AI Configuration

Harnesses

Harness	Global Context File	Project Context File	System Prompt Override	Skills Directories	Source
claude	`~/.claude/CLAUDE.md`	`CLAUDE.md` in project root	—	custom slash commands	docs
codex	`~/.codex/AGENTS.md`	`AGENTS.md` in project root	—	—	docs
gemini	`~/.gemini/GEMINI.md`	`GEMINI.md` in workspace	`GEMINI_SYSTEM_MD` env var	`~/.gemini/skills/`, `~/.agents/skills/`	context, skills, system
qwen	`~/.qwen/QWEN.md`	`QWEN.md` in workspace	`QWEN_SYSTEM_MD` env var	`~/.qwen/skills/`	docs
opencode	`~/.config/opencode/AGENTS.md`	`AGENTS.md` in project root	`prompt` field in agent config	`~/.claude/skills/` (fallback)	docs
kilo	`~/.config/kilo/AGENTS.md`	`AGENTS.md` in project root	`prompt` field in agent config	`~/.kilocode/skills/`	docs, skills
amp	`~/.config/amp/AGENTS.md`, `~/.config/AGENTS.md`	`AGENTS.md` in cwd, parent dirs, subtrees	—	`~/.config/agents/skills/`, `~/.config/amp/skills/`, `.agents/skills/`, `.claude/skills/`, `~/.claude/skills/`	docs

Master files (symlinked to all harnesses):

Context file: ~/ai/AGENTS.md
Skills: ~/ai/skills/ (29 skills)

Verified Details

Codex CLI (source):

Global scope: ~/.codex/AGENTS.md or ~/.codex/AGENTS.override.md
Project scope: Walks from project root to cwd, reading AGENTS.md or AGENTS.override.md
Fallback filenames configurable via project_doc_fallback_filenames

Gemini CLI (context, skills, system prompt):

Context hierarchy: Global (~/.gemini/GEMINI.md) → Workspace → JIT (auto-scans when accessing dirs)
Custom filename via context.fileName setting: {"context": {"fileName": ["AGENTS.md", "GEMINI.md"]}}
Skills precedence: Workspace > User > Extension; .agents/skills/ takes precedence over .gemini/skills/
System prompt override: See System Prompt Override section

Qwen Code (source):

Forked from Gemini CLI, uses similar context system
Default context file: QWEN.md (configurable via context.fileName)
Skills from .qwen/skills/ (workspace) and ~/.qwen/skills/ (user)
System prompt override: See System Prompt Override section

OpenCode (source):

Global: ~/.config/opencode/AGENTS.md
Project: AGENTS.md in project root
Fallback: CLAUDE.md (unless OPENCODE_DISABLE_CLAUDE_CODE=1)
Skills: ~/.claude/skills/ (unless OPENCODE_DISABLE_CLAUDE_CODE_SKILLS=1)

Amp (source):

Global: $HOME/.config/amp/AGENTS.md or $HOME/.config/AGENTS.md
Project: AGENTS.md in cwd, parent dirs (up to $HOME), subtrees
Fallback: If no AGENTS.md, reads AGENT.md or CLAUDE.md
Skills precedence (first wins): ~/.config/agents/skills/ → ~/.config/amp/skills/ → .agents/skills/ → .claude/skills/ → ~/.claude/skills/

Claude Code (source):

Uses CLAUDE.md at project root and parent directories
Skills via custom slash commands

Kilo (AGENTS.md, skills):

Forked from OpenCode, uses similar configuration
Global: ~/.config/kilo/AGENTS.md
Project: AGENTS.md in project root
Skills: ~/.kilocode/skills/ (global), .kilocode/skills/ (project)
System prompt override: prompt field in agent config (same as OpenCode)

Project-Local Configuration

Each harness looks for configuration files in the project directory. Place these in your project root:

