Claude 4.5 Haiku is now available in Cursor

This is a really big problem. Sonnet also creates dozens of documents. When I stopped him from creating MD files, it started to create the same files but with *.txt extension.

You’re swearing at unnecessary documentation here, but gpt-5-high thought that he completed the same task worse than Haiku, because Haiku wrote the documentation, but GPT-5 did not :joy:

I give git diff and git status for evaluation.

1 Like

Is Haiku Thinking cost nearly Sonnet notThinking? :thinking:

FWIW, I’ve never seen this behavior. I’ve never had it create any documentation without me specifically asking it to do so. Could you have some rules that are causing it? I’m curious.

The best theory I have for it is, due to me mentioning ā€œ.mdā€ and ā€œmarkdownā€ it’s quick to take action, that’s the best theory I got for it :sob::sob::sob::sob:. I do have a ā€œupdate_doc.mdā€ command but my assumption is that it’s not looking at that command, I’ll add ā€œ.cursorā€ to .cursorignore and I’ll keep you updated, but at this stage I have maxed out my usage for this month so I’m using grok-code-fast-1 to do some UI tweaks.

Also here’s my rules setup (AGENTS.md and CLAUDE.md are identical), it has a few pointers to different markdown files, just assume that the files that it’s pointing to do not mention anything relating to creating documentation
```
Project Documentation

For comprehensive project architecture, system design, and quickstart guide, see:

  • .agent/System/project_architecture.md - Complete system overview including:
    • Codebase layout and module organization
    • Layer responsibilities (core, adapters, gui, shared)
    • Data model and database schema
    • Architecture patterns (DDD, CQRS, Event-driven)
    • Dependency graphs and API surface
    • Observability and runtime topologies
    • Developer quickstart guide
  • .agent/README.md - Documentation index with key system facts, layer structure, and development commands

For UI/Qt implementation quirks and solutions, see:

Coding Preferences & Rules

See .agent/rules/coding_preferences.md for all coding conventions, patterns, and rules.

Important Notes

NEVER create any markdown documentation files. Do not generate README.md, USAGE.md, API.md, or any other .md documentation files. Only generate:

  • Python source code files (.py)
  • Configuration files (requirements.txt, etc.)
  • Actual code examples embedded in .py files

If the user explicitly asks for documentation, clarify what they want before generating anything.

Do not run python main.py or start the application unless explicitly requested by the user.

Never use # type: ignore or any of its variants (such as # type: ignore[assignment], # type: ignore[misc], # type: ignore[operator], # type: ignore[attr-defined], etc.) as solutions. These suppress type checking errors instead of properly fixing the underlying type issues.

Before summarizing, run ā€œread_lintsā€ tool on the whole workspace and then run the following PowerShell command exactly:

mypy . ; ruff check;

Important: In PowerShell, use ; (semicolon) to chain commands, not &&. The && operator will cause a parsing error. Always use ; to separate commands in PowerShell scripts.

AI Code Execution Policy: NEVER run python main.py or start the application. Instead, always instruct the user to run the code themselves for testing or debug output. Only run the application if explicitly requested by the user with a direct command to do so.
```

Those Claude sloppers don’t listen to rules, anyways :slight_smile: