Opencode Workflows

A collection of Opencode-based command templates, global agent prompts, and workflow patterns for building sophisticated command-driven projects.

Start Here: Opencode Configurator

The Opencode Configurator (opencode-configurator/) is a meta-configuration system that makes setting up OpenCode effortless. Just ask naturally:

• "Is there a plugin for reducing token usage?"
• "Set up permissions so destructive commands require approval"
• "Create a /review command that analyzes code without making changes"
• "I want to build a skill for working with our internal API"
• "Find an MCP server for filesystem access"

The configurator agent draws on seven specialized skills:

Skill	What it does
plugin-installer	Find and install community plugins, maintain a local catalog
opencode-config	Edit opencode.json with guided setup for models, permissions, providers
command-creator	Build custom /slash commands through interactive Q&A
skill-creator	Scaffold new skills with proper structure, scripts, and references
agent-architect	Design agents with research-backed prompt engineering patterns
mcp-installer	Find, install, and configure Model Context Protocol (MCP) servers
model-researcher	Research and configure new/custom AI models not yet in models.dev, with verified specifications

Installation: Drop the skill/ and agent/ folders into ~/.config/opencode/.

See opencode-configurator/README.md for full details.

---

Create OpenCode Plugin

The Create OpenCode Plugin (create-opencode-plugin/) is a workflow bundle for AI-assisted plugin development in OpenCode. Describe what you want your plugin to do, and the workflow guides you through design, testing, and publishing.

Build custom tools, event handlers, tool interceptors, LLM parameter modifiers, authentication flows, and UI notifications.

Installation: Use this bundle inside a cloned OpenCode source repo.

Usage: Run /create-plugin [your idea].

See create-opencode-plugin/README.md for full details.

---

Plugins

• Gemini/GLM Focused Mode (plugins/gemini-glm-focused-mode/) – Injects a rigorous system prompt for GLM-4.7 and Gemini models to enforce precise, grounded, and persistent coding behavior. Activates based on model name matching. Install by copying index.ts to a local folder and adding the file path to your global opencode.json plugin array.
• GPT OAuth Prompt Enhancer (plugins/gpt-oauth-prompt-enhancer/) – Injects OpenCode environment context and CLI formatting guidelines into GPT model conversations. Ensures GPT models follow CLI-specific Markdown rendering conventions and understand the AGENTS.md discovery system. Activates for any model ID containing "gpt". Install by copying src/ folder locally and adding index.ts path to your global opencode.json plugin array.

---

Security Reviewer

The Security Reviewer (security-reviewer/) is a specialized agent for auditing codebases against vibecoding vulnerabilities. It bundles 10 framework-specific security skills:

Skill	Focus
security-ai-keys	AI API key leakage detection
security-bun	Bun runtime security patterns
security-convex	Convex auth/validators review
security-django	Django security settings audit
security-docker	Docker/container hardening
security-express	Express.js middleware security
security-fastapi	FastAPI auth/CORS validation
security-nextjs	Next.js Server Actions review
security-secrets	25+ secret type detection
security-vite	Vite env var exposure checks

Installation: Copy .opencode/agent/ and .opencode/skill/ to your global config or project.

See security-reviewer/README.md for full details.

---

Configuration Examples

• Thinking Levels Variants (thinking-variants config/thinking-levels-opencode.json) – OpenCode configuration with model variants for Ctrl+T thinking level switching. Supports OpenAI OAuth plugins (GPT-5.2, GPT-5.2 Codex), Mirrorwell's Antigravity Proxy (Gemini 3 Pro/Flash, Claude 4.5), and GLM coding plan (GLM-4.7). Enables dynamic reasoning effort and thinking budget adjustment directly from the UI.

---

Included Packs

Commands2Skills Template

Note

Partially Deprecated: OpenCode now supports Skills natively. This template remains useful if you want agents to discover and execute /commands agentically via a COMMANDS.md index file injected at session start. For new projects, consider using Skills instead.

