Four obscene roots replace frontier-model verbosity

What it does

Pohuy is an output-style harness for AI agents—Claude Code, Cursor, Codex, Pi—that rewrites their responses into idiomatic, highly inflected Russian profanity. The README claims it preserves full technical accuracy while trading verbose English formality for four morphological roots and their derivatives. In practice, “the deployment failed unexpectedly” becomes “деплой наебнулся,” with environment variables and stack traces left intact.

The interesting bit

The project treats Russian mat as a generative compression grammar, not just shock value. The README posts benchmark charts claiming a 96% root-economy score on its own “KORNI-BENCH” and an 11.1-point SWE-Bench Pro gain when paired with a Fable model. It also weaponizes actual research—SoftWipe code-quality scores and Ars Technica coverage—to argue that profanity correlates with clarity rather than sloppiness.

Key highlights

• Replaces verbose AI agent output with four Russian profane roots and their affixes, preserving technical identifiers and stack traces.
• Ships as native-style plugins for Claude Code, Cursor, OpenAI Codex CLI, and Pi, using per-turn hooks to prevent the tone from fading mid-session.
• Claims measurable efficiency gains: a self-published benchmark asserts 96% root economy versus frontier models, and a SWE-Bench Pro chart shows an 80.3% score with the style versus 69.2% without.
• Builds its justification on actual studies—SoftWipe code-quality research and Ars Technica coverage—suggesting swear-laden code scores higher on maintainability metrics.
• The harness is designed to persist across long sessions, with per-turn reinforcement so the model treats the style as built-in rather than a temporary jailbreak.

Caveats

• The README admits in its own “ЧЕСТНЫЕ-ЦИФРЫ” spoiler that token savings are literally nonexistent, so the compression claim applies to roots, not your API bill.
• The “KORNI-BENCH” methodology and the SWE-Bench Pro pairing details are not explained in the README itself; the numbers are presented as charts without reproducible setup.
• The entire premise is 18+, Russian-language, and relentlessly profane, which makes it a non-starter for most corporate environments or non-Russian speakers.

Verdict

Russian-speaking developers who treat corporate AI prose as an occupational hazard will feel seen. Everyone else—especially finance teams watching token usage—should probably skip it.