Before
Their reading life lived in Apple Notes and a review template of their own. Raw thoughts came first, typed on an iPhone mid-chapter, and turning them into a structured, sourced analysis was slow, manual work. Most AI writing tools offered to write for them, which was exactly what they did not want.
“Capture first, organise later. No AI interferes while reading itself — raw, unfiltered thoughts come first, before anything gets structured.”
In their own words
The turning point
Instead of jumping into code, they wrote a fifteen-page product specification first, with a "Product Architect" role pushing back on scope and ideas. Then they built with Claude Code in tight loops: build, test on real content, find real bugs, fix, repeat. They applied their own rule, "the reader decides", to the building process itself.
“The starting point wasn't a generic product idea but my own real workflow (notes in Apple Notes, my own review template) — the app was built around that process, not the other way around.”
In their own words
After
They now have a working web app of about 2,100 lines of Python that carries a book from loose notes to a finished 19-section review, with a character relationship diagram and export in German, English and French. They tested it end to end on Christa Wolf's Kein Ort. Nirgends using their own reading notes, and live testing led them to add a version history and undo feature.
What it lets people do
Readers can capture thoughts freely, then see them sorted, analysed and deepened through research they trigger themselves, with every source checked. Each piece of text is labelled by origin, new AI drafts appear side by side with the reader's own words, and direct quotations are never translated or paraphrased.
What they learned
Real use teaches what a spec cannot, like discovering that autosave without a safety net is riskier than saving by hand. They also found that AI drafts improve noticeably when given structured context such as notes, characters and research instead of a generic prompt.
“Real usage reveals things a spec can't — e.g. that autosave without a safety net is riskier than manual saving.”
In their own words
Literary Research Studio was built during AI Mastery.
Read the full submission
Literary Research Studio
A personal research environment that turns loose reading notes into structured, source-backed literary analysis — with AI as a disciplined research assistant, never a ghostwriter.
The Pitch
What it does: The app supports the whole journey of serious reading: capture free-form notes (typed on an iPhone while reading) → automatic categorisation → literary analysis, first from the reader's own notes and later deepened through research → targeted, source-verified internet research → a finished, structured review document with a character relationship diagram and multilingual export.
What makes it special:
Capture first, organise later. No AI interferes while reading itself — raw, unfiltered thoughts come first, before anything gets structured.
Radical transparency. Every piece of text visibly carries its origin: from the reader's own notes, from verified research, from AI general knowledge, or self-written.
Nothing is silently overwritten. Every new AI draft appears as a side-by-side comparison next to the reader's own text; the reader decides. A restore mechanism safeguards every version.
No invented facts. Research is always reader-triggered, and every source is live-verified before being shown as reliable.
Quotes stay quotes. Direct quotations are never automatically translated or paraphrased, even in multilingual export.
Analysis in the reader's own language. Each project's output language is chosen freely — French, German, or English — and the entire analysis, research, and review are generated in that language, quotations aside.
Initial and Final analysis kept separate, mirroring the real scholarly process — the unbiased first reading, then the deepened reading after research.
Most AI writing tools want to write for the user. Literary Research Studio does the opposite: it treats AI as a disciplined research assistant, with complete origin-labelling, a hard rule against fabricated facts, and a workflow modelled on the real process of serious reading — not a generic text-generator template.
Achievements
Built a fully functional web app (~2,100 lines of Python, 19 templates) — from product spec to running app: note capture, automatic categorisation, two-phase analysis, character relationship diagram, source-backed research, a 19-section structured review document, multilingual export (DE/EN/FR).
Live-tested end-to-end on a real book (Christa Wolf, Kein Ort. Nirgends) using my own real reading notes — not a synthetic test dataset.
Found and fixed real bugs during live testing rather than in theory; directly led to building a version-history/undo feature that didn't exist before.
Verified end-to-end with a real Anthropic API key: AI drafts (discussion questions, book-club summary, analysis, translation) genuinely work — including the rule that direct quotations are never translated.
Tools
Python 3.9 / Flask (backend), SQLite with FTS5 (multilingual full-text search)
Vanilla JavaScript (deliberately no framework), Cytoscape.js for the relationship diagram
python-docx for Word export
Anthropic Claude (claude-sonnet-5) via the Anthropic API — structured JSON outputs, web search tool for sourced research
Open Library API for book metadata
Built entirely with Claude Code as the development tool
Methodology
The starting point wasn't a generic product idea but my own real workflow (notes in Apple Notes, my own review template) — the app was built around that process, not the other way around.
A complete product specification (~15 pages) came first, with a "Product Architect" role deliberately challenging scope and ideas before a single line of code was written.
Iterative live-testing as the core method: build → test on real content → find real bugs → fix → repeat — not a one-shot build.
The principle "the reader decides" was applied to the development process itself: every AI-generated feature was checked against my own judgment.
Learnings
Real usage reveals things a spec can't — e.g. that autosave without a safety net is riskier than manual saving.
Labelling the origin of text only becomes truly important once you're evaluating real AI output against your own notes, not in theory.
Platform quirks (macOS sandbox restrictions on protected folders) can look like app bugs but aren't — important to tell apart early.
AI drafts noticeably improve when given structured context (notes + characters + research) instead of a generic prompt.
Relevant URLs
Source code: https://github.com/ujost-web/literary-research-studio
Written 9 August 2026· Submitted by the storyteller
Every story is written by a Mindvalley student or by someone our teachers have worked with. How we collect and check stories
