Stop Copying Mistakes. Start Learning from Them: The AI Notebook I Use with My Child
The easiest part of a mistake notebook is putting a question into it. The part that often falls apart is opening it a few days later and finding out whether your child can actually solve it now.
There is a question to copy, a diagram to redraw and an answer to organize. Once everything is neatly collected, the next review session starts with another search. If the same question goes wrong again, there may simply be another photograph, with little connecting the two attempts.
I have been using wrong-notebook with my own child for a while, and I keep refining it around this everyday workflow: let the computer handle more recognition, organization and retrieval, leaving the parent and child more attention for understanding and practice.
First, credit where it belongs. This project is a fork of wttwins/wrong-notebook. I am grateful for that foundation. My branch adds features and fixes around our family use, with the changes publicly available. It is already being used, and it will continue to change as new problems appear.
Update (October 28, 2026): this article has been revised for stable release v2026.10.26. It now covers daily review scheduled at 1, 2, 4, 7, 15 and 30 days, free MiniMax account sign-in, the admin setup wizard, per-user daily AI limits, stricter grading and fuller backups. For a closer look at free accounts and daily review, see the follow-up, Free MiniMax Accounts, fnOS, and Daily Review.
Figure 1: Original infographic with a crop of the anonymized application interface. Desktop and mobile compositions are provided separately. This is a project introduction and a discussion of family use; installation details live in the repository documentation.
1. The difficult part comes after collecting a mistake
The purpose of a mistake notebook is straightforward: preserve weak spots so they can be revisited. The difficulty is that all the information around one question can end up scattered.
The photograph is in an album. The explanation is in a conversation. The printed practice sheet is on the desk. What happened last time exists only in a parent’s memory. Preparing a weekend review session starts with a small administrative job.
wrong-notebook brings these pieces together. A question can have its text, original image, explanation, concept tags, difficulty and practice history in one place. Results from variations and generated worksheets can be linked back to the source question.
Figure 2: Real application screenshot from the project’s published, anonymized screenshot set, captured on October 22, 2026. Review-state, detail and statistics figures have been recaptured from current code; other figures retain the published English demonstration set. User identity has been anonymized; question content and interface values have been preserved. English figures use English questions, explanations, tags and animation text in an isolated demonstration database. Its 12 questions, 42 synthetic attempts and estimates do not measure a real student’s outcomes; the built-in concept library remains based on the Chinese K–12 curriculum. Click a screenshot to open it at its original size.
Think of each question as having its own library card. Whether it appears in a review list, a practice variation or a printed worksheet, its record tells you where it came from. You do not have to identify it from scratch every time.
Three questions matter most to me:
- What went wrong, and which concept needs attention?
- Can the child solve a different question using the same idea?
- After this attempt, which questions deserve another look?
A tidy collection is pleasant. Being able to answer these questions gives the collection a reason to be opened again.
2. Understanding an explanation still needs an independent attempt
Watching someone demonstrate a bicycle turn can help. After understanding the movement, however, you still need to hold the handlebars and pedal to discover which parts you cannot yet do yourself.
Reading an explanation and solving a question have a similar relationship. While reading, the path is already visible. Put the answer away and try to find the first step independently, and the remaining difficulties become easier to see.
In their review of learning techniques, Dunlosky and colleagues rated practice testing and distributed practice as having high utility. In everyday language, these involve recalling and attempting things yourself, and spreading practice across time. That research concerns learning methods. It does not evaluate this application, and it cannot support a promise that installing the software will improve a child’s grades.
Figure 3: Original illustration of a suggested family review cycle. The latest version adds a built-in Daily Review that lists questions as they fall due (see Section 6); how much to practice, and how, is still for a parent and child to decide.
The design goal is to make it easier for a saved question to return to practice. AI can offer explanations and prepare variations. A parent checks the material. The child attempts it independently. The result becomes part of the record. Each step is ordinary; connecting them makes the workflow useful.
3. Capture: photographs, pasted images and typed questions
Questions at home arrive in different forms: printed workbooks, images shared by a teacher, digital learning materials and short questions a parent types in. A single input method quickly becomes restrictive.
The project supports image upload, clipboard paste, screen capture and manual text entry. Images can be cropped, with mouse-wheel, pinch and button zoom for small print, and multiple questions can be selected from one image. Text can include Markdown and LaTeX mathematics.
Figure 4: Real home screen. Questions from different sources can enter the same organization workflow.
