🔍 Read the full analysis: 55 Lessons? No—The Real Work Was Making My AI-Built PDF Editor Reliable on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Thorsten Meyer says an AI coding agent built a first working version of his private PDF editor, Docestra, on November 6, 2026. Over six days, tests on real PDFs and an independent AI review exposed save, encryption and redaction problems; Meyer says version 1.5.1 had 745 automated unit tests. The editor is private, and its time or cost savings have not been measured.
Thorsten Meyer says six days of testing and revision turned the first working version of his AI-built PDF editor, Docestra, into a more heavily tested private tool, after real files exposed problems with saving, encrypted documents and redaction. In an account published November 13, he describes moving from a first version built November 6 to version 1.5.1 on November 12, and argues that making AI-generated software dependable takes more work than producing an initial demo.
Meyer says an AI coding agent built Docestra’s first version from a one-sentence request to make an Acrobat-like editor he could use. The local web app could open and save PDFs, add annotations and signatures, fill forms, and reorder pages. According to his account, it runs in a browser on his Mac and files remain on that machine. The code was written by Anthropic’s Claude Code running Claude Opus, while OpenAI’s Codex reviewed it from the start.
A test run across 930 real PDFs found that saving could produce a broken cross-reference table in a file made by a common PDF generator, causing markups to disappear. Meyer says he changed the save method to avoid compressed object streams. In a separate test, 52 of 69 encrypted PDFs did not permit annotations under restrictions set by their authors. Because the PDF library could not re-encrypt edited files, Docestra was changed to require confirmation before unlocking such a document, save an unprotected copy and never overwrite the original. Meyer reports that all 69 test files then unlocked.
Redaction required a more substantial redesign. The first version’s “white-out” feature covered text visually but left the underlying content in the PDF. Meyer says an independent AI review found 14 problems in the first implementation of real redaction and 13 in a second review. He replaced that approach with one that renders redacted pages as images, fills marked areas in the pixels and adds a searchable text layer that omits text touching a redaction. A test across 917 PDFs found zero leaks, according to Meyer; the account does not describe the test’s full methodology.
Docestra · Nov 6–13, 2026
From one prompt to a tool I can rely on
What the first version gave me, what real files broke, and what it took to fix it, from my project log.
What the tool rests on (schematic)
Built bottom up. The top layer is a plan, not a finished feature.
What happened, by date
- Nov 6One-sentence brief; first working version committed the same evening
- Nov 6Test over 930 real PDFs finds a save bug that dropped markups
- Nov 752 of 69 encrypted PDFs forbid annotations: unlock only into a new copy
- Nov 8Real redaction: review finds 14, then 13 problems; simpler design, 0 leaks in 917 PDFs
- Nov 12Version 1.5.1: 745 unit tests, 32 browser test suites
- Nov 13Version 2.0 starts: one shared PDF engine for our AI agents (in development)
Source: Docestra project records (development log, plans for versions 1 and 2.0, commit history), Nov 6–13, 2026. The layer stack is a schematic, not a measurement.
Graphic: Powered by Thorsten Meyer AI
Real PDFs Exposed the Reliability Gap
The account illustrates a practical distinction for people building software with AI tools: a feature that works in a small demonstration may still fail on documents created by different software, protected by restrictions or containing sensitive information. In a PDF editor, a save error can erase visible work, and a visual cover-up that leaves text intact can expose information a user believes has been removed.
Meyer’s reported fixes also show why reliability involves decisions beyond generating code. He changed how the app handles encrypted files rather than silently stripping protection, and replaced a repeatedly patched redaction method with a different design after two review rounds. Those choices matter to anyone considering AI-built tools for document workflows, though this is one developer’s account—not an independent evaluation of Docestra or a general measure of AI coding performance.
Meyer says the test and review findings were converted into automated checks. The unit-test count rose from 17 in the first version to 745 in version 1.5.1. That count indicates a larger test suite, but by itself does not establish that every failure mode is covered or that the application is suitable for other users.
Products Worth Considering
Apple MacBook Air 15-inch (2026) with M5 Chip, 16GB Memory, 1TB SSD, and Liquid Retina Display – Sky Blue
LG UltraGear 25G590B-B 25-Inch Full HD 1000Hz Gaming Monitor
Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite
As an affiliate, we earn on qualifying purchases.
From One-Day Build to Version 1.5.1
Meyer defines “vibe coding” as describing desired software in plain language and having an AI model write much of the code. He says Docestra’s initial version was committed at 9:47 p.m. Central European Time on November 6, 2026. Its initial feature list included viewing, searching, drawing, adding notes and text, filling forms, editing pages, combining or extracting pages, printing, undoing actions and saving.
During the following six days, testing and code review led to changes that were not specified in the initial brief. Meyer also says licensing informed development: he excluded fonts shipped under the GPL-2.0 licence under his project rules and later checked the licences of libraries added for text recognition and text editing. He reports that testing and rework took longer than building for several packages, but gives no time or cost measurements.
Docestra is described as a private tool that is not available to download. Meyer says development of version 2.0 began November 13 and was not complete when he published his account.
“Vibe coding gets you a working tool quickly. Turning it into a tool you can rely on is the real work.”
— Thorsten Meyer, describing the project’s main lesson
Testing Limits and Unanswered Questions
The results are reported by the developer, and the source account does not provide the test files, detailed procedures or independent verification of the findings. The zero leaks in 917 PDFs figure is a reported test result, not proof that the redaction design cannot fail on other documents or under different conditions.
It is also unclear how Docestra performs across operating systems, PDF generators and document types beyond those Meyer tested. The application is private and unavailable to download, limiting outside users’ ability to reproduce the tests. Meyer explicitly says he has not measured time or cost savings, so the account does not establish whether the project was faster or less expensive overall than other development approaches.
Version 2.0 was still unfinished as of November 13. The account does not give a release date, disclose a full independent security review or establish whether the tool will be made public.
Version 2.0 Remains in Development
Meyer says work on Docestra version 2.0 began November 13, 2026, the day his account was published, but he did not provide a completion date. The next development milestone is therefore the completion of that version; no public launch has been announced, and Docestra remains private according to the account.
For now, the reported path forward is continued development and testing. Whether version 2.0 will add features, expand the test coverage or change the tool’s availability remains unspecified. Meyer’s account also leaves open whether he will publish enough test detail for other developers to reproduce the reported results.
Key Questions
What is Docestra?
Docestra is a private, local web app that Meyer says he built with AI coding agents to view and edit PDFs on his Mac. It is not available to download.
What problems did testing uncover?
Meyer reports that one save method caused markups to disappear from some PDFs, many encrypted files did not permit annotations, and the first white-out feature left underlying text in place. He says he changed the save process and encrypted-file workflow and redesigned redaction.
Did the redaction tests prove the feature is completely safe?
No. Meyer reports zero leaks in a test of 917 PDFs, but the account does not give enough methodological detail to establish how broadly that result applies. It should not be read as proof that redaction can never fail.
Has Meyer shown that AI-built software saved time or money?
No. Meyer says he has not measured time or cost savings. He reports that testing and rework took longer than building for several packages, but does not give a project-wide comparison.
When will Docestra 2.0 be released?
No release date was given. Meyer said version 2.0 had been in development since November 13, 2026, and was not finished at the time of publication.
Source: ThorstenMeyerAI.com