A universal command template with:

• Command Integration Pattern: 4-part architecture for command-to-tool interfaces (still works, but deprecated - OC now supports skills natively)
• Template Structure: Organized framework for command-based projects
• Example Commands: Repository exploration and mathematical calculation tools

See commands2skills/README.md for implementation details, architecture overviews, and usage guidance.

Agent Catalog

Agents are organized into two collections under agents/:

• generic/ – Reusable global agents that belong in ~/.config/opencode/agent/:
  • repo-navigator-creator – Builds lean AGENTS.md navigation guides.
  • subagent-orchestrator – Dispatches specialists and enforces scope isolation. (might need some tweaks - dumber models hallucinate agents)
  • openspec-orchestrator – Enforces strict OpenSpec formatting/validation and orchestrates subagents for executing OpenSpec proposals (requires OpenSpec installed and openspec init).
• vite-react-ts-convex-tailwind/ – Stack-specific experts for the modern Vite + React 19.2 + TS 5.9 + Tailwind 4.1 + Convex stack (also supports Bun). Each agent references its matching documentation file (CONVEX.md, REACT19.md, TAILWIND4.md, TS59.MD, CODING-TS.md) so it can cite authoritative answers on demand:
  • VRTCT-orchestrator – Master coordinator for the entire stack, orchestrating subagents (does not write code).
  • VRTCT-brain – Stack knowledge base and implementation lead (writes code).
  • convex-database-expert – Schema/query/mutation/action specialist grounded in Convex docs.
  • react-19-master – React 19.2 implementation reviewer focused on Server Components, Actions, and compiler rules.
  • tailwind-41-architect – Tailwind CSS 4.1 designer that enforces the CSS-first workflow.
  • typescript-59-engineer – Strict TS 5.9 engineer who guards erasable syntax and configuration hygiene.

See agents/README.md for full tables, usage details, and the complete directory tree.

Commands Catalog

The commands/ directory provides shareable command files for Opencode users. Currently available:

• /howto (commands/.opencode/command/howto.md): An /init-style command that scans the cloned repository, searches for official documentation, and generates an AGENTS.md focused on helping end users set up, operate, and troubleshoot the software (not for development work). Run /howto right after cloning a repo so your assistant knows how to install, run, and support that project.
• /improve:run (commands/.opencode/command/improve:run.md): Transforms any task into a production-ready prompt using official prompt engineering guides from OpenAI GPT-5.1/5.1-Codex, Anthropic Claude 4.5, and Google Gemini 3 Pro, then executes it immediately.
• /improve:save (commands/.opencode/command/improve:save.md): Same enhancement as /improve:run but saves the optimized prompt as a markdown file for review and refinement before execution. Runs as a subagent to save context. When finished, @ the prompt file in the main session.
• /refactor (commands/.opencode/command/refactor.md): Refactors code with strict modularity, file headers, and cleanup. Breaks large files into focused modules, removes slop (emojis, chatty comments, console logs), enforces DRY principles, and adds concise 2-3 sentence file headers. Targets files from current session or worst offenders (250+ lines) if unspecified.
• /init (commands/.opencode/command/init.md): Creates or enhances AGENTS.md documentation while preserving human-crafted content. Analyzes repository structure and generates navigation guides for LLMs.
• /refactor-rfc-xml (commands/.opencode/command/refactor-rfc-xml.md): Converts markdown files to RFC 2119 + XML tag structure for precise, machine-readable documentation.

Scripts Catalog

The scripts/ directory contains utilities that commands or agents can reuse.

• perplexica-cli.js – Node-based CLI wrapper for Perplexica’s search API that handles long-running requests (300s timeout), prints answers with sources, and exposes focus modes (web, academic, Reddit, YouTube, Wolfram) via --mode flags. Point it at your Perplexica instance (default http://localhost:3000/api/search).
• /perplexica-search (scripts/.opencode/command/perplexica-search.md) – A specialized command that orchestrates the CLI to perform research with smart mode selection, structured summaries, and full citations.

