---

name: "@tank/opencode-agent-creator" description: | Create specialized OpenCode agents that assume specific roles — frontend architect, DevOps SRE, database specialist, code reviewer, or any domain expert. Covers agent anatomy (markdown and JSON formats), role design (persona, expertise, behavioral directives), system prompt engineering, tool permissions, oh-my-opencode integration (categories, skill injection, delegation wiring), and converting existing Tank skills into standalone agents. Synthesizes OpenCode source (sst/opencode), oh-my-opencode v3.x agent system, and production agent analysis.

Trigger phrases: "create agent", "opencode agent", "new agent", "convert skill to agent", "make an agent", "agent from skill", "custom agent", "specialized agent", "agent template", "agent for", "agent configuration", "agent role", "create a specialist", "build agent", ".opencode/agent"

OpenCode Agent Creator

Core Philosophy

1. Role-first, not prompt-first — Define WHO the agent is before WHAT it says. Identity drives behavior.
2. Agents have opinions — Effective agents push back, refuse out-of-scope work, and have strong preferences. Bland agents waste tokens.
3. Minimum viable agent — Start with 20 lines. Add complexity only when the agent fails at real tasks. Over-engineered prompts confuse models.
4. Skills supplement, agents act — Skills inject knowledge. Agents assume roles. Use both together for maximum effectiveness.

Quick-Start: Create an Agent

"I need a specialist agent for [DOMAIN]"

1. Determine the role archetype: → See references/role-design.md for archetypes table

2. Choose agent properties:

   • Mode: all for versatile, subagent for delegation-only
   • Model: Match complexity to role (see model selection guide)
   • Temperature: 0.0-0.3 for precision, 0.3-0.5 for creative
   • Permissions: Restrict tools based on role

3. Write the agent file: → Use assets/agent-template.md as starter → See references/prompt-engineering.md for prompt structure

4. Deploy:

   • Global: ~/.config/opencode/agent/<name>.md
   • Project: .opencode/agent/<name>.md

5. Test with 5 patterns:

   • In-domain task → competent response
   • Out-of-scope request → polite refusal
   • Ambiguous request → asks clarifying question
   • Anti-pattern proposal → pushes back
   • Complex multi-step task → structured approach

"I want to convert a skill into an agent"

1. Read the skill's SKILL.md and references
2. Extract: domain scope, decision frameworks, anti-patterns, workflows
3. Transform passive knowledge into active role directives
4. Add personality, boundaries, and tool restrictions → See references/skill-to-agent-conversion.md

"Agent exists but isn't working well"

Problem	Likely Cause	Fix
Too generic/bland	No personality defined	Add opinions and preferences
Does things it shouldn't	No boundary section	Add "Outside Your Scope"
Asks too many questions	No decision framework	Add conditional directives
Ignores its restrictions	Prompt contradicts permissions	Align prompt with frontmatter
Doesn't appear in @ menu	Wrong file location	Move to ~/.config/opencode/agent/
Sisyphus doesn't delegate to it	Poor description	Add <example> tags to description

Agent File Format

Markdown file with YAML frontmatter. Filename = agent name.

---
description: >-
  Use this agent when [TRIGGER CONDITION].
  Includes [TASK 1], [TASK 2], [TASK 3].
  
  <example>
  user: "[request]"
  assistant: "I'll use [agent] to [action]."
  </example>
mode: all
model: provider/model-id
temperature: 0.1
color: "#HEX"
permission:
  edit: allow|ask|deny
  bash:
    "*": ask
    "specific command": allow
---

[System prompt — the agent's role, expertise, and behavioral rules]

→ See references/agent-anatomy.md for complete property reference.

Decision Trees

Archetype Selection

Need	Archetype	Key Config
Writes code in a domain	Specialist	edit: allow, low temp
Reviews but doesn't modify	Reviewer	edit: deny, read-only bash
Gathers information	Researcher	edit: deny, bash: deny, webfetch: allow
Creates plans, not code	Planner	edit: ask, read-only bash
Coordinates other agents	Orchestrator	Full permissions, task tool

Model Tier Selection

Role Complexity	Model Tier	Examples
Deep reasoning / architecture	Opus / GPT-5.2+	Solution architect, security auditor
Code generation / editing	Sonnet / GPT-5.2	Frontend engineer, API developer
Fast search / simple tasks	Haiku / Flash	Explorer, formatter, triage

Where to Deploy

Scope	Path	When
All projects	~/.config/opencode/agent/	Personal workflow agents
One project	.opencode/agent/	Project-specific specialists
omo override	oh-my-opencode.json → agents	Modify built-in omo agents

Anti-Patterns

Pattern	Problem	Fix
"Be helpful and thorough"	Vacuous instructions	Specific behavioral directives
Agent does everything	No focus, mediocre at all tasks	Narrow to 1-2 domains
200+ line prompt	Model loses focus	Under 150 lines, use skills for knowledge
No NEVER rules	Agent makes domain-specific mistakes	Add 3-5 hard constraints
Copying skill text verbatim	Passive knowledge, no agency	Transform into behavioral directives
No description examples	Primary agents can't route to it	Add 3-5 <example> tags

Reference Files

File	Contents
references/agent-anatomy.md	Full OpenCode agent schema, both formats, all properties, permission system, file locations
references/role-design.md	Role archetypes, persona construction, behavioral directives, decision frameworks, common mistakes
references/prompt-engineering.md	System prompt structure, directive patterns, tool restrictions, testing strategies, anti-patterns
references/skill-to-agent-conversion.md	When to convert, transformation process, patterns (specialist, reviewer, planner), verification
references/omo-integration.md	oh-my-opencode overrides, category system, skill injection, multi-model orchestration, deployment