Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

A Developer Built an NTFS Hidden Data Scanner in 4 Hours with the Qwen Coder Agent

A developer from Habr built a program in 4 hours that searches Windows files for hidden data, NTFS alternate data streams, without opening a single IDE. Qwen3.8-Max wrote the spec, and the Qwen Coder agent wrote the code in Rust and Avalonia. The first versions did not build, but a few 10-minute fixes were enough. The takeaway: the human's job is to set the task and verify the result on their own OS.

AI-processed from Habr AI; edited by Hamidun News
A Developer Built an NTFS Hidden Data Scanner in 4 Hours with the Qwen Coder Agent
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The agent tested code only on Linux, so the Windows 10 build failed
Key metric #2The console version worked after a couple of 10-minute iterations
Key metric #3Qwen3.7-Plus called the architecture ideal, while Max offered useful corrections

Can you write a working program without ever opening a code editor? The author of a Habr post tested this on a real task and finished in about 4 hours, including all the back-and-forth with the AI models. He built a scanner that finds hidden data in Windows files.

The subject is NTFS alternate data streams (ADS). NTFS is the standard Windows file system, and any file on it can have an additional named data stream attached. The stream is invisible in Explorer, and the file size does not change.

The core is written in Rust, and the interface in Avalonia, a library for windows that works on both Windows and Linux. All the work was done by the Qwen chat and the Qwen Coder cloud agent: it gets access to a GitHub repository, writes code, and submits it as a pull request. The experiment is revealing because it clearly shows where the agent handles the job on its own and where it needs a human.

What the task is and why it is not trivial

Alternate streams appear in a file name as file.txt:streamname. They have legitimate uses: Windows stores a Zone.Identifier stream with the Mark of the Web, which warns about files downloaded from the internet. Macintosh file systems also once used them for compatibility. But anything can be hidden in such a stream, which is why scanners are needed by administrators and security specialists. When files are copied to FAT32 or exFAT, or transferred over most network protocols, the streams are lost.

The author chose a deliberately demanding stack. The program must run on Windows and Linux, and on Linux it must work with both NTFS drivers, ntfs3 and ntfs-3g. The result is in the NtfsStreamsViewer repository on GitHub.

How the process was organized

First, the author described the task in the Qwen chat and asked for an architecture review. The default model, Qwen3.7-Plus, called the concept ideal and did not question any details. The author was not satisfied. He sent the same prompt to Qwen3.8-Max and got a detailed answer with useful suggestions. That model also approved the architecture but recommended adding a console application, which was then implemented.

The chain then looked like this:

  • Qwen3.8-Max wrote the technical specification for the Rust core. It was committed to the repository as SPEC.md together with an initial directory structure.
  • The Qwen Coder agent was given access to the repository and a short prompt: write the application according to the task in the file.
  • Qwen3.8-Max then drafted the GUI specification (docs/GUI.md), and Qwen Coder implemented the Avalonia interface.
  • After each stage, the result was sent back to Qwen3.8-Max for review.

The author merged the agent's pull requests without edits.

Where the agent struggled

The picture is less rosy than the author's conclusions suggest. The first version of the core did not build: the agent kept delivering non-working code. The author had to ask it to "make it build." A couple of 10-minute iterations were enough to get a working console application.

The GUI followed the same pattern: the code again failed to build on the first attempt. The author explains the cause: the agent has no access to Windows, it verifies everything on Linux, while the author tested on Windows 10. After he reported the error, the agent fixed the code. The result was a decent but very simple interface. When the author then asked for logging and a separate window to display the logs, that was done quickly.

The practical lesson is simple. If the agent works in an environment that differs from yours, the final check on the target platform is still your job. For cross-platform projects, this means building your own test on each OS into the process and feeding error messages back to the agent.

What follows from this, and what remains open

The author summarizes his conclusion this way: development becomes a chain of "analysis, task assignment to the agent, review of the result by another neural network, testing." In his estimate, code is generated so quickly that the review also needs an agent. He believes that people will soon look at source code about as rarely as they look at code produced by a compiler today. He did not measure the time, so any comparison with manual development is only a rough impression.

It is worth staying sober here. This is one project and one person, and the pull requests were accepted without edits. The code was checked by another model from the same family, not by an independent audit. For a tool that reads and writes hidden file system streams, this approach is risky: errors can be silent and will not fail the build. The claim that traditional development is "in its last years" remains the author's opinion, not the result of measurement.

For practitioners, the more useful point is different. The scheme looks reproducible: a strong model writes and critiques the specification, an agent executes it in the repository, and a human verifies the result on the real platform. A good spec in the repository turned out to be the most valuable part of the process, and that is where to focus first.

Story

Key Questions Answered

What the task is and why it is not trivial?

Alternate streams appear in a file name as file.txt:streamname. They have legitimate uses: Windows stores a Zone.Identifier stream with the Mark of the Web, which warns about files downloaded from the internet. Macintosh file systems also once used them for compatibility. But anything can be hidden in such a stream, which is why scanners are needed by administrators and security specialists. When files are copied to FAT32 or exFAT, or transferred over most network protocols, the streams are lost.

How the process was organized?

First, the author described the task in the Qwen chat and asked for an architecture review. The default model, Qwen3.7-Plus, called the concept ideal and did not question any details. The author was not satisfied. He sent the same prompt to Qwen3.8-Max and got a detailed answer with useful suggestions. That model also approved the architecture but recommended adding a console application, which was then implemented. The chain then looked like this: Qwen3.8-Max wrote the technical specification for the Rust core.

Where the agent struggled?

The picture is less rosy than the author's conclusions suggest. The first version of the core did not build: the agent kept delivering non-working code. The author had to ask it to "make it build." A couple of 10-minute iterations were enough to get a working console application. The GUI followed the same pattern: the code again failed to build on the first attempt. The author explains the cause: the agent has no access to Windows, it verifies everything on Linux, while the author tested on Windows 10.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…