I Built A PDF Editor With AI. Making It Work Reliably Took The Real Effort
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: I Built A PDF Editor With AI. Making It Work Reliably Took The Real Effort on ThorstenMeyerAI.com

Before you orderOffer from Amazon

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
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

Thorsten Meyer reports that an AI coding agent built a working PDF editor prototype, Docestra, on Nov. 6, 2026, but real-file tests exposed saving, encryption and redaction problems. By version 1.5.1 on Nov. 12, he says the private app had 745 unit tests; its time or cost savings have not been measured, and version 2.0 remains unfinished.

Thorsten Meyer says an AI coding agent produced a working first version of his PDF editor, Docestra, on Nov. 6, 2026, but six days of testing and revision were needed to address problems exposed by real documents. In an account published Nov. 13, Meyer describes issues involving saved annotations, encrypted files and redaction, arguing that reliability work—not the initial prototype—was the larger effort.

Meyer says the first version was committed at 21:47 Central European Time on Nov. 6. It could open and search PDFs, add highlights, drawings, notes and signatures, fill forms, rearrange pages and save files. The app was built as a local TypeScript web application running in a browser on his Mac; he says files do not leave the machine. The code was generated using Anthropic’s Claude Code with Claude Opus, while OpenAI’s Codex reviewed it from the start.

A test run across 930 real PDFs found that the chosen PDF library could create a broken cross-reference table for files made by a common PDF generator. Meyer says that could cause markups to disappear after saving. He changed the save process to avoid compressed object streams. The test also covered 69 encrypted PDFs: according to Meyer’s account, 52 did not permit annotations under their authors’ restrictions, and the library could not re-encrypt edited files. Docestra was changed to ask for confirmation before unlocking those files, then save an unprotected copy without overwriting the original. Meyer says all 69 test files could then be unlocked.

Redaction prompted a separate redesign. The original “white-out” tool only covered text visually; Meyer says the underlying PDF content remained. After he built a first version of actual redaction on Nov. 8, an independent AI review identified 14 problems in its first round and 13 in a second. He says he abandoned repeated patching in favor of converting redacted pages to images, applying the redactions to the pixels and adding a searchable text layer that omits lines touching a marked area. A test of 917 PDFs found zero leaks, according to Meyer. That result describes his reported test, not a guarantee for all documents or use cases.

At a glance
reportWhen: Published Nov. 13, 2026; development ac…
The developmentThorsten Meyer published an account of six days of testing and revisions that followed an AI coding agent’s rapid build of his private PDF editor, Docestra.
Crypto market snapshot
Fear & Greed Index
20/100 — Extreme Fear
Bitcoin BTC$69,145▼ 4.5%
Ethereum ETH$1,953▼ 2.4%
Tether USDT$0.9992▲ 0.0%
BNB BNB$639.9▼ 2.2%
USDC USDC$0.9996▲ 0.0%
XRP XRP$1.02▼ 0.9%
Solana SOL$80.62▼ 0.8%
TRON TRX$0.3126▼ 1.0%
Live data · CoinGecko · alternative.me (24h change)

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

  1. Nov 6One-sentence brief; first working version committed the same evening
  2. Nov 6Test over 930 real PDFs finds a save bug that dropped markups
  3. Nov 752 of 69 encrypted PDFs forbid annotations: unlock only into a new copy
  4. Nov 8Real redaction: review finds 14, then 13 problems; simpler design, 0 leaks in 917 PDFs
  5. Nov 12Version 1.5.1: 745 unit tests, 32 browser test suites
  6. Nov 13Version 2.0 starts: one shared PDF engine for our AI agents (in development)
Working is not the same as reliable. The first version worked on the files I tried. Real files decided what had to change.

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

Testing Exposed the Prototype’s Limits

The account illustrates a gap between software that works in a demonstration and software that can be trusted with everyday documents. A save failure can silently remove annotations; a visual cover-up can leave sensitive text in a file; and changing an encrypted document can raise questions about the author’s restrictions and the protection on the saved copy. Those are practical risks for anyone using an AI-built tool to handle contracts, forms or other sensitive material.

