TL;DR
Thorsten Meyer says an AI coding agent produced a working first version of his private PDF editor, Docestra, on Nov. 6, 2026. Over the next six days, tests on real files exposed save, encryption and redaction problems; the tool reached version 1.5.1 on Nov. 12, while version 2.0 remains in development.
Thorsten Meyer says an AI coding agent built the first working version of his PDF editor, Docestra, on Nov. 6, 2026, but tests on real documents soon exposed problems that a short demonstration had not. His project log tracks six days of fixes through version 1.5.1 and describes a broader challenge for AI-built software: a tool can work quickly on sample files while still failing on documents people rely on.
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Meyer describes Docestra as a local web app written in TypeScript that runs in a browser on his Mac, with files kept on the machine. He says AI coding agents wrote the code, with Anthropic’s Claude Code running Claude Opus, while OpenAI’s Codex reviewed it from the start. The initial version could open and search PDFs, add highlights and other annotations, fill forms, sign, rearrange pages and save files.
A robustness run over 930 real PDFs found that saving some files could produce a broken cross-reference table, causing markups to disappear from the saved document. Meyer says he changed the save behavior to avoid compressed object streams. His tests also included 69 encrypted PDFs; according to his account, 52 had author-set restrictions that disallowed annotations. Because the PDF library could not re-encrypt edited files, Docestra was changed to require confirmation and save an unprotected copy without overwriting the original. Meyer reports that all 69 test files then unlocked.
The project also distinguished visual cover-up from actual redaction. The original “white-out” tool concealed content visually but left underlying text in the PDF. Meyer says a first implementation of real redaction received 14 issues in an independent review, followed by 13 in a second round. He then switched to a simpler design and reports zero leaks across 917 PDFs. By Nov. 12, version 1.5.1 had 745 unit tests and 32 browser test suites, according to the project records he cites.
The Work Behind a Reliable PDF Tool
The project account illustrates why a convincing first version does not establish that software is dependable. PDF files can behave differently depending on how they were generated, what restrictions they carry and whether sensitive information is truly removed. The reported save bug could make annotations disappear, while a visual white-out could leave private text recoverable from the file.
For readers building or adopting AI-generated tools, the practical point is that reliability depends on testing representative files and making careful choices when the software’s libraries cannot preserve expected protections. Meyer’s account is a single developer’s project report, not an independent evaluation of AI coding tools or a measured productivity study. He says he has not measured time or cost savings.
Products Worth Considering
Samsung Odyssey OLED G5 27-Inch QHD Gaming Monitor
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
As an affiliate, we earn on qualifying purchases.
Six Days of Docestra Changes
Meyer’s timeline begins with a one-sentence brief on Nov. 6 and a first working version committed that evening at 21:47 Central European Time. The 930-file robustness run took place the same day. On Nov. 7, he changed how encrypted files were handled. On Nov. 8, he developed the redaction feature and sought review of its design.
The account says version 1.5.1 was reached on Nov. 12, followed by work on version 2.0 beginning Nov. 13. Meyer describes that next version as an effort to build one shared PDF engine for AI agents. It remains unfinished. Docestra is a private project and is not available to download, so readers cannot independently inspect or reproduce the reported tests from the article.
“Vibe coding gets you a working tool quickly. Turning it into a tool you can rely on is the real work.”
— Thorsten Meyer
What the Project Log Cannot Show
The reported test results come from Meyer’s own project records; the article does not provide an external audit or enough detail to independently reproduce the tests. It is also unclear how Docestra performs across the wider range of PDF software, operating systems and document types beyond the files described. The reported zero leaks applies to the 917 PDFs tested, not every possible document.
Meyer says he has not measured the project’s time or cost savings. Version 2.0 is still in development, and the account does not specify when it will be finished or what features it will include beyond a shared PDF engine for AI agents.
Version 2.0 Remains in Development
Meyer says development of Docestra 2.0 began Nov. 13, with the goal of creating one shared PDF engine for the project’s AI agents. No release date is given, and the version is not finished. The next milestone is therefore completion and further testing of that work; the account does not provide a schedule or release plan.
Key Questions
What is Docestra?
It is Meyer’s private PDF editing app, described as a local TypeScript web app that runs in a browser on his Mac. He says files stay on the machine; the app is not available to download.
What did the 930-PDF test find?
Meyer says it found a save issue affecting a real PDF made by a common generator. The broken cross-reference table could cause markups to disappear from the saved file.
Did the white-out tool remove the text beneath it?
No. Meyer says the first tool covered content visually while leaving the underlying text in the PDF. He later built a separate redaction feature.
Is Docestra 2.0 available?
No. Meyer says version 2.0 entered development on Nov. 13, 2026, and remains unfinished. He has not announced a release date.
Source: Thorsten Meyer AI