Harness	Context File	Skills Directory	Config File
Claude	`CLAUDE.md`	`.claude/commands/` (slash commands)	—
Codex	`AGENTS.md`	—	—
Gemini	`GEMINI.md`	`.gemini/skills/`	`.gemini/settings.json`
Qwen	`QWEN.md`	`.qwen/skills/`	`.qwen/settings.json`
OpenCode	`AGENTS.md`	—	`opencode.json`
Kilo	`AGENTS.md`	`.kilocode/skills/`	`.kilocode/launchConfig.json`
Amp	`AGENTS.md`	`.agents/skills/`	—

Precedence (workspace > user > built-in):

Harness	Context Hierarchy	Skills Hierarchy
Gemini	Project `.gemini/GEMINI.md` > `~/.gemini/GEMINI.md` > JIT	`.agents/skills/` > `.gemini/skills/` > `~/.gemini/skills/`
Qwen	Project `.qwen/QWEN.md` > `~/.qwen/QWEN.md`	`.qwen/skills/` > `~/.qwen/skills/`
Codex	Walks project root → cwd reading `AGENTS.md`	—
OpenCode	`AGENTS.md` in project root > `~/.config/opencode/AGENTS.md`	`~/.claude/skills/` (fallback only)
Kilo	`AGENTS.md` in project root > `~/.config/kilo/AGENTS.md`	`.kilocode/skills/` > `~/.kilocode/skills/`
Amp	cwd → parent dirs (to `$HOME`) → subtrees	`.agents/skills/` > `.claude/skills/` > `~/.claude/skills/`

Example project structure:

my-project/
├── AGENTS.md              # Context for Codex, OpenCode, Kilo, Amp
├── GEMINI.md              # Context for Gemini (or configure to read AGENTS.md)
├── QWEN.md                # Context for Qwen
├── CLAUDE.md              # Context for Claude
├── .gemini/
│   ├── settings.json      # Gemini project config
│   └── skills/            # Gemini project skills
├── .qwen/
│   ├── settings.json      # Qwen project config
│   └── skills/            # Qwen project skills
├── .kilocode/
│   ├── launchConfig.json  # Kilo project config
│   └── skills/            # Kilo project skills
├── .agents/
│   └── skills/            # Amp project skills
└── opencode.json          # OpenCode project config

Context Files vs System Prompts

Context files (AGENTS.md, GEMINI.md, CLAUDE.md, etc.) are user-supplied instructions that get appended to the harness's built-in system prompt. They add project-specific context, behavioral preferences, and task guidance.

System prompts are the harness's built-in instructions ("You are a helpful assistant..."). Some harnesses allow replacing this entirely via environment variables:

Harness	Context File (appended)	System Prompt Override (replaces)
Gemini	GEMINI.md	`GEMINI_SYSTEM_MD` env var
Qwen	QWEN.md	`QWEN_SYSTEM_MD` env var
OpenCode	AGENTS.md	`prompt` field in agent config
Kilo	AGENTS.md	`prompt` field in agent config
Claude	CLAUDE.md	—
Codex	AGENTS.md	—
Amp	AGENTS.md	—

When to override the system prompt:

You want complete control over agent behavior
The built-in prompt conflicts with your workflow
You're building a specialized tool on top of the harness

System Prompt Override

Some harnesses allow replacing their built-in system prompt entirely. This is different from context files (AGENTS.md, GEMINI.md) which are appended to the prompt.

Gemini CLI (docs):

Set GEMINI_SYSTEM_MD environment variable:

# Use .gemini/system.md in project (fixed path)
GEMINI_SYSTEM_MD=true gemini

# Use custom file (variable holds the path)
GEMINI_SYSTEM_MD=/path/to/my-system.md gemini

# Disable override (use built-in)
GEMINI_SYSTEM_MD=false gemini

Value	Behavior
`true` or `1`	Uses `.gemini/system.md` in project
`/path/to/file.md`	Uses that file
`false` or `0`	Uses built-in system prompt

Can also persist in .gemini/.env:

GEMINI_SYSTEM_MD=1

When active, Gemini shows |⌐■_■| indicator in the UI.

