ProgramAsWeights Releases 'Claudish' Translator to Strip AI Verbosity
The open-source tool uses a specialized framework to compile natural language specs into tiny, local neural programs.
ProgramAsWeights (PAW) has released an English ↔ Claudish translator designed to decode the dense, jargon-heavy communication style of the Claude AI assistant. The tool allows users to translate between plain English and the formulaic patterns often associated with Anthropic's models.
Built on the ProgramAsWeights framework, the translator compiles English specifications into pseudo-programs and LoRA adapters. These programs run locally on small pretrained interpreters, such as GPT-2 124M or Qwen3 0.6B. By executing locally, the tool ensures user privacy and eliminates the cost per call typically associated with large-scale AI APIs. The project has been released as open source under the MIT License.
The Rise of 'Claudish'
The term "Claudish" has emerged within the AI community to describe the specific linguistic markers—often repetitive or overly verbose—characteristic of Claude's outputs. This project follows a trend of users seeking to strip away "AI-speak" to reach direct answers. It was inspired by a previous effort from GitHub user gvzdv, who developed a "claudish-to-english" plugin for Claude Code that utilized local LLMs via Ollama to rewrite assistant messages into plain English.
Technical Implications
Beyond the novelty of translating AI jargon, the project demonstrates a shift in how small-scale AI functions are deployed. The PAW approach utilizes a compiler model to generate lightweight, deterministic local functions from natural language. This bypasses the need for constant API calls to massive frontier models, proving that specific linguistic transformations can be handled by highly efficient, localized neural programs.
What's Next
As AI assistants become more integrated into professional workflows, the demand for "AI-to-Human" translation tools is likely to grow. Observers will be watching to see if the PAW framework is applied to other distinct AI "dialects" or if the trend toward local, compiled neural programs gains wider adoption among developers seeking to reduce latency and API dependency. This shift toward localized execution suggests a future where the "personality" of an AI can be filtered or modified entirely on the client side, granting users more control over the interface of their digital assistants.