A Habr author tested whether ChatGPT can reconstruct articles from short prompts

A Habr author checked whether ChatGPT can restore articles from short prompts

The author on Habr decided to test a bold claim that any text can be compressed into a short prompt and then restored almost without loss through ChatGPT. Instead of relying on references to "Cambridge researchers," he took real articles from the platform and conducted his own test.

How the test was set up

The trigger was a reaction to an earlier translation of The Prompt article, which claimed that some Cambridge researchers can reduce any text to a minimal prompt with 98% recovery accuracy. Readers quickly noticed that the story looked like a stylization of "British scientists," which prompted the author to test not the legend, but the principle itself. For the experiment, he chose Habr as a platform with a strict technical audience where weak arguments and stretches usually don't last long.

The scheme was simple: take two recent high-rated articles, compress them into a prompt, and ask ChatGPT to write a new text in a similar genre without internet search. The author looked not only at the overall tone but also at more concrete things: whether the structure, key numbers, important episodes, and order of arguments were preserved. He separately tracked where the model would start filling in the story on its own, because such insertions are easily mistaken for real details.

Where it worked

The first test turned out to be almost a demonstration of the power of template-based recovery. It was about an article with the thesis that indispensable employees are not a defect of team architecture, but a valuable resource. For such a publication, a prompt of just 67 words was enough, after which ChatGPT generated text of 651 words. According to the author's assessment, the match was so high that next to the original, the reconstruction looked frighteningly convincing.

• The main thesis about the value of "indispensable" employees was preserved
• Criticism of standard ways to reduce bus factor returned
• The model reproduced recommendations: pay 1.5–2 times more, document through process, hire independent people
• Two practical scenarios and the exact figure of a 40% increase remained in place
• ChatGPT even added a plausible detail — a nine-month timeline of failure that wasn't in the original

The compression ratio in this case was approximately 10:1. But along with the impressive result came a problem: the model doesn't just recover familiar logic, but confidently fills in gaps with what "sounds like truth." For a reader who doesn't have the original nearby, the difference between the restored and real article can be almost imperceptible, especially if the text is built on common management patterns that the model has long known.

Where the method has limits

The second article gave a completely different result. The material was devoted not to abstract conclusions, but to an analysis of Telegram blockages, DPI, and how the developer community manually fixed specific errors in FakeTLS implementation. To get closer to the original logic, the author had to write 357 words of prompts — almost five times more.

ChatGPT output 914 words, but the important part of the text still dissolved. What was missing was exactly what distinguishes a retelling from real engineering work: specific TLS extension values, discrepancies between stated and actual key size, pull request numbers, commit hashes, the name of the community that brought the fix, and other artifacts mined from traffic and code. In other words, the model recovered the general argument but couldn't return what was found by hand in a live system, not what lay in the training data.

"ChatGPT recovered the argument.

It didn't recover the work."

This is where the author draws the line between convenient imitation and real knowledge. If the value of a text rests on structure, familiar theses, and typical conclusions, it compresses well and decompresses reasonably well. But if an article consists of manual reverse engineering, measurements, packet interception, and new observations, those same "2% losses" turn into almost everything the material was read for in the first place.

What this means

The experiment on Habr doesn't prove that LLMs are already capable of replacing authors entirely, but it shows quite accurately where the practical boundary lies. Columns, explanatory texts, and management essays with common patterns models reassemble much more easily than materials with unique facts and their own measurements. For editors and readers, a simple test follows: what matters is not just how convincingly an article sounds, but whether there's something in it that the author really found, checked, and proved themselves. This part of the content currently withstands compression into a prompt and reverse assembly the worst.