Meyer says the work included adding tests for problems found during reviews. The unit-test count rose from 17 in the first version to 745 in version 1.5.1. The account also notes a licensing decision: the PDF viewer library includes fonts under GPL-2.0, which Meyer says he did not bundle because his rules excluded GPL code. He says he later checked the licences of libraries added for text recognition and editing.

These details are Meyer’s report about a private project, not an independent assessment of Docestra. He says he has not measured time or cost savings. The article therefore offers a case study of the engineering and review burden after AI-assisted code generation, rather than evidence that the approach is faster or cheaper overall.

Products Worth Considering

Apple MacBook Pro 14.2-inch with Tandem OLED Display, Space Black

Apple MacBook Pro 14.2-inch with Tandem OLED Display, Space Black

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 to an AI model and letting it write the code, rather than writing the code directly. Docestra began with a one-sentence brief on Nov. 6: to build a usable Acrobat-style editor. The first version included a broad set of editing features, and subsequent requests came from Meyer’s own workflow.

He describes development from version 1 on Nov. 6 through version 1.5.1 on Nov. 12. A feature to unlock restricted PDFs, for example, followed a Nov. 7 request to edit an encrypted file. The project remained private and was not available to download when Meyer published his account. He says work on version 2.0 began Nov. 13 and was still underway.

“Vibe coding gets you a working tool quickly. Turning it into a tool you can rely on is the real work.”

— Thorsten Meyer, writing on ThorstenMeyerAI.com

Limits of the Reported Tests

The test results come from Meyer’s own account; the supplied material does not include independent verification, test files or a full methodology. The reported zero leaks across 917 PDFs applies to that test set and does not establish how the redaction method performs on every PDF or in other workflows. It is also unclear how the app handles documents beyond the reported cases or whether later testing has found further faults.

Docestra is private and unavailable for download, so readers cannot independently examine or reproduce the reported results. Meyer says he has not measured time or cost savings. The account does not establish whether the tool is suitable for organizational use, nor does it provide an independent security assessment of the local application.

Version 2.0 Remains in Development

Meyer says development of Docestra 2.0 began on Nov. 13, 2026, and that the release is not finished. He does not give a release date or describe a public download plan. For now, the next milestone is continued development; the source account does not specify which features or tests are planned.

Any broader assessment will depend on what future versions include and whether their behavior can be checked beyond Meyer’s own test runs. The current account establishes a development timeline and the problems he says he addressed, but not measured productivity gains or independent confirmation of reliability.

Key Questions

What is Docestra?

Docestra is a private PDF editor Thorsten Meyer says he built with AI coding agents. It runs locally in a browser on his Mac, and he says documents stay on the machine.

What problems did testing find?

Meyer reports a save issue that could make annotations disappear, restrictions affecting encrypted PDFs, and an early redaction tool that covered text without removing it from the PDF. He says he revised the save and encryption handling and replaced the first redaction design.

Does the redaction test prove every document is safe?

No. Meyer reports zero leaks in a test of 917 PDFs, but the account does not provide independent verification or establish results for every file and use case.

Can the public download Docestra?

No public download is identified in Meyer’s account. He says Docestra is private and that version 2.0 was still in development as of Nov. 13, 2026.

Did Meyer measure time or cost savings?

No. Meyer explicitly says he has not measured time or cost savings. His account describes the development process and testing, not a quantified comparison with conventional software development.

Source: ThorstenMeyerAI.com

Nothing in this article is financial or investment advice. Cryptocurrency and precious-metal investments carry significant risk — do your own research and consider a licensed advisor.
You May Also Like

DeepSWE – The benchmark that made the models spread out again

DeepSWE, released May 26, 2026, reveals wider gaps among AI coding models, challenging previous benchmarks that underestimated differences between top models.

The NVIDIA Earnings Preview: What Q1 FY27 Will Reveal About the AI Cycle

NVIDIA reports Q1 FY27 earnings on May 20, 2026, with analysts watching for revenue near $78B, revealing key trends in the AI cycle and infrastructure demand.

What Is CBDCS

Discover the transformative potential of CBDCs and how they could reshape our financial landscape—what implications do they hold for the future?

The Real Trade-offs Of Downscaling AI Models To Four Bits

Examining the effects of reducing AI model precision to four bits, including capabilities lost and the implications for deployment and reliability.