🔍 Read the full analysis: I Used AI To Build A PDF Editor—and Learned Where The Work Begins on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Thorsten Meyer says an AI coding agent produced a working first version of Docestra, a private PDF editor, on Nov. 6, 2026. Over six days of testing and revisions, he found save failures, limits on encrypted files and risks in an early redaction design; he says the project is still private and has no measured time or cost savings.
Thorsten Meyer says an AI coding agent built the first working version of his PDF editor, Docestra, on November 6, 2026, but tests on real documents quickly exposed problems that a working demo had missed. In an account published Nov. 13, Meyer describes six days of revisions to the private, local web app, including a save failure, restrictions on encrypted PDFs and a redaction feature that underwent repeated independent review.
Meyer says he gave an AI coding agent a one-sentence brief to build an Acrobat-like tool for his own use. The first version was committed at 9:47 p.m. Central European Time that day. It could open PDFs and support tasks including highlighting, drawing, notes, signatures, form filling, page reordering and saving. Meyer describes the project as a TypeScript web app that runs in a browser on his Mac, with files staying on the machine.
The first substantial test was a run over 930 real PDFs. Meyer says it revealed that saving some documents could produce a broken cross-reference table, causing annotations to disappear. He changed the save method to avoid compressed object streams. His tests also found that 52 of 69 encrypted PDFs did not permit annotations under restrictions set by their authors; the library Docestra used could not re-encrypt edited files. Meyer says the app was changed to ask for confirmation before unlocking these files, save an unprotected copy and leave the original untouched. He reports that all 69 test files then unlocked.
Redaction required a larger redesign. The first version’s “white-out” tool covered text visually but did not remove it from the PDF. Meyer says an independent AI review found 14 problems in its first round reviewing a new in-place redaction design and 13 in a second round. He then replaced that approach: pages with redactions are converted to images, marked areas are filled in the pixels, and a searchable text layer is added without text that touches a marked area. Meyer reports that tests on 917 PDFs found zero leaks. That result is his account of the test, not an independently published evaluation.
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
Why Document Testing Changed the Build
The project illustrates a distinction between producing a functional prototype and establishing that a document tool behaves safely across varied files. A PDF editor can appear to work while silently losing annotations, mishandling author-set restrictions or leaving supposedly redacted information in the document. Those failures matter because users may rely on the saved file, not just what appears on screen.
Meyer’s account also shows the value and limits of using a second AI model as a reviewer. He says OpenAI’s Codex reviewed code generated with Anthropic’s Claude Code running Claude Opus. The review surfaced issues, but the reported findings also led him to stop patching an intricate design and replace it with a simpler one. His report does not establish that AI review alone can certify document security; it describes one developer’s process, tests and decisions.
For readers considering AI-built tools, the practical lesson is narrower than a claim that such software is ready for general use: testing against representative files, checking output and preserving originals can expose failures that a short demonstration will not. Meyer says he has not measured time or cost savings, so the account does not quantify whether the approach was more efficient overall.
Products Worth Considering
Bose QuietComfort Wireless Headphones – Twilight Blue (Limited Edition)
Samsung 27-inch Odyssey OLED G6 QHD Gaming Monitor
Bose QuietComfort Wireless Noise-Cancelling Over-Ear Headphones, Cypress Green
As an affiliate, we earn on qualifying purchases.
From One-Sentence Brief to Version 1.5.1
Meyer defines “vibe coding” as building software mainly by describing desired behavior in plain language and asking an AI model to write the code. In Docestra’s case, he says AI coding agents produced the code and a second model reviewed it from the first day. The initial app already had a broad set of editing functions, but Meyer says it had been tried on only a small number of files before the larger robustness run.
From version 1 on Nov. 6 to version 1.5.1 on Nov. 12, Meyer reports changes to saving, encrypted-file handling and redaction, alongside other additions. He says the automated unit-test count grew from 17 in the first version to 745 in version 1.5.1. He also says he checked the licences of bundled libraries, including excluding fonts distributed under GPL-2.0 because his project rules did not allow GPL code.
These details come from Meyer’s first-person account on ThorstenMeyerAI.com. Docestra is described as a personal project, not a publicly downloadable product. His test counts and reported outcomes are not presented in the source as an outside audit or peer-reviewed assessment.
“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 about the project
What the Report Does Not Establish
The results are Meyer’s reported findings; the source does not describe independent replication of the 930-file robustness run or the 917-file redaction test. It also does not give enough detail to assess the composition of each test set, how the PDFs were selected or how “zero leaks” was defined and checked.
Docestra remains private and is not available to download, according to Meyer. He says he has not measured time or cost savings, and his account does not provide a full comparison with commercial PDF editors. The report also does not establish how the app would perform across other operating systems, browsers, document types or users’ workflows.
Meyer says version 2.0 was in development as of Nov. 13 and was unfinished. Its planned capabilities and release timing are not specified in the source material.
Version 2.0 Remains in Development
Meyer says development of Docestra version 2.0 began on November 13, 2026, the day his account was published, but the release was not finished. He does not give a launch date or say whether he plans to make the software public.
For now, the next step described is continued development and testing of his personal tool. The account leaves open whether version 2.0 will add features, expand the test process or change the app’s availability. Further claims about reliability or savings would require measurements and evidence beyond the results Meyer reports here.
Key Questions
What is Docestra?
Docestra is a private PDF editor that Thorsten Meyer says he built with AI coding agents for his own use. It runs as a local web app in a browser on his Mac, and he says files stay on the machine.
What did Meyer say the first version could do?
The Nov. 6 version could open PDFs and, according to Meyer, support highlighting, drawing, notes, signatures, form filling, page changes and saving. Later testing found problems that were not apparent from the initial working version.
What problems did testing uncover?
Meyer reports a save issue that caused annotations to disappear from some PDFs, restrictions affecting 52 of 69 encrypted test files, and shortcomings in an early redaction method that only covered text visually.
Did the redaction test prove Docestra is secure?
No. Meyer reports that a revised redaction approach had zero leaks in a test of 917 PDFs, but the source does not describe an independent audit or provide enough test-method detail to treat that result as a general security guarantee.
Can the public download Docestra?
No public download is described. Meyer says the app is private and unavailable to download. He also says version 2.0 is in development and unfinished, without giving a release date.
Source: ThorstenMeyerAI.com
