TL;DR
Thorsten Meyer says an AI coding agent produced a working PDF editor, Docestra, on November 6, 2026. Tests across real files exposed problems with saving, encrypted documents and redaction; the development log describes fixes through version 1.5.1 on November 12. The author has not measured time or cost savings, and Docestra remains private.
Thorsten Meyer says an AI coding agent built a working version of his PDF editor, Docestra, on November 6, but testing on real documents quickly exposed defects that the initial build had not revealed. In a development account published November 13, Meyer describes six days of revisions through version 1.5.1, including changes to saving, encrypted-file handling and redaction. The account illustrates the gap between producing a functioning prototype and making a tool dependable in daily use.
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
Meyer says he gave an AI coding agent a one-sentence brief to build an Acrobat-like editor he could use. The first version was a local web app written in TypeScript, running in a browser on his Mac. He says files stay on the machine. The initial app could open PDFs, search, highlight, draw, add notes and signatures, fill forms, rearrange pages and save edited files. Meyer identifies the coding agent as Claude Code running Claude Opus, with OpenAI’s Codex reviewing code from the start.
A test run over 930 real PDFs found that saving some documents could produce a broken cross-reference table, causing markups to disappear from the saved file. Meyer says he changed the save approach to avoid compressed object streams. The same testing covered 69 encrypted PDFs. According to the account, 52 restricted annotations under settings chosen by the documents’ authors, while the PDF library could not re-encrypt files after editing. Meyer decided the app would ask for confirmation before unlocking such a file and save an unprotected copy, leaving the original untouched. He says all 69 test files then unlocked.
Redaction required a separate redesign. The initial “white-out” feature covered content while leaving underlying text in the PDF, a limitation Meyer says was stated in the project plan. On November 8, he built a redaction feature intended to remove content. An independent AI review identified 14 problems in its first review and 13 in a second review, which Meyer says mostly concerned gaps in earlier fixes. The source account says he then chose a simpler design instead of continuing to patch the initial approach. It reports zero leaks across 917 PDFs, but does not describe the test method in detail.
Testing Changed the Editor’s Design
The development log matters because a successful demonstration on a few sample documents can conceal failures that affect a tool’s core purpose. In Docestra’s case, a save defect could make annotations disappear, encrypted-document restrictions created a decision about preserving user protections, and a visual cover could be mistaken for actual redaction. These are practical reliability and privacy concerns for anyone editing contracts or other sensitive records.
The account also shows why review and testing do not end when an AI-generated app runs. Meyer describes testing a large collection of real files, reviewing failures and changing design decisions. The project log records 745 unit tests and 32 browser test suites for version 1.5.1. Those counts document the project’s checks; by themselves, they do not establish how the app performs for all users or every kind of PDF.
Meyer says he has not measured time or cost savings. The editor is private and unavailable to download, so outside users cannot independently assess its behavior. The reported development experience is therefore evidence about one project, rather than a general measure of what AI coding tools can deliver.
Popular right now
De'Longhi Magnifica Duo Fully Automatic Espresso Machine with Milk Frother, Silver
Ninja CREAMi 7-in-1 Ice Cream Maker, NC301
KitchenAid Artisan Series 5-Quart Tilt-Head Stand Mixer, Dried Rose
As an affiliate, we earn on qualifying purchases.
Six Days of Document Testing
The chronology in Meyer’s account runs from November 6 to November 13, 2026. He says the first working version was committed at 21:47 Central European Time on November 6. The 930-file robustness run that day surfaced the save bug before he encountered it in daily use. On November 7, he settled on unlocking restricted encrypted files only after confirmation and saving them as a new, unprotected copy.
On November 8, work shifted to redaction. The review findings prompted a simpler design, and the account reports no leaks in a set of 917 PDFs. By November 12, Docestra had reached version 1.5.1, with the test counts listed in the project records. Version 2.0 work began November 13. Meyer describes its direction as building one shared PDF engine for AI agents, and says that work remains in development.
Vibe coding, as Meyer defines it, means describing a desired tool in plain language and having an AI model write much of the code. His account distinguishes that fast start from the later work of testing and revising. It is based on his project records, including the development log, plans and commit history, rather than an external evaluation of Docestra.
““Vibe coding gets you a working tool quickly. Turning it into a tool you can rely on is the real work.””
— Thorsten Meyer, describing his conclusion
What the Project Log Cannot Show
Docestra is private and unavailable to download, so readers cannot reproduce the reported tests or inspect the application. The account provides test counts and outcomes but gives limited detail about the 930-file collection, how documents were selected, or how the reported redaction leaks were checked. The results are attributed to Meyer’s project records and should be read on that basis.
It is also unclear how the editor handles PDF files outside the tested sets, including documents with different structures or restrictions. Meyer says he has not measured time saved or financial costs, so the account does not establish whether the tool is more efficient than existing alternatives. Version 2.0 is described as underway, not complete, and its planned shared engine is not yet a finished feature.
Version 2.0 Remains in Development
Meyer says development of Docestra 2.0 began November 13, with a plan for one PDF engine shared by the project’s AI agents. The account does not give a release date or say what capabilities will be included in a finished version. Version 1.5.1 is the latest completed release described in the source material.
The next useful evidence will be the project’s progress from plan to finished software, along with further details about how it is tested. Until then, the reported milestones show what Meyer says changed during the prototype’s first six days, while broader performance, savings and availability remain unsettled.
Key Questions
What is Docestra?
Docestra is a local web app for viewing and editing PDFs that Thorsten Meyer says he built with help from AI coding agents. It runs in a browser on his Mac, and he says files stay on the machine.
What problem did testing find in saved PDFs?
A run across 930 real PDFs found that saving some files could break a cross-reference table and make markups disappear. Meyer says he changed the save method to avoid compressed object streams.
How does Docestra handle encrypted PDFs?
Meyer says the app asks for confirmation before unlocking encrypted files with restrictions that prevent annotations. It then saves an unprotected copy and does not overwrite the original.
Can users download or evaluate Docestra?
No. Meyer describes the app as private and unavailable to download, so the reported results cannot be independently reproduced by outside users from the information provided.
Has Meyer measured whether Docestra saves time or money?
No. The account says time and cost savings have not been measured. It describes the development and testing of the editor, not a comparison with other PDF tools.
Source: Thorsten Meyer AI
