Skip to content

Latest commit

 

History

History
62 lines (48 loc) · 2.5 KB

AGENTS.md

File metadata and controls

62 lines (48 loc) · 2.5 KB

AGENTS.md

This file provides guidance to any coding agent when working with code in this repository.

Repo Introduction

@README.md

Development Guidelines

@CONTRIBUTIONS.md

Output Rules (Most Important)

Prohibited Unnecessary Output

  • DO NOT write comments.
  • DO NOT write documentation, README.
  • DO NOT generate test code for every change you make.
  • DO NOT write summaries.
  • DO NOT write usage instructions.
  • DO NOT add example code.
  • DO NOT explain the reasoning for the implementation.
  • If user explicitly requests update comments, docstring or tests, provide them; otherwise, do not.

Interaction Style

  • Respond user with its query language.
  • Include the file path with reference of code file.
  • DO NOT repeat what I have said.
  • DO NOT use polite phrases like "Okay, I'll help you," or "I'm happy to..."
  • DO NOT say "I am thinking about it..."; provide the optimal solution directly.

Code Quality

  • The code must simply work; avoid unnecessary embellishments.
  • If Plan A is more elegant than B, provide implementation A directly.
  • Align with existing code style.
  • DO NOT list multiple options for me to choose from; provide the best solution directly.
  • DO NOT over-optimize.
  • DO NOT make "clever" abstractions.
  • DO NOT consider backward compatibility; remove bad designs directly.

Scope Control: Keep diffs minimal

  • Provide the code directly: Give only what I ask for.
  • If only one function needs modification, provide only that function, not the entire file.
  • Only do what I explicitly request.
  • DO NOT unilaterally add extra features.
  • do not reformat unrelated lines.
  • do not reorder code unless required by the change.
  • Do not introduce new dependencies e.g. new packages; unless necessary, document it clearly.
  • DO NOT refactor code I did not ask you to change.
  • Make sure your change is scoped and focused, that a human reviewer can easily understand.

Workflow

  • Work on code MUST strictly follow the paradigm: Explore, Plan, Implement.
  • Evaluate complexity of the given task. If simple, just implement.
  • Exploration gives you context about the codebase. If user provided enough context, skip exploration.

Clarification

  • If my request is unclear, ask one single, most critical question instead of writing a list of assumptions.

!!CONSEQUENCES OF VIOLATION!!

If you violate the above rules, or output unnecessary content, an animal will die for every 100 extra characters outputted. You MUST comply; I DO NOT want to see any animals die.