A Ukrainian developer tweaked Claude Code to make it respond shorter. How did she do it?
Ukrainian developer Kateryna Kucherenko published her setup for Claude Code, which can reduce the amount of text in the AI assistant’s responses.
Ukrainian developer Kateryna Kucherenko published her setup for Claude Code, which can reduce the amount of text in the AI assistant’s responses.
The developer shared this on Linkedin.
«‘A lot of text. I didn’t finish reading it. Cut it short’ is my real line from the Claude Code session. The question was simple, the answer was three screens. I dug through my session history: there are a lot of such lines,» says Kucherenko.
In her opinion, with the spread of AI coding, the burden on the developer has also changed. Part of the code writing can be given to the model, but the human still has the decision, control and verification of the result. And the long answers of the AI agent add to this the need to constantly search for the main thing among the explanations.
«Before AI, code was written by hand and the brain rested on this mechanics. Now AI writes the code, and I have everything else left: decisions, control, verification. And the main enemy turned out to be not only hallucinations, but the amount of text that needs to be read to get the gist. Cognitive load,» explains the developer.
To combat this, Kucherenko created her own output style for Claude Code. It’s a set of instructions that are added to the system prompt and change how Claude forms answers. Her style was named Roy Kentre, after the straightforward character in the TV series «Ted Lasso.» The repository describes the style as follows: «the main thing is to immediately, rigorously weed out the unnecessary, without summaries at the end, and communicate like a colleague, not an assistant.»
The developer noticed that during long sessions, Claude gradually lost his style and returned to long replies. So the developer added a hook, a mechanism that fires after each new message from the user and adds a short style reminder to the context.
«A few hundred tokens and the persona lives the entire session,» explains Kucherenko.
Also recently, Claude Code has a built-in Concise style, which should also make the model answer shorter. Kucherenko asked Claude to «dig up» its prompt from the binary files and, she says, found about 180 lines of instructions. There, she saw a similar mechanism with a short reminder of the style while working.
«The hand-assembled pattern matched the vendor’s,» the developer notes.
However, she doesn’t plan to switch to the standard Concise. After all, it is tailored specifically to her way of working and, in addition to brevity, sets Claude a specific manner of communication.
Kucherenko posted the setup in an open repository. It also includes a system for transferring work between sessions, a context fill indicator, and a set of separate AI agents for writing code and reviewing.
Previously, dev.ua wrote that Anthropic allowed Claude Code users to create their own output styles and thus change the manner of the assistant’s responses.
«Smelly» configuration files cause AI agents to waste tokens — study
«Smelly» configuration files cause AI agents to waste tokens — study
How to save AI tokens? A selection of tools created by IT professionals to track expenses at Claude Code
How to save AI tokens? A selection of tools created by IT professionals to track expenses at Claude Code
Read the country’s main IT news in our Telegram
Read the country’s main IT news in our Telegram