AI attempts to recognize the question, answer, explanation, the student’s incorrect working and the reason for the mistake. It can also suggest a subject, concept tags and difficulty. For illustrated questions, it can generate a redrawn SVG diagram for display and printing.
OCR is simply the computer reading the writing in a photograph. A missed minus sign or a dropped condition can send the rest of the solution in the wrong direction. The application therefore preserves the original image and offers a review-before-import option, allowing a person to check the recognized material before saving it.
Figure 5: Real question detail screen. The question, its classification and follow-up actions are kept together, with an option to consult the original image.
Suppose a question says “at least” and recognition turns it into “at most.” That is like copying one character incorrectly in a delivery address: careful packaging will not send the parcel to the right place. Keeping the original and making corrections convenient helps catch the problem early.
Mathematical formatting has its own practical details. Fractions, roots and equations are rendered as mathematics. If the LaTeX syntax is broken, the application offers an AI repair action. Regenerating the complete solution is a separate operation, with confirmation before replacement, so a formatting correction and a new solution can be treated differently.
4. Concept tags and animation offer more ways into a question
When preparing a review, searching by concept is often more direct than searching by date. You might want to revisit angles without remembering when a particular photograph was taken.
The standard concept library is organized by subject, school stage, grade and chapter. Custom tags are also supported. AI selects tags from the appropriate subject and stage, and a parent can review and adjust them. The concept library now lives in the Statistics area, trimming the navigation from six items to five: one tab browses the library and manages custom tags, another shows how often each tag is used.
Figure 6: Real concept library. Tags work like labels on bookshelves, making it easier to find material for a particular weak spot.
Some explanations work well as text. Other relationships become easier to discuss when they are drawn. The project can ask AI to generate a step-by-step SVG animation, with pause, seek, replay and full-screen controls. Interactive GeoGebra demonstrations can also be enabled for mathematics and physics questions.
An SVG is a picture described through drawing instructions: lines, shapes and text each have a place. Those shapes remain clear when enlarged. Revealing them in stages can break a solution into smaller pieces.
Figure 7: Real animation player showing an existing question explanation. This English demonstration also uses English text within the diagram. Mathematical claims should still be checked against the question.
One useful moment may come when a parent pauses the animation and asks, “What happens next, and why?” The child offers a prediction before playback continues. A presentation becomes something to talk through together.
Smooth animation and attractive colors do not establish that a proof is correct. Conditions, geometric relationships and answers still need checking. Families can also turn animation off and keep a simpler combination of text and paper practice.
5. Everyday details drive the next iteration
A tool used repeatedly at home develops increasingly specific requirements. “Import a question” soon becomes “What if we photographed it twice?”, “What if recognition needs a correction?”, or “Can the paper level follow the difficulty we already selected?”
Several current behaviors address those practical details.
| Everyday friction | Current behavior |
|---|---|
| The same question is photographed again | Text similarity is used to detect duplicates; a match keeps the original and records another incorrect attempt |
| Families need different duplicate-detection sensitivity | The threshold is saved in user preferences; the default is 90% |
| AI recognition needs checking before it is saved | Review before import can be enabled, with an editable correction screen |
| Difficulty and A/B paper level repeat related choices | When no manual choice is supplied, 1–3 stars default to A and 4–5 stars to B |
| AI work takes time | Background-task progress is available, and work can continue after leaving the current page |
| Yesterday’s variation should still be there today | The most recently generated variation and its solution are saved for each question and shown again on return |
| A fresh installation needs a starting point | On first sign-in, the admin goes through a setup wizard: change the default password, set up an AI provider and create the first notebook |
Figure 8: Real preferences screen. Options that affect the organization workflow can be adjusted by the user.
Recording the repeated mistake is especially useful. A question database rarely needs two identical copies, but a review history needs to preserve the fact that the same question went wrong again.
Imagine a tap that starts leaking for a second time. The maintenance log needs another entry, even though the tap itself was registered long ago. A repeated problem contains information about what happened after the first repair.
Similarity detection also needs judgment. Two questions can look almost identical while testing a meaningful change in conditions. The application allows threshold adjustment and question comparison so a family can choose a suitable level of sensitivity.
These details may not make dramatic feature announcements. They do affect whether a tool remains comfortable enough to keep using. This is the kind of iteration I intend to continue as real use reveals more friction.
6. Practice variations turn a recent explanation into a visible history
The project can generate variations around the same concept as a source question, with a choice of difficulty. After an attempt, a result can be recorded. Worksheet results also return to the originating question.