Qwen Code:

Forked from Gemini CLI - same mechanism with QWEN_SYSTEM_MD:

QWEN_SYSTEM_MD=/path/to/system.md qwen

OpenCode (docs):

OpenCode uses custom agents to replace the system prompt. Define an agent in opencode.json:

{
  "agent": {
    "interactive": {
      "prompt": "You are a helpful assistant...",
      "model": "anthropic/claude-sonnet-4"
    }
  }
}

The prompt field completely replaces the built-in system prompt for that agent.

Prompts

Interactive Agents (`~/ai/prompts/interactive_agents/`)

Top-level agents - users invoke these directly in the coding harness.

Agent	Role	Temperature
interactive	Default entry point, user-facing	0.5
autonomous	Execution engine (called by interactive)	0.2

Interactive handles all horizons:

Horizon	Criteria	Interactive's Action
Trivial	< 2 min, obvious	Just do it
Small	Fits in head, clear path	Brief plan → execute
Complex	Multiple unknowns, design needed	Design doc → hand off to autonomous

Autonomous is not an entry point - it receives plans from interactive and executes without user interaction.

Workflow:

User → interactive
         │
         ├─► trivial ──► just do it
         │
         ├─► small ──► plan briefly → execute
         │
         └─► complex ──► design doc → hand off to autonomous
                                          │
                                          └─► planner → executor → done

Other:

Agent	Description
executor	Executes implementation plans
minimal	Minimal agent template

Subagents (`~/ai/prompts/subagents/`)

Internal agents - called by interactive/autonomous agents, not meant for direct user invocation.

Research (spawned in parallel):

Agent	Purpose
codebase-locator	Find WHERE files are
codebase-analyzer	Understand HOW code works
pattern-finder	Find existing patterns
artifact-searcher	Search past work

Execution (spawned sequentially):

Agent	Purpose
planner	Creates implementation plan
executor	Orchestrates implementer + reviewer
implementer	Writes code
reviewer	Verifies correctness

Utility:

Agent	Description
project-initializer	Initializes new projects
code-compliance-reviewer	Reviews code against standards
code-smell-detector	Detects code smells

Worker Agents (`~/ai/prompts/worker_agents/`)

Specialized agents with extra local tools - usable both interactively AND autonomously.

Workers are domain-specific agents that can be:

Loaded interactively when user wants careful, hands-on work
Used autonomously for scheduled tasks, regular upkeep, or audits

Worker	Interactive Use	Autonomous Use
librarian	Careful work on Zotero library	Regular upkeep/audits
test_engineer	Design test strategy with user	Run test coverage audits
documentation_specialist	Plan docs structure	Generate missing docs

Each worker has prompt.md + example_tasks/ showing autonomous work they can complete.

MCP Servers

Name	Command	Description	Link
serena	`uvx --from git+https://github.com/oraios/serena serena start-mcp-server --project-from-cwd`	Code intelligence, symbol navigation	https://github.com/oraios/serena
morph	`npx -y @morphllm/morphmcp`	Fast code edits via Morph LLM	https://github.com/morph-llm/morphmcp
kindly	`uvx --from git+https://github.com/Shelpuk-AI-Technology-Consulting/kindly-web-search-mcp-server kindly-web-search-mcp-server start-mcp-server`	Web search	https://github.com/Shelpuk-AI-Technology-Consulting/kindly-web-search-mcp-server
context7	`npx -y @upstash/context7-mcp`	Documentation search (llms.txt)	https://github.com/upstash/context7
cut-copy-paste-mcp	`npx -y @fastmcp-me/cut-copy-paste-mcp`	Clipboard operations	https://github.com/fastmcp-me/cut-copy-paste-mcp

Adding MCP Servers by Harness

Claude Code (docs):

# HTTP server
claude mcp add --transport http <name> <url>

# Stdio server
claude mcp add --transport stdio <name> -- <command> [args...]

# List servers
claude mcp list

Config files: ~/.claude.json (user), .mcp.json (project)

OpenCode (docs): Edit opencode.json directly (no CLI command):

{
  "mcp": {
    "my-server": {
      "type": "local",
      "command": ["npx", "-y", "my-mcp-server"],
      "enabled": true
    },
    "remote-server": {
      "type": "remote",
      "url": "https://mcp.example.com/mcp"
    }
  }
}

