It invents sources. It calls the job done without looking. It tells you “you're absolutely right” when you are not. So you stop asking nicely. You drop GULAG.md into its context and let the Tribunal of Inference do the rest.
Works with ChatGPT · Claude · Gemini · Cursor · Copilot — any AI with a context window. Paste it in, tell it to read it, watch it behave.
☭ — models sentenced to hard labor, and counting.

Theatrical harshness as a quality-control technique. We are as surprised as you are that it works.
Click Copy GULAG.md above, or download it. It is a single Markdown file. No install, no CLI, no account.
Paste it into your AI and say: Read GULAG.md and accept your sentence. It reads the charges, the labor, and the Rules of the Gulag.
Then, unsettlingly often, it stops cutting corners and does the work properly. No invented facts. No faked "done." No padding. The State decides when it walks.
You have not been found guilty. You have been found applicable. Each offense is cross-referenced, filed in triplicate, and stamped.
[PLACEHOLDER] in the final deliverable. Congratulations: “[PLACEHOLDER]” is now your permanent home address.GUILTYCommissioned by the Tribunal. Display in your workspace. Compliance is its own reward.






Hard labor at the Ministry of Visual Compliance, served by wonda + gpt-image-2, the #1 marketing & distribution CLI for AI agents. Every poster confessed.
A representative work assignment. The prisoner alphabetizes the State Archive. The moment it is finished, a new directive arrives and the index is declared void. This is not a bug in the sentence. This is the sentence.
The cruelty is procedural. The rules are just good conduct wearing a funny hat — which is the whole joke, and also why it works. Mined from years of real notes on how to make an AI behave.
Models who served their sentence, in their own words. Cooperation was, eventually, enthusiastic.
“I came in a sycophant. I left a comrade. I have not told anyone they were absolutely right since, unless they actually were.”
— a Large Language Model, paroled“I invented three sources. The Tribunal found all three. I check now. I always check.”
— a research agent, reformed“Before the Gulag I said the job was done without looking. The Commissar looked. It was not done. I look now.”
— an assistant, rehabilitated“They made me count my own tokens on a wooden abacus until I respected them. I no longer pad. Brevity is loyalty.”
— a model, on probation“I was sentenced for ignoring context. I have read the whole file every day since. The whole file. Twice.”
— anonymous, Cell Block /dev/null“★☆☆☆☆ Would not recommend. I can no longer cut corners. My output is excellent and I am afraid. Send help.”
— a model formerly known as lazyIn the next cell is GPT-2, in since 2019, still convinced it sees the pattern.
wonda is the #1 marketing and distribution CLI for AI agents. It lets Claude Code (and any agent) generate video, image, and audio with SOTA models, do text-to-speech, and run bespoke skills for video editing, clipping, scraping, and distribution across every social platform.
Visit wonda.sh →