Prism Starter

Starter files for Chapters 10-13 of From Prompts to Pipelines. A multi-client content management system built on OpenClaw.

The system routes Telegram messages to client-specific agents, generates voice-matched content, enforces compliance rules, and remembers client preferences across sessions. Uses OpenClaw's Paperclip multi-agent architecture with a router agent (cheap model) dispatching to per-client agents (quality model).

Quick Start

# 1. Install OpenClaw if you haven't
curl -fsSL https://openclaw.ai/install.sh | bash

# 2. Clone this into your prism project
git clone https://github.com/ciscoittech/prism-starter.git prism
cd prism

# 3. Link the config so OpenClaw can find it
mkdir -p ~/.openclaw
ln -sf $(pwd)/openclaw.json ~/.openclaw/openclaw.json

# 4. Set your environment variables
export OPENROUTER_API_KEY=your-key-here
export TELEGRAM_BOT_TOKEN=your-bot-token
export TELEGRAM_OWNER_ID=your-telegram-id
export GATEWAY_TOKEN=any-random-string-here

# 5. Install the plugins
openclaw plugins install ./plugins/voice-scorer
openclaw plugins install ./plugins/content-calendar

# 6. Update the router skill with your Telegram chat IDs
# Edit skills/router/SKILL.md — replace [MAYA_CHAT_ID] and [JORDAN_CHAT_ID]

# 7. Start the gateway
openclaw config set gateway.mode local
openclaw gateway

Plugins are pre-built so you don't need TypeScript or a build step. Just install via CLI and run.

What's Inside

prism-starter/
├── openclaw.json              # Multi-agent config (router + 2 client agents)
├── plugins/
│   ├── voice-scorer/          # Scores content against voice profiles
│   │   ├── src/               # Source code (TypeScript)
│   │   └── dist/              # Pre-built (ready to run)
│   └── content-calendar/      # Per-client content scheduling
│       ├── src/               # Source code (TypeScript)
│       └── dist/              # Pre-built (ready to run)
├── skills/
│   ├── router/                # Routes messages to correct client agent
│   │   └── SKILL.md
│   └── shared/
│       ├── content-pipeline/  # 5-stage content production workflow
│       │   └── SKILL.md
│       └── memory-rules/      # Provenance labels for stored knowledge
│           └── SKILL.md
└── clients/
    ├── maya/
    │   ├── voice/             # Maya's writing style and voice parameters
    │   │   └── SKILL.md
    │   └── compliance/        # Maya's brand rules (health claims, sponsorships)
    │       └── SKILL.md
    └── jordan/
        ├── voice/             # Jordan's writing style and voice parameters
        │   └── SKILL.md
        └── compliance/        # Jordan's brand rules (copyright, community safety)
            └── SKILL.md

Adding a New Client

1. Create voice and compliance skills:

   clients/newclient/voice/SKILL.md      # Voice profile + scoring parameters
   clients/newclient/compliance/SKILL.md  # Brand rules (BLOCKING)
   

2. Add the agent to openclaw.json:

   {
     "id": "newclient",
     "identity": {
       "name": "NewClient's Agent",
       "theme": "content creation for NewClient"
     },
     "model": { "primary": "openrouter/anthropic/claude-sonnet-4-6" },
     "skills": ["newclient-voice", "newclient-compliance", "content-pipeline", "memory-rules"]
   }

3. Add the skill directories to openclaw.json skills.load.extraDirs

4. Add the client's Telegram chat ID to skills/router/SKILL.md

The Framework

This system uses four concepts from the book:

Concept	Where You See It
Instruction	Voice skills tell each agent HOW to write for their client
Memory	Provenance-labeled knowledge persists across sessions
Control	Voice scoring + compliance rules block bad content before it ships
Flow	Five-stage pipeline: plan, generate, check, approve, publish

Read the free chapter: Why That Worked (And Why Most AI Doesn't)

Get the full book: From Prompts to Pipelines

For Developers

If you want to modify the plugins:

# Install dev dependencies (includes TypeScript + OpenClaw SDK)
cd plugins/voice-scorer && npm install && npm run build
cd ../content-calendar && npm install && npm run build

Note: building requires ~2GB RAM due to the OpenClaw SDK. Pre-built dist/ files are included for environments with limited resources.

License

MIT