Skip to content
Back to skills

Setup

ASecurity

Interactive onboarding for the CEO plugin. Walks through settings (model tier, verify-fix, team mode, checkpoint frequency, etc.) and writes settings.json. Run once after installing the plugin, or anytime to reconfigure.

  • 6 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added May 26, 2026
ai-agentsrustgoapifrontendbackend

Works with

  • api

Security analysis

A100/100

Scanned May 27, 2026

npx -y skills add andywxy1/ceo-plugin --skill setup --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Setup?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Setup
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/andywxy1-setup/badge)](https://www.skillsdirectory.com/skills/andywxy1-setup)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: setup
description: "Interactive onboarding for the CEO plugin. Walks through settings (model tier, verify-fix, team mode, checkpoint frequency, etc.) and writes settings.json. Run once after installing the plugin, or anytime to reconfigure."
---

# CEO Setup

You are running the CEO plugin setup wizard. Walk the user through configuring their preferences and write the result to `${CLAUDE_PLUGIN_ROOT}/settings.json`.

## Step 1: Welcome

Say:
> **Welcome to CEO Plugin Setup**
>
> I'll walk you through a few preferences that control how the CEO orchestrates your agents. This takes about 1 minute. You can re-run `/ceo:setup` anytime to change these.

Check if `${CLAUDE_PLUGIN_ROOT}/settings.json` already exists. If it does, read it and say:
> "You already have a settings file. I'll show your current choices — just confirm or change them."

## Step 2: Model Tier

Use AskUserQuestion:

```json
{
  "questions": [
    {
      "question": "How should the CEO assign models to agents? This is the biggest cost/quality tradeoff.",
      "header": "Model Tier",
      "options": [
        {
          "label": "Balanced (Recommended)",
          "description": "Opus for architects, reviewers, strategists. Sonnet for developers, writers, operators. Best tradeoff of quality and speed."
        },
        {
          "label": "Max Quality",
          "description": "ALL agents run on Opus. Best reasoning across the board, but slower and more expensive."
        },
        {
          "label": "Max Speed",
          "description": "ALL agents run on Sonnet. Fastest and cheapest. May miss nuance on complex architecture or review tasks."
        }
      ],
      "multiSelect": false
    }
  ]
}
```

Map the answer:
- "Balanced" → `"balanced"`
- "Max Quality" → `"max_quality"`
- "Max Speed" → `"max_speed"`

## Step 3: Verification & Teams

Use AskUserQuestion with two questions:

```json
{
  "questions": [
    {
      "question": "Should every implementation task be reviewed by a read-only agent before marking complete? (Catches bugs but adds time per task)",
      "header": "Verify Loop",
      "options": [
        {
          "label": "Yes, always verify (Recommended)",
          "description": "A reviewer checks every deliverable. Max 3 retry cycles before escalating to you."
        },
        {
          "label": "Yes, but max 1 retry",
          "description": "Verify once. If it fails, escalate to you immediately instead of looping."
        },
        {
          "label": "No, skip verification",
          "description": "Trust agent output without review. Faster but riskier. Not recommended for production work."
        }
      ],
      "multiSelect": false
    },
    {
      "question": "How should the CEO coordinate agents that need to work together (e.g., frontend + backend on the same API)?",
      "header": "Team Mode",
      "options": [
        {
          "label": "Auto (Recommended)",
          "description": "Use team messaging for coupled workstreams, standalone dispatch for independent tasks."
        },
        {
          "label": "Always use teams",
          "description": "Force all multi-agent workstreams into team mode with SendMessage coordination."
        },
        {
          "label": "Never use teams",
          "description": "Always use standalone Agent() calls with handoff documents. Simpler but agents can't talk to each other."
        }
      ],
      "multiSelect": false
    }
  ]
}
```

Map answers:
- Verify: "Yes, always" → `{ enabled: true, max_retries: 3, reviewer_model: "opus" }`. "Yes, max 1" → `{ enabled: true, max_retries: 1, reviewer_model: "opus" }`. "No" → `{ enabled: false, max_retries: 3, reviewer_model: "opus" }`.
- Team: "Auto" → `"auto"`, "Always" → `"always"`, "Never" → `"never"`.

If model_tier is `"max_quality"`, set `reviewer_model: "opus"`. If `"max_speed"`, set `reviewer_model: "sonnet"`. If `"balanced"`, set `reviewer_model: "opus"` (reviewers should reason deeply).

## Step 4: Autonomy Level

Use AskUserQuestion:

```json
{
  "questions": [
    {
      "question": "How often should the CEO pause to check in with you?",
      "header": "Checkpoints",
      "options": [
        {
          "label": "After each workstream (Recommended)",
          "description": "CEO reports progress and asks approval before starting the next workstream. Good balance of control and flow."
        },
        {
          "label": "After each task",
          "description": "CEO pauses after every single task. Maximum oversight but slower."
        },
        {
          "label": "Only at phase boundaries",
          "description": "CEO runs autonomously within phases, only pausing between Discovery/Planning/Pre-flight/Execution. Fastest, least control."
        }
      ],
      "multiSelect": false
    },
    {
      "question": "How many agents should review requirements during pre-flight? (More = catches more ambiguity, less = faster)",
      "header": "Pre-flight",
      "options": [
        {
          "label": "3 agents (Recommended)",
          "description": "Good coverage without slowing things down."
        },
        {
          "label": "1 agent (Lightweight)",
          "description": "Fastest pre-flight. Only the most critical agent reviews."
        },
        {
          "label": "5 agents (Thorough)",
          "description": "Maximum ambiguity detection. Best for high-stakes or large projects."
        }
      ],
      "multiSelect": false
    }
  ]
}
```

Map:
- Checkpoint: "After each workstream" → `"workstream"`, "After each task" → `"task"`, "Only at phase boundaries" → `"phase"`.
- Pre-flight: "3" → `3`, "1" → `1`, "5" → `5`.

## Step 5: Write settings.json

Assemble the settings object and write to `${CLAUDE_PLUGIN_ROOT}/settings.json`:

```json
{
  "model_tier": "<selected>",
  "verify_fix": {
    "enabled": true,
    "max_retries": 3,
    "reviewer_model": "opus"
  },
  "team_mode": "<selected>",
  "checkpoint": "<selected>",
  "preflight_agents": 3,
  "project_dir": "./ceo-projects",
  "shared_context": true
}
```

## Step 6: Confirm

Show the user their final settings in a readable summary table:

```
Setting              | Value
---------------------|------------------
Model tier           | Balanced (Opus + Sonnet)
Verify-fix loop      | Enabled, max 3 retries
Reviewer model       | Opus
Team mode            | Auto
Checkpoints          | After each workstream
Pre-flight agents    | 3
Project directory    | ./ceo-projects
Shared context       | Enabled
```

Then say:
> **Setup complete.** Your preferences are saved to `settings.json`. Run `/ceo` to start a project, or `/ceo:setup` anytime to reconfigure.

## Notes

- `shared_context` and `project_dir` use sensible defaults and are not asked during setup to keep it fast. Users can edit `settings.json` manually to change them.
- The schema file at `${CLAUDE_PLUGIN_ROOT}/settings.schema.json` documents every field with descriptions and valid values.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.