New OSS Tool Tackles AI Coding Inconsistency

Same Prompt, Different Code — Every Single Time

Ask an AI coding assistant to 'write a function to fetch the list of users' today, and you might get getUsers(). Ask again tomorrow and it returns fetchUserList(). By next week, it has moved on to loadAllUsers(). The logic works fine each time — but the codebase slowly devolves into a patchwork of conflicting naming conventions, architectural patterns, and style choices.

This is the dirty secret of AI-assisted development in 2025. Tools like Claude Code, Cursor, and Amazon's newly launched Kiro are remarkably capable at generating functional code. But consistency — the kind that separates a maintainable professional codebase from a tangled mess — remains an unsolved problem for most teams.

One developer decided to tackle it head-on by building an open-source solution.

The Problem Nobody Talks About

After six months of heavy AI-assisted coding, the developer behind the project hit a wall. The inconsistencies weren't random — they were systematic. Every AI session essentially started from scratch, unaware of the team's preferred conventions.

'My initial reaction was maybe I need to write better prompts,' the developer explained. 'I wrote better prompts. The functions got slightly better. New inconsistencies appeared elsewhere.'

The root cause wasn't the AI's capability. It was the absence of what every experienced engineering team accumulates over time: unwritten rules. Use get not fetch. Always return arrays wrapped in a result object. Put repository logic in /services, not /utils. These micro-decisions define a codebase's coherence, and no amount of prompt engineering reliably captures them all.

This realization — that the AI had never been given the team's implicit conventions — sparked the creation of a dedicated open-source tool designed to codify and enforce those rules across AI coding sessions.

How It Works

The project operates on a straightforward principle: if unwritten rules are the problem, make them written — in a format AI assistants can actually consume.

Rather than relying on linters or static analysis alone, the tool creates a structured conventions layer that sits between the developer and the AI assistant. Think of it as a team style guide specifically optimized for LLM consumption. It captures naming conventions, architectural patterns, file organization rules, error handling preferences, and dozens of other micro-decisions that typically live only in senior developers' heads.

The tool integrates with the major AI coding environments — Claude Code via its CLAUDE.md conventions, Cursor through its rules system, and Kiro through its spec-driven workflow. When an AI assistant generates code, it references these codified rules, dramatically reducing the 'drift' that occurs across sessions.

Why This Matters Now

The timing is significant. AI coding assistants are experiencing explosive adoption. GitHub Copilot surpassed 1.8 million paid subscribers in early 2025. Anthropic's Claude Code launched to strong developer interest. Amazon's Kiro entered the market specifically targeting the consistency problem with its 'spec-driven development' approach. Cursor has become a favorite among indie developers and startups.

But adoption has outpaced best practices. Most teams are still figuring out how to integrate AI coding tools without sacrificing code quality. A 2024 GitClear study found that AI-assisted codebases showed measurably higher rates of 'code churn' — code that gets written and then quickly rewritten — suggesting that inconsistency is creating real downstream costs.

The open-source approach is particularly appealing because it remains tool-agnostic. As the AI coding landscape fragments across multiple assistants, a portable conventions layer prevents teams from being locked into one vendor's solution.

The Bigger Picture: From Prompts to Protocols

This project reflects a broader shift in how developers think about AI collaboration. The early era of AI coding was dominated by prompt engineering — crafting the perfect instruction to get the perfect output. That approach is giving way to something more systematic.

Teams are increasingly recognizing that effective AI-assisted development requires infrastructure, not just clever prompts. Convention files, architectural decision records, and structured context documents are becoming first-class artifacts in modern development workflows.

As one commenter on the project noted, 'We don't ask junior developers to just figure out our conventions by reading the codebase. Why would we expect AI to do that?'

What to Watch

The project is still in its early stages, but it addresses a pain point that virtually every team using AI coding tools has encountered. If it gains traction, expect the major AI coding platforms to incorporate similar conventions-first approaches natively.

Anthropic and Cursor have already moved in this direction with built-in rules and memory features. Amazon's Kiro launched with specification documents as a core concept. The trend is clear: the next frontier in AI-assisted coding isn't smarter models — it is smarter context.

For teams struggling with AI-generated inconsistency, the immediate takeaway is actionable: stop writing better prompts and start writing down your rules.