Strict Linting Dramatically Improves LLM Code Quality: TypeScript Go, Oxlint, and Oxfmt for the AI Era

A technical article by Christoph Nakazawa—Meta engineer and OpenClaw developer—titled “Fastest Frontend Tooling for Humans & AI” is gaining attention for its practical implications on working with AI coding agents. The central thesis: “Humans and LLMs both perform significantly better in codebases with fast feedback loops, strict guardrails, and strong local reasoning.”

Recommended Toolchain (2026)

TypeScript Go (tsgo)

A Go-language rewrite of TypeScript delivering up to 10x faster type checking. Nakazawa has deployed it across 20+ projects ranging from 1,000 to 1,000,000 lines, reporting that “tsgo detected type errors that the original JS implementation missed.”

Migration requires installing @typescript/native-preview and replacing tsc with tsgo.

ESLint → Oxlint

A Rust-based linter capable of running ESLint plugins directly through shims. The notable contribution is @nkzw/oxlint-config, which systematizes an opinionated approach to improving LLM-generated code quality:

• Errors only, no warnings: Warnings are easily ignored; remove them
• Strict, consistent code style: Enforce modern language features
• Bug prevention: Ban problematic patterns like instanceof; prohibit console.log and test.only in production

Prettier → Oxfmt

A drop-in Prettier replacement with import sorting and Tailwind CSS class sorting built in—no plugins required. Automatically tidies AI-generated code formatting diffs.

Controlled Experiment: Strict Guardrails Dramatically Improve GPT 5.2 Codex Output

The most compelling section is Nakazawa’s controlled experiment: the same UI framework migration task given to GPT 5.2 Codex under two conditions:

1. Empty repository (no guardrails)
2. Template with strict guardrails pre-configured (Oxlint + tsgo)

Result: the guardrailed condition produced “significantly fewer bugs and clearly superior results.”

This demonstrates that AI coding agent output quality depends not just on prompt quality but heavily on codebase quality and toolchain design.

Why Strict Rules Work for LLMs

When generating code, LLMs self-correct more effectively against clear constraints (errors) than ambiguous feedback (warnings). Warnings signal “this works either way,” while errors from type checking or static analysis require the agent to address them definitively.

With faster type checking (10x via tsgo), the agent’s feedback loop shortens, enabling more iterations within the same time window.

Practical Migration Guidance

The article includes specific prompt text for “ESLint→Oxlint migration” and “Prettier→Oxfmt migration” that can be passed directly to an AI agent—a deliberate practical design choice.

Related article:

• Unlock Claude Code’s 1M Token Context Window

Source: cpojer.net / Hacker News