no exceptions.

The full workflow runs as follows. A user uploads photos, several allowed at once. A vision model transcribes multiple questions from the page, capturing the stem in LaTeX, answer options, figure placeholders, printed answer keys, and handwritten marks. The tool then automatically crops the authoritative figure from each detected bounding box, a step that consumes 0 token. A human reviewer then inspects the image in a side-by-side view, with the original photo on the left and a question card on the right rendered live via KaTeX. On confirming entry, the user checks the item into the bank and tags the error cause, using one of three labels: carelessness, not understood, or other, and may choose whether the figure is taken from the original image or drawn by hand. As described in the post, this covers the design rationale, the three-stage verification chain, and every step of the end-to-end workflow.