OpenCode Cheatsheet for Software Engineers

OpenCode is a terminal-native AI coding agent that gives you powerful composability through agent modes, custom commands, and shell integration. This cheatsheet collects the most impactful tricks you need – from token-saving workflows to multi-agent setups.

---

Core Modes & Model Selection

Trick	What it does
Build (default) / Plan modes	Press Tab to toggle between Plan (research & architect) and Build (implement).
Cheap model for Plan, expensive for Build	Set a fast, low-cost model (e.g., GPT-4o-mini) for planning and a powerful one (e.g., Claude Opus) for building. Saves tokens without losing quality.
/model command	Switch the active model mid‑session without leaving the terminal.

---

File & Context Tricks to Save Tokens

• @filename references
  Type @ followed by a file path to attach files directly. This avoids the agent searching the whole codebase, saving context tokens and getting more accurate results.

• /compact
  Summarises the conversation so far, freeing up context window space while keeping the agent fully informed. Use it before long chains of changes.

• /diff
  Shows a clean diff of the changes the agent wants to make before applying them. Review fast, then accept or deny.

• Comments in prompts with //
  Add inline comments to your message (e.g., // don't touch the database layer). The agent respects them as meta-instructions.

---

Shell Mode & Workspace

• !command to enter shell mode
  Type !npm run test and it runs in the project’s context. Start with ! to stay in the shell. Exit shell mode with exit or Ctrl+D.

• Multiple terminal instances
  Open other terminal windows pointing to the same or a different project. Share context across workspaces or isolate risky experiments.

• @terminal
  Pipe the output of the last terminal command directly into the agent’s prompt – use it to feed logs, errors, or test results.

---

Multi-Agent Superpowers

OpenCode ships with built‑in agents and lets you create your own, all runnable in parallel.

Agent	Usage
@general	Handles broad coding tasks (default).
@explore	Searches the codebase, reads docs, gathers context.
@reviewer	Reviews code, suggests improvements, catches bugs.
Custom agents	Drop Markdown files into .opencode/agents/.

Custom Agent Example

.opencode/agents/security.md contains the agent’s personality and tools. Then invoke it with @security.
You can run @explore to research a library while @security audits the auth flow – true parallel work.

---

Custom Commands & Project Setup

• /init — Bootstraps a fine‑tuned AGENTS.md file that describes your project’s stack, conventions, and goals. Start every project with this.
• Custom slash commands — Create .opencode/commands/super-commit.md. Inside, define a prompt like:
  ”Review all changes, write a concise git commit message following conventional commits, and execute git commit -m '…'.”
  Then just type /super-commit.
• .opencode/config.json — Store default models, temperature, and other settings. Example:

  {
    "models": {
      "plan": "gpt-4o-mini",
      "build": "claude-sonnet-4-20250514"
    }
  }
  

---

Skills & Design

• skills.sh – Download a curated skill pack for a specific technology (e.g., React, FastAPI). Run once, and the agent gains deep, current knowledge of that stack.
• autoskills (github.com/midudev) – Analyses your project and suggests the best skill packs to install. Perfect when you’re starting on an unfamiliar codebase.
• DESIGN.md from Google – Drop a copy of Google’s open‑source DESIGN.md template into your repo. Use it with /compact to enforce thoughtful architectural decisions during long sessions.

---

Bonus: Show Off Your Project

Use Heygen HyperFrames to generate a dynamic video demo of your project directly from the command line. Great for READMEs, hackathons, and stakeholder updates.

---

Final flow that saves you hours:

1. Tab → Plan, ask the cheap model to design the feature.
2. /init if no AGENTS.md exists yet; then skills.sh or autoskills.
3. Run @explore to map the relevant parts of the code.
4. Switch to Build mode with an expensive model, @file the key modules, ! test, @reviewer check, /compact when context fills up.
5. Use /super-commit and keep shipping.

Happy building!