Notebook Navigator. Zero lines written by hand.

Johan wrote his first Evernote note in 2009. The app was fast, and the idea was simple: everything you want to remember in one place, searchable in seconds. For more than a decade he kept everything there, from research ideas to receipts.

In 2020 Evernote shipped a major rewrite of its clients under the banner "uniting the Evernote apps". For Johan, daily use got slower and heavier after the switch. In 2022 the company was sold to Italy's Bending Spoons, which closed the US office and kept a smaller development team in Italy. In the years after the sale the app gained features, including AI integration. What Johan wanted was the opposite: for it to feel like 2009 again.

So began the search for a replacement, and it came with one non-negotiable requirement. Notes collected over seventeen years must be exportable in their original format, with nothing lost on the way. Apple Notes failed exactly there: export requires third-party services, and the results rarely match what you see in the app. Bear, OneNote, Joplin and UpNote all got honest attempts. Each was missing some of the efficiency of the original Evernote client.

In the end only one option remained. He would have to write the app himself.

Out of the box, Obsidian shows a vault as a file tree: a narrow panel of file names, and that is it. This is a deliberate design choice, the core stays minimal and the rest is left to plugins. But for anyone arriving from Apple Notes or Evernote, the first meeting is a cold shower. No previews, no thumbnails, no dates. Just names.

Notebook Navigator turns that same panel into something anyone coming from Apple Notes, Bear or Evernote recognizes instantly. On the left, a tree of folders and tags. Next to it, a file list with title, text preview, date and thumbnail. Then the open note.

Out of the box

With Notebook Navigator

The list does the heavy lifting. Every note shows its first lines of text and a thumbnail of its first image, whether that image sits in the note, in a PDF, in an SVG or behind a YouTube link. Notes can be shown from all subfolders at once, grouped by date, and pinned to the top. Folders and tags take custom colors and icons. If you live in tags rather than folders, the tag tree can be your primary view, and if you write long-form you can track word counts against targets.

Then there is everything a screenshot cannot show: full keyboard navigation, a built-in calendar for daily notes, support for 21 languages, and a public API for other plugin developers to build on. The whole interface is also tuned for mobile, with swipe gestures and layouts of its own for phones and tablets.

For anyone who has dismissed Obsidian as a tool for technical people, this is the bridge. It looks and works like a regular notes app, while every note stays a plain file you own. The full feature list lives at notebooknavigator.com.

That one person can build, maintain and keep improving a product of this size comes down to how the work is run. Johan does not write the code, he runs the work: sets direction, reviews, checks quality and ships releases. The code itself is written by whichever frontier model is suited to the task at hand. Over fourteen months, that has meant three different models.

Claude Opus 4 built the platform. From May 22, 2025, and through the summer, Opus 4 in Claude Code laid the foundation: the architecture, the two-pane interface, the tag tree, the cache. 538 commits landed in June alone. In late July it hit a wall. The codebase had grown past what the model could hold in its head, just as Anthropic tightened how much context Claude Code could use. New features started breaking other parts of the code, and progress stalled.

GPT-5 and Codex took the baton. OpenAI released GPT-5 in August 2025, and by the end of that month the GPT-5 series in Codex had taken over as Johan's daily driver. It stayed that way for ten months, and feature work flowed: the directory launch, 2.0 with automatic thumbnails, 3.0, the calendar, the API. The limitation was context memory. Improvements that required the model to see large parts of the codebase at once, like sweeping performance work, remained out of reach.

Anthropic Fable 5 got to pay down the performance debt. Three weeks ago Johan switched primary model again, and everything in releases 3.2.0 through 3.2.3 is written by Fable 5. Three weeks produced four releases.

The models also review each other. Ahead of the performance series, Johan commissioned two independent platform reviews of the entire codebase, one from Codex and one from Fable 5, then had them reconciled claim by claim against a pinned revision of the code. Of 296 individual observations, 55 survived validation and became the backlog for the performance work. The ones that fell were archived together with the evidence against them, so the same wrong idea does not resurface in the next review. Letting AI review AI, always with a different model than the one that wrote the code, is a fixed part of the method, in this project exactly as in our client work.

The most common objection to AI-written code is that it will not hold up over time. Notebook Navigator is the counterexample, and the recipe is old-fashioned engineering discipline rather than magic.

Every change passes a battery of automated checks that stop the release at the first error or warning: TypeScript, ESLint with Obsidian's official rule set, Prettier, the test suite and a production build. Security is watched by CodeQL and the OpenSSF Scorecard, with public results. Release notes are structured data in the code, categorized as new, improved, changed or fixed, and shown to users on update.

The documentation is written for two kinds of readers. The architecture documents explain the system to humans. The repo's AI instructions tell the models where the truth lives: which files own which behavior, how the build is verified, and which rules must never be broken. When a new model takes the baton, that file is what makes the handover cheap.

Notebook Navigator is our proof, in production for over a year, that agentic development works at scale: hundreds of thousands of real users, a codebase past a quarter of a million lines, millisecond-level performance requirements, and a maintenance responsibility that never ends.

The most transferable part is the most recent one. When Fable 5 paid down the performance debt, it did so across 270,000 lines of code it had not written itself. That is exactly the situation inside an existing codebase: someone else's architecture, someone else's decisions, and a demand for improvements without breaking anything.

The method behind it is the one we bring to our clients. Clear scope and acceptance criteria. Test suites and builds that refuse to let warnings through. AI models writing the code and other AI models reviewing it. Releases at a high cadence, so every improvement reaches users while it is still fresh. And a human who sets the direction and owns every decision.

The only difference on your side is who owns the result. Notebook Navigator belongs to Johan. What we build together with you, in your teams and on your infrastructure, belongs to you.