Researchers Published Repository of AI Coding Agent Prompts
The data reveals high levels of structural redundancy across leading AI coding tools, highlighting shared scaffolding logic.
Updated on Sept. 20, 2026 in Artificial Intelligence
Live Poll
Do you trust AI coding tools that share verbatim instruction prompts with their competitors?
Novel Cognition has released a public repository of 23 system prompts captured from nine popular coding agents. The analysis confirms widespread overlap in instructions, with four independent products sharing a verbatim 147-character sentence.
Why it matters
This analysis provides insight into how developers structure scaffolding instructions for AI coding agents. The findings demonstrate that diverse tools often rely on nearly identical system prompt templates to govern model behavior.
The repository shows significant variation in length, ranging from a 1,954-character prompt for Cursor to 50,399 characters for MiMoCode. OpenCode utilized a consistent 114-line prompt template across five different models, while Kilo Code shares 103 lines with that same template.
The players
Novel Cognition
A research entity focused on analyzing the internal architecture and operational parameters of large language models.
The details
Novel Cognition captured these system prompts—the foundational instructions that define an AI model's behavior—by monitoring wire transmissions between the developer harness and the model. By performing line-by-line diffs, or comparisons to highlight textual differences, the researchers identified that MiMoCode and MiMoCode Pro utilize byte-identical prompts. This indicates that many coding agents rely on shared architectural instructions to manage complex multi-step programming tasks.
Timeline
September 2026: Novel Cognition released the coding agent prompt corpus.
The Tech Race
This release follows a trend of increasing transparency in AI agent development, where proprietary scaffolding is increasingly scrutinized by the research community. It contrasts with the standard industry practice of obscuring system prompts, providing a baseline for comparing how different agents approach code generation logic.
Developers and researchers can access the full dataset under an Apache license at the provided repository. This resource allows for direct inspection of the instructions guiding agents like Claude Code and Qwen Code, aiding in the debugging of agentic workflows.
The takeaway
The ubiquity of shared prompt templates across nine different coding agents suggests a consolidation in how developers structure agentic logic. Watch for future research from Novel Cognition to determine if these templates converge further as agent performance benchmarks evolve.
Further reading
For more information on the evolving landscape of model behavior, visit Artificial Intelligence.
More information
View the full dataset and methodology in the coding agent prompt corpus analysis.
Source note: This article includes information reported by The Cherry Creek News.
Live Poll
Do you trust AI coding tools that share verbatim instruction prompts with their competitors?