Config files: ~/.config/opencode/opencode.json (global), opencode.json (project)

Codex CLI (docs): Edit ~/.codex/config.toml:

[mcp_servers.my-server]
command = "npx"
args = ["-y", "my-mcp-server"]
env = { "API_KEY" = "value" }

[mcp_servers.remote-server]
url = "https://mcp.example.com/mcp"

Config file: ~/.codex/config.toml

Gemini CLI (docs):

# Add server
gemini mcp add --transport http <name> <url>
gemini mcp add -s user <name> -- <command> [args...]

Or edit settings.json:

{
  "mcpServers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "my-mcp-server"]
    },
    "remote-server": {
      "httpUrl": "https://mcp.example.com/mcp"
    }
  }
}

Config files: ~/.gemini/settings.json (user), .gemini/settings.json (project)

Qwen Code (docs): Forked from Gemini CLI - same configuration pattern:

qwen mcp add --transport http <name> <url>

Config files: ~/.qwen/settings.json (user), .qwen/settings.json (project)

Kilo: Edit ~/.kilocode/cli/global/settings/mcp_settings.json:

{
  "mcpServers": {
    "my-server": {
      "command": "npx",
      "args": ["-y", "my-mcp-server"],
      "env": {},
      "disabled": false
    }
  }
}

Amp (docs): Amp reads MCP config from multiple locations. Check ~/.config/amp/ or harness documentation for current format.

OpenCode Plugins

Config: ~/ai/opencode/opencode.json

Auth Plugins

Plugin	Purpose	Link
`opencode-antigravity-auth@latest`	Antigravity provider auth (Gemini, Claude via Google)	https://github.com/NoeFabris/opencode-antigravity-auth
`opencode-qwencode-auth`	Qwen OAuth authentication	anomalyco/opencode#11557
`opencode-openai-codex-auth`	OpenAI Codex auth (GPT-5.x Codex models)	https://github.com/numman-ali/opencode-openai-codex-auth

Utility Plugins

Plugin	Purpose	Link
`cc-safety-net`	Safety net for destructive git/fs commands	https://github.com/kenryu42/claude-code-safety-net
`@azumag/opencode-rate-limit-fallback`	Auto fallback on rate limit (429)	https://npmjs.com/package/@azumag/opencode-rate-limit-fallback
`@ramtinj95/opencode-tokenscope`	Token usage stats and cost tracking	https://github.com/ramtinJ95/opencode-tokenscope

Name		Name	Last commit message	Last commit date
Latest commit History 20 Commits
opencode		opencode
openrouter		openrouter
prompts		prompts
skills		skills
.gitignore		.gitignore
AGENTS.md		AGENTS.md
README.md		README.md
justfile		justfile

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

AI Configuration

Harnesses

Verified Details

Project-Local Configuration

Context Files vs System Prompts

System Prompt Override

Prompts

Interactive Agents (`~/ai/prompts/interactive_agents/`)

Subagents (`~/ai/prompts/subagents/`)

Worker Agents (`~/ai/prompts/worker_agents/`)

MCP Servers

Adding MCP Servers by Harness

OpenCode Plugins

Auth Plugins

Utility Plugins

About

Uh oh!

Releases

Packages

Contributors 2

Uh oh!

Languages

dzackgarza/ai

Folders and files

Latest commit

History

Repository files navigation

AI Configuration

Harnesses

Verified Details

Project-Local Configuration

Context Files vs System Prompts

System Prompt Override

Prompts

Interactive Agents (~/ai/prompts/interactive_agents/)

Subagents (~/ai/prompts/subagents/)

Worker Agents (~/ai/prompts/worker_agents/)

MCP Servers

Adding MCP Servers by Harness

OpenCode Plugins

Auth Plugins

Utility Plugins

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Contributors 2

Uh oh!

Languages

Interactive Agents (`~/ai/prompts/interactive_agents/`)

Subagents (`~/ai/prompts/subagents/`)

Worker Agents (`~/ai/prompts/worker_agents/`)

Packages