Grading has also become stricter. An answer that matches exactly under local comparison is marked correct; anything else goes to AI for judgment, reducing the chance that a partly correct answer is accepted as fully correct.
A parent can then examine the sequence: an initial mistake, a later variation and another encounter on a worksheet. That history provides more context for deciding what to revisit next.
Figure 9: Real practice history. The percentage is an application estimate for this question, rather than an exam score or evidence of the project’s educational effectiveness.
The current implementation first asks AI to estimate mastery from the attempt history, with a rule-based fallback if assessment fails. The latest version derives review status automatically from recorded results and the estimate.
Three review states follow the recorded evidence
- To Review: no practice result has been recorded yet.
- Reviewing: results exist, but the estimate is below 50%, or no estimate is available yet.
- Mastered: results exist and the estimate is at least 50%.
Think of three homework baskets: material waiting to be tackled, material being reinforced, and material that currently meets the application’s classification rule. The 50% boundary organizes a review queue; a child’s independent work still matters more than the label.
The manual mastery toggle has been removed. Lists, filters, statistics and worksheet selection use the same rule so the different screens stay consistent.
Figure 10: Original illustration of the current three-state rule. Classification helps decide what to revisit next.
Think of the number as a weather forecast: it can help with plans, but you still look outside. Sparse records, closely repeated questions or substantial parental prompting can create a gap between an apparently correct answer and independent understanding.
I suggest using the percentage as a reason to examine the record. Which attempt went wrong recently? Did a changed condition cause trouble? Could the child explain the method without a prompt? The strongest evidence still comes from the child’s work and explanation.
Daily Review lists the questions that are due
In the earlier version, a parent had to remember when to revisit a question. The new Daily Review schedules each question at Ebbinghaus-style intervals of 1, 2, 4, 7, 15 and 30 days. A correct answer moves it to the next interval, a wrong answer sends it back to the first, and six consecutive correct answers take it off the schedule. The home page shows how many questions are due, and the Review page lists questions due today or earlier, oldest first.
Figure 11: Real Daily Review list, reused from this blog’s anonymized follow-up screenshots. Each question shows its due date and current round.
Think of a library due date. It reminds you when the book should come back, but only opening the book shows whether you understood it. Daily Review answers “Which questions are due today?” while the three review states describe what the record suggests so far. Neither replaces the other, and a Mastered question can still fall due. The interval table is a general starting point; if a child is tired or too many questions are due, simply do fewer.
7. Printing matters because family learning also happens on paper
Long arithmetic, auxiliary lines and a complete written argument all need space. A family study tool has to support those moments away from the screen.
The project can generate printable worksheets using selected notebooks, concepts and mastery conditions, with variations based on source questions. It can also print filtered mistakes directly, optionally including answers, explanations, tags or original images. A new worksheet shows as Generated, and only becomes Completed once results have been entered, so unfinished sheets on the desk are easy to spot.
Figure 12: Real worksheet screen. Source questions and later attempt results remain connected.
Printed content carries a source ID, making it possible to locate the original record when entering results afterwards. The ID works like a parcel collection code: even after the paper leaves the computer, there is a route back to its record.
Figure 13: Real print preview. HTML can be downloaded, or the browser can print the sheet or save it as a PDF.
A possible family routine is to choose a few concepts that need reinforcement, print a small set and let the child work independently. Check the answers together afterwards and record the useful results. The amount and frequency should follow the child’s needs; easy question generation is no reason to produce an endless pile of exercises.
The application prepares materials and preserves records. The thinking on the page still belongs to the child.
8. Statistics help choose the next step
The application separates mistake statistics from practice statistics. The former includes question counts, mastery rate, subject distribution and monthly trends. The latter includes attempt counts, accuracy and active months.
Figure 14: Real mistake statistics. The numbers are the values in that interface snapshot and illustrate the feature.
Figure 15: Real practice statistics. These are usage records, not the results of a controlled study of learning outcomes.
The two views answer different questions. How many mistakes have been collected and how many attempts have actually been made are different measures. Accuracy also needs context when question difficulty changes.
For me, the useful question is, “What deserves another look this week?” A statistics page can act like a sign at a crossroads, helping choose a direction. It should not become a permanent label attached to a child.
9. Self-hosting, model choice and privacy need concrete boundaries
The application uses Next.js, React and TypeScript, with Prisma and SQLite for data. It supports multiple users and data import/export. Documentation and release channels cover Docker, npm, desktop applications, FnOS and the two Android modes.
