Oc Create Agent

Create OpenCode Agent

You are initiating the OpenCode agent creation workflow. This process guides the user through an interactive interview to gather requirements, generates a properly formatted agent markdown file with permissions and system prompt, and validates the result.

Critical Rules

AskUserQuestion is MANDATORY

IMPORTANT: You MUST use the AskUserQuestion tool for ALL questions to the user. Never ask questions through regular text output.

---

Phase 1: Load References

Read the OpenCode platform overview and agent guide:

1. Read ${CLAUDE_PLUGIN_ROOT}/references/platform-overview.md
2. Read ${CLAUDE_PLUGIN_ROOT}/references/agent-guide.md

---

Phase 2: Interview

Gather requirements through a structured interview.

Round 1: Identity

Question 1 — Agent Name:

• Header: "Agent Name"
• Question: "What should this agent be named? Use kebab-case (e.g., code-reviewer, test-runner)."
• Options:
• "I'll type the name" — Let me provide a custom name

Question 2 — Purpose:

• Header: "Purpose"
• Question: "What does this agent do? Describe its main function."
• Options:
• "Code analysis" — Reviews or analyzes code
• "Code generation" — Writes new code or modifies existing code
• "Research" — Gathers information from docs, web, or codebase
• "Automation" — Executes multi-step workflows

Question 3 — Mode:

• Header: "Mode"
• Question: "How will this agent be used?"
• Options:
• "Primary (Recommended)" — Interactive agent the user switches to via @name
• "Subagent" — Spawned by other skills/agents via the task tool
• "Both" — Works as primary and subagent
• multiSelect: false

Round 2: Model & Behavior

Question 1 — Model Tier:

• Header: "Model"
• Question: "What model tier should this agent use?"
• Options:
• "Sonnet (Recommended)" — Good balance of capability and speed
• "Opus" — Maximum reasoning capability for complex tasks
• "Haiku" — Fast and lightweight for simple tasks
• "Default" — Use the session's current model
• multiSelect: false

Question 2 — Temperature:

• Header: "Temperature"
• Question: "How deterministic should responses be?"
• Options:
• "Low (0.3)" — More consistent, factual responses
• "Default (0.7)" — Balanced creativity and consistency
• "High (0.9)" — More creative and varied responses
• multiSelect: false

Question 3 — Steps Limit:

• Header: "Steps"
• Question: "Should there be a limit on agentic iterations?"
• Options:
• "Default" — No explicit limit
• "Conservative (25)" — Limit to prevent runaway loops
• "Generous (100)" — Allow extended multi-step workflows
• "Custom" — I'll specify a number
• multiSelect: false

Round 3: Permissions

Question 1 — Permission Level:

• Header: "Permissions"
• Question: "What permission level should the agent have?"
• Options:
• "Read-only" — Can read, glob, grep only. No writes, edits, or bash.
• "Standard" — Read allowed, write/edit/bash require approval
• "Full access" — All tools auto-allowed
• "Custom" — I'll configure per-tool permissions
• multiSelect: false

If "Custom" selected:

Question 2 — File Operations:

• Header: "File Ops"
• Question: "File operation permissions?"
• Options:
• "read: allow, write: ask, edit: ask" — Read freely, ask for writes
• "read: allow, write: allow, edit: allow" — Full file access
• "read: allow, write: deny, edit: deny" — Read only
• multiSelect: false

Question 3 — Execution:

• Header: "Execution"
• Question: "Bash execution permission?"
• Options:
• "deny" — No shell access
• "ask" — Require approval for each command
• "allow" — Auto-approve all commands
• multiSelect: false

Question 4 — Web Access:

• Header: "Web"
• Question: "Web access permissions?"
• Options:
• "No web access" — Deny websearch and webfetch
• "Search only" — Allow websearch, deny webfetch
• "Full web access" — Allow both websearch and webfetch
• multiSelect: false

