LLM Code Quality: The Agent.md Solution for AI-Assisted Development

Fabien Sanglard, a well-known developer, has shared his approach to improving the quality of LLM-generated code. His method centers on a simple file called agent.md, which is automatically loaded by coding harnesses at the start of a session. This file contains a set of rules that 'super fine-tune' the model's output, addressing the common problems of spaghetti code, missing comments, and poor structure.

Sanglard's journey with LLMs began in mid-2025, with disappointing results on his Rust project libadbmdns. However, by January 2026, he saw significant improvements, though the code quality remained below production standards. The breakthrough came in March 2026 when he started using agentic IDEs like Antigravity and VS Code's Claude Code plugin, which allowed him to iterate on the generated code. This is when he realized the need for a persistent set of instructions to avoid repeating himself in every session.

Key Rules for Better Code

Sanglard's agent.md is a comprehensive list of rules, each targeting a specific quality issue. Some of the most impactful rules include:

Brief communication: For comments and commit messages, use as few words as possible. Avoid superlatives and praise; focus on the cold, hard truth.
Avoid magic numbers: Extract recurring values into descriptive constants or enums. If a value comes from a spec (like HTTP 200 OK), always use a constant.
Reduce indentation: Avoid the 'Arrow Anti-Pattern' by using early returns and continue statements.
Short function names: Keep them under 30 characters.
Use enums over booleans: For function parameters, this improves readability and type safety.
Add breathing room: Use empty lines between logical blocks of code.
Explain the 'what' and 'why': Add small, to-the-point comments, using examples or ASCII drawings for complex systems.
Respect visibility: Keep fields and functions private unless absolutely necessary. Prompt the user before changing access modifiers.
Program to levels of abstraction: Encapsulate low-level mechanics (e.g., raw I/O) in dedicated drivers, exposing clean high-level APIs.
Minimize changes: Don't touch unrelated code blocks, and keep the diff small when implementing features.
Follow commit message rules: Use the 7 rules from Chris Beams, including a 50-character subject line and imperative mood.
Test-first for bug fixes: Write a failing test first, then the fix, and observe the test pass.

These rules address the most common pain points in LLM-generated code. Sanglard notes that while this trick has improved code quality, it's not a magic bullet. You still need to read and verify the code, as LLMs constantly hallucinate. However, the focus shifts from style to architecture and design.

continue reading below...

Dealing with Context Dilution

One of the biggest challenges with LLMs is context dilution, also known as 'Lost in the Middle.' As the context grows, the model pays less attention to instructions in the middle. Sanglard suggests two ways to minimize this:

Keep the context short: Start a new session per feature.
Explicitly ask the harness to reload agent.md: Saying 'Reload agent.md' when you see code quality dropping can help.

These strategies are practical for developers who rely on long coding sessions.

Auto-Updating agent.md

Updating agent.md doesn't require manual editing. Sanglard recommends asking the agent to update the file itself when you find yourself repeating a suggestion. This creates a feedback loop where the agent learns from your corrections, making future sessions more efficient.

This approach aligns with the broader trend of using LLMs in development, as seen in tools like Abacus.AI's AppLLM and Desktop, which offer browser-based development and multiple modes for coding. The industry is moving towards more integrated AI assistance, and having a structured way to guide the AI is crucial.

Why This Matters

Sanglard's agent.md is a simple yet powerful tool for anyone using LLMs for coding. It addresses the gap between impressive demos and production-ready code. By codifying style preferences, developers can save time and effort, focusing on higher-level design decisions.

This is especially relevant as AI-assisted coding becomes more prevalent. The challenges of code quality, context dilution, and maintaining standards are universal. Sanglard's approach offers a practical solution that can be adopted by individual developers and teams alike.

As the AI landscape evolves, tools like agent.md will become standard practice. They bridge the gap between the raw capabilities of LLMs and the meticulous standards of professional software development.