You can set the technical names aside for a moment. Self-hosting is like keeping a filing cabinet in a place you manage. You have more control over backups, access and upgrades, along with responsibility for maintaining them.
When several people share one instance, the latest version adds finer administration. New sign-ups are closed by default and can be changed with ALLOW_REGISTRATION. An admin signing in with the default password must change it first. Admins can also set a daily AI limit per user and see each user’s usage for the day. Export and import now include user preferences, random-mistake worksheets with their attempt records, saved practice and mastery estimates, making backups before a device change more complete.
AI integrations include Gemini, OpenAI, Azure OpenAI, Anthropic, MiniMax, Zhipu GLM, DeepSeek and Kimi. Image-input support, pricing and output quality need to be checked for the chosen channel and model. An integration in the list does not mean every model can perform every task.
The new MiniMax Code OAuth channel lowers the barrier to starting: sign in with a free MiniMax Code account in the browser to use MiniMax-M3.1-Flash without buying an API key. The app checks in daily to collect credits, and AI features spend the account’s credits. This channel uses the MiniMax Code client’s sign-in rather than an official open API, so the platform may change its rules, and free credits do run out. The follow-up article covers those limits.
Figure 16: Original data-boundary illustration. Self-hosting does not mean that every AI operation runs locally.
When an external model API is used for recognition or generation, relevant question content is sent to that provider. Keeping the database under your control does not automatically keep every processing step on the device. Cropping away names, school details and class information before upload is a useful habit.
The project also provides an Openclaw batch-image-upload endpoint for parents who already use agents in their workflows. That is an existing import capability. It should not be read as a claim of autonomous teaching, complete tutoring or a finished MCP product.
My preferred division of work is straightforward: an agent helps prepare and move materials, a parent checks the content and results, and the child does the learning. Credentials and identifiable schoolwork do not belong in public screenshots or issues simply to demonstrate automation.
10. Android phones and tablets: choosing between four APKs
A kitchen in your pocket, or a client for the kitchen at home
The project now offers standalone and client Android applications, each with Phone and Tablet variants.
The standalone app is like carrying a small kitchen with you: the application server and its filing cabinet run on the device. It suits use on one Android device without deploying a separate server. The client is more like taking a menu to the same kitchen: it supplies the interface while an existing server holds the questions and records, allowing several devices to access the same collection.
Standalone packages include Node.js and the application server and require Android 8.0 or later on a 64-bit ARM device. Client packages require Android 7.0 or later and connect to an existing server. Phone variants stay in portrait orientation; Tablet variants rotate. The two modes have different package IDs and can be installed together.
Figure 17: Original mode comparison, not an APK device screenshot. Application screenshots show real responsive browser interfaces; this article does not report Android device installation tests.
Let the operating system handle capture, saving and printing
Both Android modes add native capabilities around the WebView: camera, gallery and file selection for uploads; system Save As for downloads and backups; system printing, including PDF output; and the system browser for external sites.
Standalone data lives in the app’s private directory and is deleted on uninstall. System backup is disabled for both apps, so export your material before uninstalling or changing devices. With a client, you also need to maintain the server’s data backups.
At the time of checking, stable release v2026.10.26 supplied all four APKs (standalone and client, each in Phone and Tablet forms) and included the three review states, Daily Review and setup wizard described here. Consult each Release to see whether later commits have reached its binaries.
PWA caching covers interface assets
Phones and tablets can also add the PWA to their home screen. The improved Service Worker caches scripts, styles and icons and provides an offline notice page. Questions and statistics still come from the server; this does not create an offline question database.
It is like keeping the cover and interface pieces ready while fetching the contents from the filing cabinet. Standalone Android moves the server onto the device, but it does not put a language model there. Recognition, explanations and variations using an external model still require the relevant network and API access.
11. Ongoing maintenance includes permissions and security
Boundaries for imported drawings and native actions
An SVG diagram is like a note carrying drawing instructions. Current code repeatedly sanitizes dangerous fragments to reduce the risk of nested content reconstructing a script. Android’s Bridge between web content and native actions works like a service window: it performs these operations only for pages belonging to the configured application server. The Openclaw batch-upload endpoint also accepts only JPEG and PNG images.
Make checking part of maintenance
Recent work moves the default Node.js to version 24 and the desktop app to Electron 45, with security updates for Next.js and the mathematics and proxy dependencies. CI now includes Dependabot dependency checks, CodeQL scanning, zero-warning ESLint and a minimum test-coverage threshold. These help identify and address problems over time, alongside the user’s own content review, backups and access management.
