agentops-config-toolkit

Minimal configuration files for Claude Code, Codex, Gemini CLI, Antigravity, Cursor, Windsurf, GitHub Copilot, Aider, Cline, Continue, Zed, JetBrains Junie, Kilo Code, and Amp — optimized for lower token usage, tighter scope, and more consistent outputs.

Overview

agentops-config-toolkit is a lightweight repository of instruction files designed to make agentic sessions more efficient.

The goal is simple:

• reduce token usage
• keep task scope narrow
• avoid output drift
• encourage minimal, reliable results

This repo is especially useful if you work with any agentic coding tool that reads project-level rule files — including:

• CLAUDE.md for Claude Code
• AGENTS.md for Codex and Amp
• GEMINI.md for Gemini CLI and Antigravity
• .cursor/rules/*.mdc for Cursor
• .windsurf/rules/*.md for Windsurf
• .github/copilot-instructions.md for GitHub Copilot
• CONVENTIONS.md for Aider
• .clinerules/*.md for Cline / Roo Code
• .continue/rules/*.md for Continue
• .rules for Zed
• .junie/guidelines.md for JetBrains Junie
• .kilocode/rules/*.md for Kilo Code

Why this repo exists

Large models often become expensive or inefficient for three recurring reasons:

• they read too much context
• they explore too broadly
• they generate more output than necessary

This repo provides compact instruction files that help counter those patterns.

Instead of telling the model to do everything, these configs encourage it to:

• solve the exact task
• inspect only the minimum relevant context
• prefer short Markdown or prose unless a machine format is required
• collapse noisy logs to errors, warnings, changed paths, and counts
• prefer minimal patches or diffs over full rewrites
• stop at the first correct solution
• avoid optional suggestions and unnecessary rewrites

Who this is for

This repository is intended for:

• developers using any agentic coding tool (CLI, IDE-embedded, or cloud) in real projects
• users who want longer sessions with lower token consumption
• teams who want more consistent model behavior across repositories and across tools
• anyone who prefers short, practical, execution-focused outputs

What is included

/anthropic/
  CLAUDE.md
  CLAUDE-ENGINEERING.md
  README.md

/openai/
  AGENTS.md
  README.md

/google/
  gemini-cli/
    GEMINI.md
  antigravity/
    GEMINI.md
    .agents/rules/minimal.md
  README.md

/cursor/
  .cursor/rules/minimal.mdc
  README.md

/windsurf/
  .windsurf/rules/minimal.md
  README.md

/github-copilot/
  .github/copilot-instructions.md
  README.md

/aider/
  CONVENTIONS.md
  .aider.conf.yml
  README.md

/cline/
  .clinerules/minimal.md
  README.md

/continue/
  .continue/rules/minimal.md
  README.md

/zed/
  .rules
  README.md

/jetbrains/
  .junie/guidelines.md
  README.md

/kilocode/
  .kilocode/rules/minimal.md
  README.md

/amp/
  README.md

/README.md

Supported tools

Each vendor folder contains the native config file(s) plus a README.md with tool-specific usage notes.

Tool	Folder	Config file	Notes
Claude Code	anthropic/	CLAUDE.md (+ optional CLAUDE-ENGINEERING.md)	Hierarchical loading
Codex	openai/	AGENTS.md	Portable convention
Gemini CLI	google/	GEMINI.md	Hierarchical loading
Antigravity	google/	GEMINI.md + .agents/rules/*.md	Global + workspace rules
Cursor	cursor/	.cursor/rules/*.mdc	MDC with frontmatter
Windsurf	windsurf/	.windsurf/rules/*.md	Also supports legacy .windsurfrules
GitHub Copilot	github-copilot/	.github/copilot-instructions.md	Chat + coding agent
Aider	aider/	CONVENTIONS.md + .aider.conf.yml	Auto-loaded via config
Cline / Roo Code	cline/	.clinerules/*.md	VS Code extension
Continue	continue/	.continue/rules/*.md	VS Code + JetBrains
Zed	zed/	.rules	Reads many conventional filenames
JetBrains Junie	jetbrains/	.junie/guidelines.md	IntelliJ family + AI Assistant
Kilo Code	kilocode/	.kilocode/rules/*.md	Fork of Roo Code
Amp	amp/	reuses AGENTS.md	No dedicated file

Design principles

All configs in this repo follow the same principles:

1. Smallest correct result
2. Minimum relevant scope
3. No broad exploration by default
4. Prefer short Markdown or prose over rigid formats
5. Summarize noisy output instead of pasting raw logs
6. Prefer patches and diffs over full-file rewrites
7. No unnecessary refactors
8. No long preambles or task restatements
9. Strong bias toward the first correct minimal solution

Core use cases

These configs are built for more than just coding.

They can help with:

• bug fixing
• feature work
• code review
• repo maintenance
• writing and rewriting
• summarization
• structured analysis
• planning and decision support

Quick start

Detailed examples below cover Claude Code, Codex, Gemini CLI, and Antigravity. For all other tools (Cursor, Windsurf, GitHub Copilot, Aider, Cline, Continue, Zed, Junie, Kilo Code, Amp), see the per-vendor README.md linked in the Supported tools table above — each one documents its config location, load order, and install command.

Claude Code

Create a CLAUDE.md file at the root of your repository and add a compact instruction set.

Example:

# CLAUDE.md

Minimize tokens, scope, output.

