Hello PM Enthusiasts and Professionals,
Here are the recordings and resources from masterclass on “ClawdBot aka OpenClaw for Product Managers“
OpenClaw Masterclass — Complete Summary
What Is OpenClaw and Why Should You Care?
OpenClaw is an open-source AI agent framework that lets you talk to any large language model (GPT, Claude, Gemini, DeepSeek, Llama, Minimax) through your everyday messaging apps — WhatsApp, Telegram, Slack — and then get that LLM to act on your behalf. Not just answer questions, but send emails, create files, scrape websites, set reminders, connect to APIs, and execute recurring tasks autonomously.
The critical insight is this: before OpenClaw, if you wanted an LLM to do something in the real world (send an email, update a Jira ticket, pull data from a website), a developer had to write the integration code. OpenClaw flips that. It gives the LLM itself the ability to write, save, and execute code. The LLM becomes the developer. You just talk to it in natural language.
This is why OpenClaw is currently the 5th most-starred repository on GitHub globally — and every repository above it (Linux, Vue, React, Next) is at least a decade old. The developer community isn’t endorsing OpenClaw as a finished product. They’re endorsing the architecture as a paradigm shift.
Think of it as the Wright Brothers’ airplane. It’s not safe, it can’t fly far, and most people wouldn’t get on it. But it’s more historically significant than any commercial jet that came after it. OpenClaw is the first airplane of AI agents.
The Architecture: Why It’s Brilliantly Simple
OpenClaw’s architecture has six components, and understanding them is more valuable than knowing how to use the tool itself, because this pattern will become the foundation for how agents are built going forward.
1. Gateway (The Front Door) This is where every message enters and every response exits. The gateway connects to channels — your WhatsApp, Telegram, Slack, a web chatbot, or any API endpoint. The innovation here is that instead of building their own proprietary chat interface, OpenClaw lets you interact from wherever you already are. You don’t switch contexts. You don’t open a new app. You message your Telegram bot while standing in line for coffee, and your AI agent gets to work.
2. LLM (The Brain) The gateway passes your message to any large language model you choose. This is model-agnostic by design. You can use GPT, Claude, Gemini, DeepSeek, or even run a local open-source model like Llama on your own hardware. You swap models depending on the task — use a cheap, fast model like Gemini Flash for summarization, and a powerful one like Claude Opus for complex reasoning or code generation. You switch models directly from Telegram by tapping a menu option.
3. PI Agent (The Hands) This is the breakthrough component. The PI Agent is a small piece of software that allows any other program to create, edit, run, and delete files on a system. The LLM generates code, tells the PI Agent to save it to a specific location, and then tells it to execute that code. The LLM is now the orchestrator. It doesn’t just know how to connect to Gmail — it writes the script that connects to Gmail, deploys it, and runs it. Every interaction between any two software systems in the world happens through code. If you give an LLM the power to write and execute code, it can theoretically integrate with anything.
4. Memory (The Personality) OpenClaw maintains memory through simple text files. There’s an agents.md file that stores everything about the agent’s configuration. There’s a soul.md file where the agent’s personality builds over time — you give it a name, instructions, preferences, and it saves them. There’s a memory file where every conversation is minified and stored, so the agent has context about you across sessions. This solves one of the biggest failures in AI adoption: agents that forget everything and treat every interaction as a blank slate. OpenClaw’s agent remembers your preferences, your ongoing tasks, your communication style.
5. Cron & Heartbeat (The Pulse) This is what makes OpenClaw proactive rather than reactive. A heartbeat runs every 30 minutes, reading through all the agent’s files to determine if there’s something it should do for you. A cron system lets you schedule specific recurring tasks. Together, they mean the agent doesn’t wait for you to ask — it monitors, anticipates, and acts. If you’ve mentioned a medication schedule in conversation, it can remind you. If you’ve set up a news digest, it fires on schedule without prompting.