@At Reference Files

Use everything inside at/ as prefix instructions during development sessions:

• @coding-ts (at/CODING-TS.MD): Universal engineering guidelines emphasizing DRY principles, type safety, and clean architecture. Reference it in prompts (e.g., “Implement feature X following @coding-ts principles”) before starting any new feature or refactor so the LLM stays aligned. These reference files are meant for direct @ mentions in user instructions and are not invoked by subagents.

MCP Configurations

• Authenticated Chrome DevTools MCP (mcp-configs/authenticated-chrome-dev-tools-mcp/) – Enables Chrome DevTools MCP to work with authenticated browser sessions. Normally Chromium blocks Google account login in automated browsers, and concurrent sessions can interfere with existing Chrome processes, potentially closing background windows. This wrapper creates a temporary "shadow profile" that clones your cookies, sessions, and login data, allowing AI agents to access your logged-in context without disrupting your main browser.

⚠️ Security Warning: This enables agentic AI to access all your browser's data including logged-in accounts, cookies, and session information. Use with extreme caution.

Other Opencode Projects

• Agent Swarm Demo (repo) – Shows how to launch a multi-server swarm where Claude Code handles background process management while OpenCode runs specialized agents. The swarm orchestrator coordinates inter-agent communication across multiple OpenCode instances and currently requires Claude Code because it relies on Claude’s background bash processes to spawn the servers.

Getting Started

1. Clone this repository locally.
2. Explore commands2skills/ for command templates and tooling patterns.
3. Install desired agents globally (~/.config/opencode/agent/) or copy them into your project’s .opencode/agent/ directory.
4. Install OpenSpec and run openspec init if you plan to use the OpenSpec orchestrator or compatible workflows.

Enabling Skills

Important

All agents in this repository have skills disabled by default. This prevents context pollution and keeps agent behavior predictable. You MUST explicitly enable the skills you want each agent to use.

How to Enable Skills

Add a permission: skill: section to the agent's YAML frontmatter:

---
agent: your-agent-name
description: Your agent description
mode: primary
permission:
  skill:
    "skill-name-1": "allow"
    "skill-name-2": "allow"
    "*": "deny"
---

• List each skill you want to enable with "allow"
• Always end with "*": "deny" to block all other skills
• Skills MUST be installed in ~/.config/opencode/skill/ or .opencode/skill/ for the agent to use them

Refer to each agent's documentation for the specific skills it supports.

Additional Documentation

• commands2skills/README.md – command architecture, tooling integration, and usage instructions.
• agents/README.md – agent descriptions, guardrails, and setup guidance.

About

This repository provides tested Opencode command patterns, global agent prompts, and workflow templates that demonstrate best practices for creating maintainable, scalable command-based projects with external tool integration.

---

RFC 2119 + XML Tag Structure

All agent prompts, skills, and commands in this repository use RFC 2119 keywords and XML tags for structure. This isn't arbitrary formatting—it's research-backed prompt engineering.

Why XML Tags?

XML tags provide clear boundaries that LLMs parse reliably:

• Clarity: Distinct separation between instructions, examples, and context reduces misinterpretation
• Hierarchy: Nested tags create logical groupings (e.g., <workflow> containing numbered steps)
• Consistency: Standardized tags like <instructions>, <rules>, <examples> work across all models
• Parseability: Structured output can be extracted programmatically when needed

<instructions>
1. Read the file
2. Apply changes
</instructions>

<rules>
- MUST preserve existing formatting
- SHOULD use surgical edits
</rules>

Why RFC 2119 Keywords?

RFC 2119 (BCP 14) defines precise requirement levels used in internet standards since 1997:

Keyword	Meaning
MUST	Absolute requirement. No exceptions.
MUST NOT	Absolute prohibition.
SHOULD	Strong recommendation, but valid reasons may exist to deviate.
SHOULD NOT	Discouraged, but acceptable in specific circumstances.
MAY	Truly optional.

This eliminates ambiguity. "You should validate input" is vague. "The agent SHOULD validate input before processing" is precise—it's strongly recommended but not a hard failure if skipped with reason.

Style Guide

See RFC-XML-STYLE-GUIDE.md for the complete reference, including tag catalogs, keyword definitions, and before/after examples. Use /refactor-rfc-xml to convert existing files.