vibecoding-assistant

Vibecoding Assistant

This skill provides workflows and guidance for AI-assisted software development, optimized for OpenClaw agents.

Core Workflow

Follow this loop for vibecoding:

Intent → Explore → Plan → Build → Review → Iterate → Ship

1. Explore (Understand the Task)

Before generating code, explore the codebase to understand:

• Existing patterns and conventions
• File structure and architecture
• Dependencies and configuration

Prompt pattern:

"Explore the codebase to understand [specific area]. Look for [patterns/conventions]."

2. Plan (Define the Approach)

Always ask for a plan before implementation. This catches architectural issues early.

Prompt pattern:

"Create a step-by-step plan for [feature]. Include: files to modify, interfaces to define, edge cases to handle. Don't write code yet—describe the approach."

3. Build (Iterative Implementation)

Build in small, focused iterations:

• One feature or file at a time
• Run tests after each piece
• Review before proceeding

4. Review (Quality Gates)

Review every change before proceeding:

• Architecture and patterns
• Error handling
• Edge cases
• Tests

5. Ship (Deployment)

• Add environment setup instructions
• Configure deployment
• Document the changes

OpenClaw Configuration

Core Files

OpenClaw uses these files in each workspace:

File	Purpose
SOUL.md	Personality and communication style
IDENTITY.md	Agent's specific purpose and role
MEMORY.md	Long-term facts and accumulated knowledge
HEARTBEAT.md	Scheduled, proactive behaviors
TOOLS.md	User-defined conventions for tool usage
AGENTS.md	Core operational rules

Gateway

The Gateway is a WebSocket server (default: localhost:18789):

• Routes messages from channels to agents
• Handles session management
• Enforces Lane Queue (serial execution by default)

Start Gateway:

openclaw gateway

Agent Configuration

Configure agents in ~/.openclaw/openclaw.json:

{
  "agents": {
    "defaults": {
      "workspace": "~/.openclaw/workspace",
      "model": { "primary": "anthropic/claude-sonnet-4" },
      "thinkingDefault": "low",
      "timeoutSeconds": 600,
      "sandbox": { "mode": "non-main", "scope": "agent" }
    },
    "list": [
      {
        "id": "coder",
        "workspace": "~/.openclaw/workspace-coder",
        "model": { "primary": "anthropic/claude-sonnet-4" },
        "theme": "helpful coding assistant"
      }
    ]
  },
  "bindings": [
    { "agentId": "coder", "match": { "channel": "cli" } }
  ]
}

Multi-Agent Setup

For multi-agent workflows:

{
  "agents": {
    "list": [
      {
        "id": "lead",
        "role": "coordinate tasks, assign work"
      },
      {
        "id": "researcher",
        "role": "gather information, validate facts"
      },
      {
        "id": "coder",
        "role": "implement features, fix bugs"
      },
      {
        "id": "reviewer",
        "role": "quality checks, review output"
      }
    ]
  }
}

Policy Stack (7 Layers)

OpenClaw enforces permissions through 7 layers:

1. Profile: Auth profile permissions
2. Provider: Model provider support
3. Global: Global enable/disable
4. Channel: Channel-specific rules
5. Workspace: Workspace boundaries
6. Agent: Agent-specific rules
7. Tool: Individual tool permissions

Sandbox Modes

• all: Sandbox everything
• non-main: Sandbox non-main sessions
• off: Full host access

Skills in OpenClaw

Skills are discovered at runtime and injected based on eligibility:

• Dependencies: Binary availability checked
• OS Compatibility: Platform-specific skills
• Context: Relevant skills loaded per turn

Agent-to-Agent Communication

Use sessions_spawn tool for recursive spawning:

• Parent agent spawns child agents for subtasks
• Child agents have isolated workspaces
• Results aggregated back to parent

Prompting Best Practices

Structure Prompts Clearly

• Be specific about requirements
• Include examples of expected input/output
• Define constraints and edge cases
• Specify non-goals to avoid scope creep

Example Prompt Structure

## Task
[Clear description of what to build]

## Context
[Relevant background about the codebase]

## Requirements
- [Specific requirement 1]
- [Specific requirement 2]

## Constraints
- [Performance requirements]
- [Security requirements]
- [Compatibility requirements]

## Examples
[Show example input/output if applicable]

Chained Prompting

For complex tasks, break into sequential prompts:

1. "Create the interface/types for [feature]"
2. "Implement [component A] using the interfaces"
3. "Implement [component B] that uses A"
4. "Add tests for [feature]"

Debugging Prompts

When stuck, prompt for systematic analysis:

Generate a diagnostic report:
1. List all files involved in [feature]
2. Explain the role of each file
3. Identify why [error] is occurring
4. Propose 3 different debugging approaches

OpenClaw-Specific Tips

Making Agents Feel "Alive"

OpenClaw agents can feel proactive using:

• Heartbeats: Timer-triggered turns (e.g., every 30 minutes)
• Cronjobs: Scheduled tasks
• Webhooks: External event triggers
• Agent-to-Agent: Inter-agent messaging

Workspace Isolation

Each agent has its own workspace:

• Separate file systems
• Independent memory
• Isolated tool access

Running Multiple Agents

1. Define agents in openclaw.json
2. Set up bindings to route messages
3. Use agent IDs to reference specific agents

References

For detailed guidance, see:

• openclaw-config skill - OpenCode/OpenClaw configuration
• prompts.md - Prompt templates
• debugging.md - Debug patterns with real error traces

Expanded References

Real project structures:

• node-express.md - Express.js project structure
• react-nextjs.md - Next.js App Router patterns
• python-fastapi.md - FastAPI clean architecture
• go-clean-arch.md - Go clean architecture

Common patterns with real code:

• error-handling.md - Production error handling
• state-management.md - React Query, Zustand, forms
• api-design.md - REST API best practices

Anti-patterns with fixes:

• god-files.md - Breaking up large files
• circular-deps.md - Fixing circular imports
• spaghetti-services.md - Layered architecture

Quick references:

• 00-ask-first.md - What to ask before coding
• 01-debug-errors.md - Common errors quick fix
• 02-review-checklist.md - Code review checklist

Real prompts that worked:

• prompts-history.md - Prompt patterns from real sessions

Skill optimization:

• skill-loading.md - Best practices for loading skills efficiently
• skill_indexer.py - Parses skills.yaml for fast lookup
• skill_loader.py - Just-in-time content loading with caching
• skills.yaml - Lightweight skill index for fast selection

Documentation:

• docs/architecture.md - Conceptual architecture and design
• docs/concept.md - Design reasoning and metaphors
• docs/api-reference.md - Full API documentation
• docs/integration-guide.md - OpenCode/OpenClaw integration
• docs/examples.md - Real usage examples
• docs/troubleshooting.md - Common issues and solutions