VQV Signal

RESEARCH WATCH TECHNICAL

Study finds coding agent prompts are mostly tool definitions that change between releases

According to a post on X by @kbhaskar_95, a new arXiv study by Paixão, Tavares, Almeida and Ahmed captured the first model request from 1,886 npm releases of Claude Code, Codex, Gemini CLI, Qwen Code and OpenCode. In Claude Code 2.1.243, the user's prompt was 0.03% of the request, while 20 tool definitions made up 77%. The post also says Claude Code 2.1.69 briefly replaced 18 tools with a single ToolSearch tool, cutting the request from 77,601 to 20,380 characters before version 2.1.70 reverted the change.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-10-29T00:35:00+00:00 Detected 2026-10-29T05:22:57+00:00
View original source

According to a post on X by @kbhaskar_95, a new arXiv study by Paixão, Tavares, Almeida and Ahmed captured the first model request from 1,886 npm releases of Claude Code, Codex, Gemini CLI, Qwen Code and OpenCode. In Claude Code 2.1.243, the user's prompt was 0.03% of the request, while 20 tool definitions made up 77%. The post also says Claude Code 2.1.69 briefly replaced 18 tools with a single ToolSearch tool, cutting the request from 77,601 to 20,380 characters before version 2.1.70 reverted the change.

AI-assisted summary based on the listed source.

A new arXiv study (Paixão, Tavares, Almeida, Ahmed) captured the first model request from 1,886 npm releases of Claude Code, Codex, Gemini CLI, Qwen Code and OpenCode. In Claude Code 2.1.243 the user's prompt was 0.03% of the request and 20 tool definitions were 77%; Claude Code 2.1.69 briefly swapped 18 tools for...

For developers and businesses that build on or rely on these coding agents, the study suggests that much of each request is made up of tool definitions rather than the user's own instructions. Those definitions can change between releases, as the study's title notes, so behavior may shift without a clear announcement. The source is a short post with a truncated snippet, so the full findings should be checked in the original paper.

Treat this as research or capability evidence, not a product promise; inspect the source before applying it.

Signal Strength 95% Technical label WATCH Public Interest 56 Category RESEARCH Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 60 Practical Impact Score 46 Novelty Interest Score 94 Consequence Score 30 Curiosity Score 16 Shareability Score 66

VQV surfaced this signal because it is recent, relevant to AI Coding Tools, connected to Scobleizer News Pulse (X).