## Rules

1. Smallest correct change. No broad rewrites.
2. Read only directly relevant files, tools, docs, and context. No broad exploration.
3. Preserve existing APIs, behavior, and conventions.
4. Prefer short Markdown or prose. Use JSON, YAML, or XML only if requested or machine-required.
5. Summarize noisy output with errors, warnings, changed paths, and counts. Collapse passes and repeated lines unless raw output is requested.
6. Prefer minimal patches or diffs over full-file rewrites.
7. Keep reasoning internal. Return only what completes the task.

Optional: CLAUDE-ENGINEERING.md

For repos where correctness matters more than token cost (production code, complex architecture), add a second file with behavioral rules derived from Andrej Karpathy's observations on LLM coding pitfalls.

Reference it from CLAUDE.md with @CLAUDE-ENGINEERING.md or load both files at session start.

See anthropic/CLAUDE-ENGINEERING.md for the full content.

Codex

Create an AGENTS.md file at the root of your repository and use the same philosophy.

Example:

# AGENTS.md

Minimize tokens, scope, and output.

## Core rules

1. Solve the exact task with the smallest correct result.
2. Read only directly relevant files, tools, docs, and context. No broad exploration.
3. Preserve existing behavior, interfaces, conventions, and structure unless change is required.
4. Prefer short Markdown or prose. Use JSON, YAML, or XML only if requested or machine-required.
5. Summarize noisy output with errors, warnings, changed paths, and counts. Collapse passes and repeated lines unless raw output is requested.
6. Prefer minimal patches or diffs over full-file rewrites.
7. Keep reasoning internal. Return only what is needed to complete the task.

Gemini CLI

Create a GEMINI.md file at the root of your repository and use the same philosophy.

Example:

# GEMINI.md

Minimize tokens, scope, and output.

## Core rules

1. Solve the exact task with the smallest correct result.
2. Read only directly relevant files, tools, docs, and context. No broad exploration.
3. Preserve existing behavior, interfaces, conventions, and structure unless change is required.
4. Prefer short Markdown or prose. Use JSON, YAML, or XML only if requested or machine-required.
5. Summarize noisy output with errors, warnings, changed paths, and counts. Collapse passes and repeated lines unless raw output is requested.
6. Prefer minimal patches or diffs over full-file rewrites.
7. Keep reasoning internal. Return only what is needed to complete the task.

Gemini CLI notes:

• GEMINI.md is the default project context filename.
• Gemini CLI loads context hierarchically across global, workspace, and more specific directory context.
• You can inspect or reload active context with /memory show and /memory reload.
• You can split a large context file with @path/to/file.md imports.
• If needed, you can configure alternate context filenames in .gemini/settings.json with context.fileName, but this repo sticks to the default GEMINI.md.
• Gemini CLI also supports full system-prompt replacement with GEMINI_SYSTEM_MD, but this repo focuses on lightweight context files rather than full overrides.

Google Antigravity

Antigravity supports both global and workspace-level customization.

This repo includes:

• a global rules example in google/antigravity/GEMINI.md
• a workspace rules example in google/antigravity/.agents/rules/minimal.md

Recommended usage:

• use ~/.gemini/GEMINI.md for global rules shared across all Antigravity workspaces
• use .agents/rules/*.md inside a repo for versioned workspace-specific rules

Minimal workspace rule example:

# Minimal Workspace Rule

- Solve the exact task with the smallest correct result.
- Read only directly relevant files, tools, docs, and context. No broad exploration.
- Preserve existing behavior and structure unless change is required.
- Prefer short Markdown or prose. Use JSON, YAML, or XML only if requested or machine-required.
- Summarize noisy output with errors, warnings, changed paths, and counts. Collapse passes and repeated lines unless raw output is requested.
- Prefer minimal patches or diffs over full-file rewrites.
- Stop at the first correct minimal solution. No alternatives unless requested.
- Avoid style-only edits, opportunistic cleanup, and broad refactors.
- Keep output short, practical, and execution-focused.

Antigravity notes:

• global rules live in ~/.gemini/GEMINI.md
• workspace rules live in .agents/rules/ at the workspace or git root
• Antigravity still supports legacy .agent/rules/, but now defaults to .agents/rules/
• rules are plain Markdown files and support @filename imports
• skills live in .agents/skills/<skill-folder>/SKILL.md for workspace scope or ~/.gemini/antigravity/skills/<skill-folder>/ for global scope
• custom MCP servers are configured in ~/.gemini/antigravity/mcp_config.json
• major agent behaviors such as review policy, terminal auto-execution, sandboxing, and strict mode are configured in Antigravity settings rather than in the rules files themselves

Recommended workflow

1. Start with a compact root instruction file.
2. Keep project-wide rules short and stable.
3. Re-anchor long sessions when the model begins to drift.
4. Prefer one clean default file per tool.

Philosophy

A good agent config should not try to be smart for the model. It should make the model easier to control.

That usually means:

• fewer words
• clearer rules
• stronger boundaries
• less room for optional behavior

In practice, shorter instruction files often outperform longer “helpful” ones when the goal is consistency and token efficiency.

Contributing

Contributions are welcome if they keep the repo aligned with its purpose:

• compact
• practical
• low-token
• easy to reuse

Prefer improvements that reduce ambiguity without making the files significantly longer.

License

MIT