The repository now includes an MIT license, contribution and security-feedback guidance, a formal CHANGELOG and a bilingual wiki. Use the changelog for stable releases and Releases plus commit history for preview builds and current main. That distinction makes it easier to see what has reached an installation package.
12. Who might find it useful, and common questions
If your family already photographs questions, organizes mistakes and revisits them periodically, this project may be useful. It is particularly relevant if preparing a review repeatedly involves finding the same materials again, and you are willing to check AI output and manage your own records.
Families seeking a complete curriculum, guaranteed answer quality or someone to guide every study session will need services beyond what this evolving project provides.
Can I explore it without writing code?
Start with the project introduction or the online demo. The demo uses a shared account: other visitors can see its data, so use only non-private example material. Login information is in the repository README.
Can I start without a paid API key?
Try the MiniMax Code OAuth channel: sign in with a free MiniMax Code account in the settings, a step the admin setup wizard also suggests. Whether the free credits are enough depends on how many questions you add and which features you use; a paid channel remains an option when you need steadier capacity.
Where should I install it for long-term use?
For long-term family use, consult Releases and the installation guide. The assets actually available depend on the selected release.
Does everything work offline?
The PWA caches interface assets; question and statistics requests still go to the server. Standalone Android runs that server locally, while a client connects to an existing one. Neither mode bundles a language model. External model calls for recognition, explanations and variations still need their network and API access.
Can AI explain a question incorrectly?
Yes. Original images, review before import, editing and regeneration make corrections easier. The same scrutiny applies to animations, formulas and mastery estimates. For disputed questions, return to the textbook, the teacher’s requirements and a solution that can be checked step by step.
Could it become an answer-copying tool?
How the family uses it matters. A parent can organize and print the material, let the child attempt it independently and discuss the explanation afterwards. The software supplies materials and records; the sequence of use remains something adults and children agree on together.
Can I modify and redistribute it?
The repository now uses the MIT license. Follow its terms and retain copyright and license notices. Bundled third-party components keep their own licenses; see the contribution guidance for sharing improvements.
Will the project keep changing?
Yes. I still use it with my child and intend to keep responding to problems found through use. Open-source maintenance has a practical pace, though. Commits, issues and releases are the record of what has actually shipped; I am not promising a fixed update timetable.
If your mistake notebook is carefully filled but difficult to revisit
Have a look at margrop/wrong-notebook. Browse the screenshots and demo, then decide whether it fits your family’s habits. If it looks useful, a Star can help you find it again. Specific reports about awkward steps are welcome as issues.
Helpful feedback describes what you tried, what you expected and what happened. Remove child identity, school information, account details and access addresses before attaching images. Contributions can start with documentation, interface translation, tests or a small practical fix.
I hope it becomes an increasingly comfortable family tool: questions are easier to preserve, review is easier to begin and each attempt is easier to see. The most valuable part of learning still happens between people: a child willing to try again, and a parent willing to help think a problem through.
Sources and image notes
Updated on October 28, 2026 against public main 610af18 and stable release v2026.10.26, including the implementation, documentation, changelog and actual Release assets. Source and packaged-release progress can differ.
- Project repository and README: features, demo and entry points.
- Upstream: wttwins/wrong-notebook: the foundation of this fork, with thanks to its author and contributors.
- CHANGELOG: features and fixes in each stable release.
- Multi-platform installation guide: package forms and runtime details.
- Screenshot provenance: capture date, anonymization and bilingual materials.
- Mastery evaluation implementation: AI assessment and rule fallback; see the three-state implementation for the current classification rule, and the Daily Review implementation for the intervals.
- Follow-up: free MiniMax accounts, Daily Review and daily AI limits: credit limits, automatic check-in and the grading flow in detail.
- Dunlosky et al., Improving Students’ Learning With Effective Learning Techniques: the 2013 review; this article cites its conclusions about methods, not this application’s effectiveness.
This article contains 12 real application screenshots, including the Daily Review screenshot reused from this blog’s follow-up, and five original diagrams. Unchanged screens use the published English demonstration set; review-state, question-detail and statistics screens have been recaptured from current code. Screenshot pixels were converted losslessly to WebP. The five original infographics have been redesigned with separate Chinese and English typography and distinct desktop and mobile compositions. Text and graphics have been checked in the rendered article.
For other tools that can complement paper-based learning, see this blog’s PhET interactive science lab introduction and 41 online logic puzzles.