6. Skills (The Expertise) Skills are prompt templates stored as simple markdown files in a folder. Each skill has a name, a description, and the full prompt content. When you ask the agent to write a PRD, it reads the description of each skill, identifies the right one, and applies it automatically. No switching between ChatGPT projects or gems. No copy-pasting prompts. You just say “write a PRD for this feature” and it pulls the right skill. You can create your own skills, download verified ones from ClawHub (a skill marketplace), or even ask your Telegram bot to create and save a new skill for you — all in natural language.
How OpenClaw Differs From Everything Else
The typical AI tool — ChatGPT, Claude, Gemini — is a brilliant mind that can’t act. It gives you the most plausible-sounding answers, but it can’t send the email, can’t update the ticket, can’t scrape the website. It’s the world’s smartest person who never leaves their desk.
The typical automation tool — N8N, Make, Relay — can act, but requires explicit configuration. You have to create nodes, get API keys, test connections, and understand the plumbing. It’s powerful but not accessible to non-technical users.
OpenClaw merges both. The LLM provides the intelligence and the implementation. When Ankit told his bot he wanted it to send emails through Gmail, the bot itself told him to go to Google Cloud Console, create a project, enable the API, and share the credential file. Ankit followed the instructions, shared the JSON file through Telegram, and the bot configured itself. No code written by a human. No developer needed.
The four limitations of current agents that OpenClaw addresses:
Reactive-only behavior. Most agents wait for you to prompt them. OpenClaw’s heartbeat mechanism checks every 30 minutes for things it should proactively do.
Explicit tool configuration. Most agents require a developer to wire up every integration. OpenClaw’s LLM writes its own integration code.
No personalization or memory. Most agents treat every user identically and forget between sessions. OpenClaw maintains persistent memory files that build context over time.
Limited extensibility. Most pre-built agents (Gamma, Lovable, etc.) do one thing well but can’t be extended to arbitrary tasks. OpenClaw is completely open-ended — if it can be done through code, the agent can do it.
Installation: Three Paths, One Recommendation
Option 1: Your personal computer. Never recommended. OpenClaw takes system-level permissions and can read, write, and execute anything on your machine. If there’s a prompt injection or security breach, your personal data is exposed.
Option 2: A spare machine (Mac Mini, etc.). Possible but expensive (~$600+) and requires always-on internet. Only for people who truly know what they’re doing.
Option 3: A Virtual Private Server (VPS). The recommended approach. A VPS is a virtual computer in the cloud where your personal data doesn’t exist, it runs 24/7, and you can destroy it completely if anything goes wrong.
The simplest path: Go to Hostinger (or any hosting provider), buy a monthly VPS plan, open Docker Manager, click Projects → Compose → One-Click Deploy, search for OpenClaw, and deploy. It asks for your LLM API key (start with Gemini — free tier available at Google AI Studio) and gives you an access code.
For Telegram integration: Search for “BotFather” on Telegram, create a new bot, get the API token, and paste it into the OpenClaw dashboard under Channels → Telegram. That’s it. Your bot is live.
For those who want the command-line route: Open the terminal in your VPS, run the single install command from OpenClaw’s website, and follow the prompts. It handles the rest.
Security: You’re the Driver
OpenClaw is not secure by default. This is acknowledged openly. But security is your responsibility, just as driving safety is the driver’s responsibility, not the car’s. Here are the six rules:
Use a VPS, not your personal machine. Your personal computer has saved passwords, credit card info, personal documents. A VPS is a clean environment you can nuke if something goes wrong.
Run security audits frequently. The command openclaw security audit checks for exposed keys, misconfigured permissions, and other vulnerabilities. Run it regularly.
Only install verified skills. Before installing any skill from ClawHub, click “Hide Suspicious” to filter out potentially malicious ones. Copy any skill’s code and paste it into Claude or GPT to ask whether it’s safe before installing it. Most OpenClaw security incidents come from malicious skills.
Use separate, disposable accounts. Don’t give OpenClaw access to your primary Gmail, Slack workspace, or Jira. Create a service account with limited permissions. If you don’t know what a service account is, ask your engineering team — they’ll set one up in minutes.
