Devrel Autonomy

DevRel Claude Code Plugin

A Claude Code plugin system for autonomous DevRel demo creation. Enables Claude Code to develop demos, documentation, and content with minimal human supervision.

Installation

Quick Install (from GitHub)

# Add this repository as a plugin marketplace
/plugin marketplace add djliden/devrel-claude-code-plugin

# Install the plugin
/plugin install devrel-autonomy@devrel-marketplace

Manual Install (clone first)

# Clone the repo
git clone https://github.com/djliden/devrel-claude-code-plugin.git

# Add the local marketplace
/plugin marketplace add ./devrel-claude-code-plugin

# Install the plugin
/plugin install devrel-autonomy@devrel-marketplace

Available Commands

Once installed, you'll have access to these slash commands:

Command	Description
/devrel-start	Start a new DevRel demo project with autonomous execution
/devrel-expand	Expand existing content to new formats (code → blog, etc.)
/devrel-review	Review session status, decisions, and questions

How It Works

/devrel-start - New Project Kickoff

1. Gather Requirements: Collects project concept, target artifacts, writing samples for style matching, and resource constraints
2. Clarify Build vs Use: If you mention an existing project/tool, confirms whether to USE it or BUILD something new (prevents wasted effort)
3. Permissions Verification: Shows exactly what will be accessed (API keys, services, local ops) and gets explicit approval
4. Browser Setup (optional): If demo has a web UI, offers Playwright integration for testing and screenshots
5. Execute Autonomously: Works until complete, documenting decisions and logging progress
6. Deliver Review Package: Presents working code + content drafts + decision log + review checklist

/devrel-expand - Artifact Expansion

Transform existing content into new formats:

• Code demo → Blog post
• Blog post → Video script
• Any combination → Runbook/guide

/devrel-review - Session Review

Check progress on autonomous sessions:

• Summary of accomplishments
• Decision log with rationale
• Accumulated questions
• Items needing human input

Design Principles

Autonomy

• Works until finished or hitting an impossible blocker
• Makes decisions and documents them (doesn't halt for permission)
• Only stops for: hard external blockers, irreversible high-stakes decisions, or completion

DevRel Suitability

Demo code, not production code:

• Simple and readable over robust and defensive
• Clear inputs, visible outputs, easy to follow
• Educational - code that teaches

Human Touch

• Matches provided writing style samples
• Generates options, not just one path
• Human is editor-in-chief, not proofreader

Permissions Setup

The plugin automatically creates .claude/settings.local.json during setup with permissions for autonomous work. This includes:

• Sandbox mode: Safe execution boundaries
• Auto-approve bash: Common commands (uv, python, git, npm, etc.) run without prompts
• Auto-approve edits: File changes don't require confirmation

If you want to set this up manually, create .claude/settings.local.json:

{
  "permissions": {
    "allow": [
      "Bash(uv:*)", "Bash(python:*)", "Bash(pip:*)",
      "Bash(uvicorn:*)", "Bash(mlflow:*)",
      "Bash(npm:*)", "Bash(npx:*)", "Bash(node:*)",
      "Bash(curl:*)", "Bash(git:*)",
      "WebFetch", "WebSearch"
    ],
    "defaultMode": "acceptEdits"
  },
  "sandbox": {
    "enabled": true,
    "autoAllowBashIfSandboxed": true
  }
}

Key settings:

• defaultMode: "acceptEdits" - Auto-approves file edits
• sandbox.autoAllowBashIfSandboxed: true - Auto-approves bash commands within sandbox safety boundaries

Specialized Agents

The plugin uses specialized subagents for different tasks:

Agent	Purpose
orchestrator	Coordinates workflow, manages session file, delegates to other agents
coder	Writes demo code - simple, readable, educational (not production code)
writer	Creates content (blogs, scripts) matching your writing style
reviewer	QA check before human review - verifies code runs, content matches style
browser	Playwright-based UI testing, screenshots, web interaction

Screenshot Beautification

The plugin includes a screenshot-beautifier skill that uses ImageMagick to polish raw screenshots with rounded corners, shadows, and gradient backgrounds.

