opencode-prompting

OpenCode Prompting

Workers can't read files — all context must be in the prompt. A good prompt = a good result.

---

Prompt structure

## Task
[One clear sentence: what to do]

## Context
[Paste the relevant code, diff, error message, or file content directly here]

## Output Format
[Exact shape: "List as bullets with severity", "Max 500 words", "Steps in order"]

---

Rules

1. One task per worker. Don't combine review + planning + debugging in one prompt.
2. Paste code inline. Workers cannot cat files — include the relevant excerpt.
3. Specify output format. "3 bullet points max" → worker stays focused.
4. Set length limits. "Under 400 words" → prevents rambling.
5. Avoid vague questions. "What are the 3 biggest issues?" beats "What do you think?"

---

Templates by task type

Code review

## Task
Review this code for bugs, security issues, and code quality problems.

## Code
<paste diff or file content here>

## Output Format
Bullet list: [SEVERITY] file:line — description
SEVERITY: CRITICAL | HIGH | MEDIUM | LOW
Max 10 findings, CRITICAL first.

Implementation plan

## Task
Create an implementation plan for: <description>

## Current state
<paste relevant file structure or existing code>

## Output Format
1. Files to create/modify (with paths)
2. Key decisions and tradeoffs
3. Step-by-step order
4. Risks
Max 600 words.

Debugging / Q&A

## Task
<specific question>

## Context
<error message, stack trace, or relevant code>

## Output Format
1. Root cause (one sentence)
2. Fix (code snippet if applicable)
3. Why this works (one sentence)

---

What to delegate (good fit for OpenCode)

• Code review with structured output
• Implementation planning
• Explaining code patterns
• Finding bugs and anti-patterns in a specific file/diff
• Generating boilerplate

What NOT to delegate (keep in Claude)

• Multi-file refactoring that requires real-time edits
• Tasks under 50 tokens to answer
• Anything requiring iterative back-and-forth with the user
• Tasks where you need live file access during the answer

---

Model selection

The worker uses the model configured via /swarm-code:init. Pass model="haiku" to the Agent tool — Haiku is cheap and fast, sufficient for most delegation tasks.

For complex tasks (security audit, architecture review), pass the prompt directly to Claude instead of delegating — the token savings don't justify the quality tradeoff for high-stakes decisions.

---

Composing the prompt

Claude should NOT forward the raw user message to the worker. Always:

1. Add the relevant code/context inline
2. Specify the output format clearly
3. Set length limits

Then pass the enriched prompt to the worker.