opencode-prompting

Prompting through opencode

opencode's POST /session/:id/message accepts a prompt and returns the assistant's response. The user picks the underlying model (Anthropic, OpenAI, Moonshot, OpenRouter routes, etc.) via opencode's own config or the --model <provider/model> flag. This plugin doesn't choose models or alias them.

When you're forwarding a user's task

The companion's rescue subcommand sends the user's free-form text directly. Don't add framing the user didn't ask for. Specifically:

• Don't prepend "You are an expert software engineer..." style preambles.
• Don't add output-format prescriptions ("respond in markdown with sections...").
• Don't attach extra context (file dumps, diff summaries) the user didn't include.

If the user wants context, they include it. opencode has its own tools for reading files and running shell commands inside the session — let it use them on demand.

When you're building review prompts

buildReviewPrompt and buildAdversarialReviewPrompt already produce the right framing for their respective modes. Don't wrap or rewrite them upstream.

Model-agnostic tone

Avoid model-family quirks:

• No "let me think step by step" prefixes (extended-thinking-style).
• No "respond with reasoning_effort=high" hints.
• No JSON-mode tricks unless the user explicitly asks for structured output.

Different providers handle these differently or ignore them. A clean, declarative prompt travels well across opencode's model registry.

Long context

opencode's session model already keeps conversation history. For follow-ups, prefer /opencode:rescue --resume <text> over re-uploading context — the session retains what was already discussed.