Don't blame Claude – It's me, I'm the problem, it's me

Developer Chris Greening detailed an ongoing AI coding struggle where Claude failed to adhere to strict style instructions after more than 700 commits in an open-source project, as documented in atomic14.com's report. Rather than attributing the issue to model limitations, the breakdown stemmed from conflicting context provided directly within the project workspace.

This analysis examines how accumulated project artifacts undermine AI instructions in long-running codebases and why developers must adjust existing project context to align model behavior. It details the interaction between prompt configuration files and existing code artifacts for engineers managing vibe-coded projects.

Don't blame Claude – It's me, I'm: what actually changed

Over 700 commits into developing Harmless, a browser-playable Elite-like space game vibe-coded with AI assistance, developer Chris Greening attempted to change how Claude formatted its text output. Greening requested that the model switch to ASD-STE100, also known as Simplified Technical English, and updated the project configuration file to enforce this requirement. Despite adding the explicit instruction to CLAUDE.md, the AI assistant continued generating text using short phrase naming conventions found in previous documentation and TODO items.

The breakdown occurred because historical project artifacts contradicted the newly added configuration rule. Every document, completed task record, comment, and TODO file created across the previous 700 commits contained the original phrasing style. When Claude processed the repository, the vast weight of historical context overrode the single rule defined in CLAUDE.md, leading to persistent formatting conflicts between requested instructions and observed project patterns.

Don't blame Claude – It's me, I'm: how it works

Large language models like Claude analyze the full context of a repository alongside configuration directives when generating output. In this project, CLAUDE.md specified ASD-STE100 rules, but the workspace was filled with older files using short phrases like "A citation that names nothing" or "The safe zone that only the spawner obeys". The model treated these established patterns as the true house convention of the repository.

Because the AI reads both configuration files and existing workspace documents, contradictory historical data neutralizes direct instructions. As a result, the assistant repeatedly prioritized matching the existing file style over following the output rules. Greening noted that expecting an AI model to ignore hundreds of commits of prior context is ineffective without active manual correction across the codebase.

Don't blame Claude – It's me, I'm: why it matters now

As vibe-coding workflows scale past initial prototypes into massive codebases, managing AI context drift has become a critical operational challenge for developers. Greening's experience demonstrates that changing prompt instructions late in a project lifecycle requires significant refactoring of existing documentation rather than simple configuration edits. The issue shows that AI performance failures often trace back to prompt and context misalignment rather than model bugs.

The findings highlight the limits of relying solely on files like CLAUDE.md when repository history pulls in the opposite direction. Developers attempting late-stage rule changes must actively guide the model and update legacy documentation to prevent context pollution. Understanding this dynamic helps engineering teams avoid wasted effort when attempting to alter AI behavioral patterns mid-project.

Don't blame Claude – It's me, I'm: who is affected

This structural context issue impacts software engineers, hardware designers, and open-source maintainers who utilize AI assistants for long-term project development. Greening, who also works with hardware automation and ESP32-S3 development boards using Atopile and ESP-IDF, encountered this limitation while building his browser-based game Harmless. Anyone managing vibe-coded repositories with extensive commit histories faces similar context retention challenges.

Developers who attempt to alter formatting styles, documentation standards, or architectural rules midway through a project are most vulnerable to this problem. When legacy files outnumber new instruction files, the AI will consistently regress to historical workspace norms. Teams building large systems with AI tools must recognize that configuration updates require systematic cleaning of older codebase references.

Don't blame Claude – It's me, I'm: what to watch

Moving forward, developers must treat context management as an ongoing maintenance task rather than a one-time configuration setup. Correcting an established repository trajectory requires constant manual oversight to align historical documents with new style requirements. Greening described this process as turning a massive tanker, requiring persistent effort on the tiller to steer the AI back toward ASD-STE100 compliance.

Engineering teams will need to develop better tooling and practices to prune outdated context from AI-assisted codebases. Future workflows must balance instruction files with automated refactoring of legacy documentation, comments, and task logs. Watching how developers manage long-term AI alignment across hundreds of commits will dictate the sustainability of vibe-coding methodologies.