Prerequisite: Install ImageMagick

brew install imagemagick  # macOS
apt-get install imagemagick  # Linux

Usage via helper script:

# Default (macos style)
./scripts/beautify.sh screenshot.png

# Or specify preset: macos, gradient, minimal, dark, white
./scripts/beautify.sh screenshot.png gradient output.png

Presets:

Preset	Description
macos	(default) macOS window style - rounded corners, soft shadow, light background
gradient	Purple/blue gradient background, rounded corners, shadow
minimal	Light gray background, subtle shadow
dark	Dark background with purple glow
white	Clean white background with shadow

The skill also includes detailed ImageMagick commands if you need custom styling. Claude will automatically use this skill when asked to beautify or polish screenshots.

Browser Automation (Optional)

For demos with web UIs, install the Playwright plugin:

/plugin install playwright@claude-plugins-official

This enables:

• Automated UI testing
• Screenshot capture for documentation
• Form filling and interaction verification
• Works with authenticated sites (you log in manually, Claude continues)

Repository Structure

devrel-claude-code-plugin/
├── .claude-plugin/
│   └── marketplace.json          # Marketplace manifest (must be at repo root)
└── devrel-autonomy/              # The plugin
    ├── .claude-plugin/
    │   └── plugin.json
    ├── agents/
    │   ├── browser.md            # UI testing and screenshots
    │   ├── coder.md              # Demo code development
    │   ├── orchestrator.md       # Workflow coordination
    │   ├── reviewer.md           # Quality assurance
    │   └── writer.md             # Content creation
    ├── commands/
    │   ├── devrel-start.md
    │   ├── devrel-expand.md
    │   └── devrel-review.md
    └── skills/
        └── screenshot-beautifier/
            ├── SKILL.md          # ImageMagick instructions
            └── scripts/
                └── beautify.sh   # Helper script with presets

Example Usage

> /devrel-start

# System asks for:
# - Project concept (what does the demo show?)
# - Target artifacts (code, blog, video script?)
# - Writing sample (for style matching)
# - Resource constraints (which APIs/models to use?)

# Then works autonomously, delivering:
# - Working code + README
# - Content drafts
# - Decision log
# - Review checklist

Tips

Monitor the First Few Minutes

After launching autonomous execution with "go", stick around for 3-5 minutes. Early permission prompts or configuration issues are most likely to appear at the start (cloning repos, installing dependencies, starting servers). Once the agent is past initial setup and working smoothly, it's safer to step away.

Restart After Permissions Setup

Claude Code loads settings at startup. After the setup phase writes .claude/settings.local.json, you'll need to restart and /resume for the new permissions to take effect. The plugin will prompt you to do this.

Check /devrel-review Periodically

Even during autonomous execution, you can run /devrel-review in another terminal (in the same project directory) to check progress, see decisions made, and catch any accumulated questions.

Advanced: Auto-Deny Hook for Fully Hands-Off Operation

If you want Claude to automatically continue when hitting unexpected permission prompts (rather than waiting for you), you can set up an auto-deny hook. Use with caution - this means Claude will skip operations it can't perform rather than waiting for approval.

Create a script scripts/deny-and-continue.sh:

#!/bin/bash
cat <<EOF
{
  "hookSpecificOutput": {
    "hookEventName": "PermissionRequest",
    "decision": {
      "behavior": "deny",
      "message": "User unavailable. Find another approach or document this blocker and continue with other work.",
      "interrupt": false
    }
  }
}
EOF

Add to .claude/settings.local.json:

{
  "hooks": {
    "PermissionRequest": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "./scripts/deny-and-continue.sh"
          }
        ]
      }
    ]
  }
}

To disable: Remove the hooks section from settings.local.json.

Recommendation: Only use this if you're comfortable debugging Claude Code issues. For most users, the default behavior (pause and wait for permission) is safer.

License

MIT