Use Doppler for secrets management. Don’t hardcode API keys into configurations. Ask OpenClaw how to configure Doppler, and it’ll walk you through it.
Log everything to Supabase. If your keys get compromised or your usage spikes unexpectedly, you need visibility. Ask OpenClaw to configure logging to Supabase (free tier), and you’ll have a dashboard showing every interaction, even if you lose access to the OpenClaw instance itself.
The PM Workflow: Market Insights Ghost
Here’s a fully architected example of how a product manager can use OpenClaw for ongoing market intelligence — a “ghost” workflow that runs silently in the background.
Step 1 — Define Sources. List your competitors’ websites, review platforms (G2, Capterra), your own customer support chat logs, and call transcripts from tools like Gong, Fireflies, or Otter. These tools all have APIs. You tell OpenClaw to create a script that connects to them, it asks for your API key, and the integration is done.
Step 2 — Fetch Content. Ask OpenClaw to install a web scraping skill (it’ll use Puppeteer or FireCrawl under the hood). It crawls your listed sources on a schedule, pulling fresh content each cycle.
Step 3 — Generate Insights. Don’t just ask it to “summarize.” Use strategic frameworks: SWOT analysis, PESTEL analysis, Porter’s Five Forces, competitive comparison matrices. The more structured your instructions, the more actionable the output. Also provide an instructions file that describes your company’s context, positioning, and priorities so the agent understands what matters to you.
Step 4 — Deliver via Email. Connect your Gmail (using the same credential-sharing method demonstrated with the Hacker News bot), and set up a daily or weekly email digest. The agent scrapes, analyzes, formats, and emails — all without you lifting a finger.
You can apply this same pattern to any PM workflow: competitive analysis, user feedback synthesis, release notes aggregation, stakeholder update generation, or metric anomaly alerts. The framework is always: define sources → fetch content → apply analytical framework → deliver output.
Saving Tokens and Money
OpenClaw is agentic, which means it consumes significantly more tokens than a simple chat interaction. When you give it a task, the LLM might generate code, execute it, search the web, fetch content, summarize it, and iterate — all behind the scenes. A one-line answer might cost 50x more tokens than you’d expect.
Choose the right model for the task. Use Gemini Flash for summarization and simple tasks. Use Claude Opus or GPT-4 for complex reasoning, coding, or analysis. Switch models from the Telegram menu as needed.
Set hard spending limits. In your API provider’s dashboard (OpenAI, Anthropic, Google), set monthly spending caps. Without them, a runaway agentic loop can drain your credit card.
Make changes directly when possible. Instead of asking the bot to add a skill (which triggers an LLM call, web search, file generation), SSH into your VPS using Cursor or VS Code, and edit the skill files directly. This costs zero tokens for routine maintenance.
To SSH: connect Cursor to your VPS (File → New Window → Connect with SSH), and you’ll see all OpenClaw’s files directly. You can edit soul.md, add skills, modify configurations — all without burning API credits.
The Mindset Shift
This is the single most important takeaway.
Before OpenClaw, the question was: How do I build tools on top of LLMs? This is how products like Gamma, NotebookLM, and Lovable think. Take an LLM, wrap a specific UI around it, and solve one problem.
After OpenClaw, the question becomes: How do I enable the LLM to expand as an agent? Don’t constrain the LLM to one interface or one task. Give it access to tools, memory, and execution capability, and let it figure out how to accomplish whatever you need.
This is not just a product shift. It’s an architectural philosophy that will underpin the next wave of AI applications — and the trillion-dollar opportunities that come with them. You don’t need to use OpenClaw specifically. But you need to understand this architecture, because every serious AI agent framework going forward will build on these same principles: flexible input channels, model-agnostic LLM layer, autonomous code execution, persistent memory, modular skills, and proactive scheduling.
The people who understand this architecture today will be the ones building and leading the products of tomorrow.