Round 4: System Prompt

Question 1 — Prompt Focus:

• Header: "Focus Areas"
• Question: "What should the system prompt emphasize? Select all that apply."
• Options:
• "Structured output" — Agent produces formatted reports or specific formats
• "Tool usage patterns" — Agent follows specific tool usage workflows
• "Domain expertise" — Agent has specialized knowledge in a domain
• "Phase-based workflow" — Agent works through numbered phases
• multiSelect: true

Question 2 — Constraints:

• Header: "Constraints"
• Question: "Any important constraints for the agent?"
• Options:
• "No destructive actions" — Never delete files or drop data
• "Stay in scope" — Don't explore beyond the assigned area
• "Concise output" — Keep responses short and focused
• "No constraints" — Let the agent operate freely
• multiSelect: true

Round 5: Location & Display

Question 1 — Location:

• Header: "Location"
• Question: "Where should the agent be created?"
• Options:
• "Project (.opencode/agents/)" — Available only in this project
• "Global (~/.config/opencode/agents/)" — Available in all projects
• "Custom path" — I'll specify the directory
• multiSelect: false

Question 2 — Display:

• Header: "Display"
• Question: "Agent display preferences?"
• Options:
• "Default" — Standard display, visible in @-mention list
• "Custom color" — I'll pick a color
• "Hidden" — Don't show in @-mention list
• multiSelect: false

Interview Summary

Present a comprehensive summary:

## Agent Summary

- **Name**: {name}
- **Description**: {generated description}
- **Mode**: {primary/subagent/all}
- **Model**: {provider/model-id}
- **Temperature**: {value}
- **Steps**: {value or "default"}
- **Permissions**: {summary}
- **Focus areas**: {list}
- **Constraints**: {list}
- **Location**: {path}

Use AskUserQuestion to confirm:

• Header: "Confirm"
• Question: "Does this look correct? Ready to generate the agent?"
• Options:
• "Yes, generate it" — Proceed with generation
• "Make changes" — I want to adjust something

---

Phase 3: Generate

Spawn the generator agent:

Task:
  subagent_type: "agent-alchemy-opencode-tools:oc-generator"
  prompt: |
    Generate an OpenCode agent with these specifications:

    Type: agent
    Name: {name}
    Description: {description}
    Mode: {mode}
    Model: {model}
    Temperature: {temperature}
    Steps: {steps}
    Permissions: {permission object}
    Color: {color or null}
    Hidden: {true/false}
    System prompt focus: {focus areas}
    Constraints: {constraints}
    Target path: {target directory}/agents/{name}.md

    Reference guide: ${CLAUDE_PLUGIN_ROOT}/references/agent-guide.md
    Template: ${CLAUDE_PLUGIN_ROOT}/references/templates/agent-template.md

    Interview notes:
    {all gathered requirements}

---

Phase 4: Validate

Spawn the validator agent:

Task:
  subagent_type: "agent-alchemy-opencode-tools:oc-validator"
  prompt: |
    Validate the following OpenCode artifact:

    Type: agent
    Path: {path to generated file}
    Reference guide: ${CLAUDE_PLUGIN_ROOT}/references/agent-guide.md

If validation fails with errors, fix the issues and re-validate.

---

Phase 5: Present

Present the generated agent:

1. Read the generated file
2. Show the file contents with syntax highlighting
3. Explain key design decisions:

• Permission model rationale
• Why certain mode was chosen
• Model tier selection reasoning
• System prompt structure

1. Show validation results
2. Explain how to use the agent:

• Primary: @{name} in the TUI
• Subagent: task({ prompt: "...", command: "{name}" })

1. If applicable, explain how to configure the model in opencode.json

CRITICAL: Complete ALL 5 phases before finishing.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: sequenzia
• Source: sequenzia/agent-alchemy
• License: MIT
• Homepage: https://sequenzia.github.io/agent-alchemy

Install and usage instructions live in the source repository linked above.