My OpenClaw setup that finally works (Complete Walkthrough)
The Startup Ideas Podcast
Secure with Strong Models and Limits
Moritz Kremb outlines local setup advantages, prompt injection basics, stronger model resistance, least-access permissions, and agent-owned accounts.
I sit down with Moritz Kremb, an OpenClaw power user and agency builder based in Berlin, to break down how to actually make OpenClaw useful. Moritz walks through a 10-step optimization guide covering everything from troubleshooting and memory management to model selection and security basics. He then demos two real systems he built with OpenClaw: a full short-form video content pipeline and a conversational CRM. This episode is for anyone who tried OpenClaw, hit a wall, and wants a clear path to turning it into a superhuman digital employee.
Timestamps
00:00 – Intro and episode promise
02:17 – What is OpenClaw
03:17 – OpenClaw vs. ChatGPT vs. Claude Code
07:43 – Where Claude Cowork and Dispatch fit in
09:47 – Why choose OpenClaw over Cowork
11:03 – Step 1: Setting up OpenClaw
14:46 – Step 2: Personalize your workspace files
18:04 – Step 3: Fix and optimize memory
22:43 – Step 4: Choose the right model (OAuth method)
25:56 – Anthropic ban and model provider gray areas
27:33 – Step 5: Organize Telegram groups and topics
30:19 – Step 6: Understand the three browser modes
35:18 – Step 7: Skills — built-in, marketplace, and custom
39:03 – Step 8: Optimize the heartbeat file
42:00 – Step 9: Security basics and prompt injection
48:08 – Step 10: Least access principle and agent-owned accounts
49:52 – Use case 1: No AI Slop content system
58:37 – Use case 2: Conversational CRM
01:01:15 – Final thoughts on the future of personal agents
01:02:55 – Jensen Huang's take: OpenClaw as the new computer
Key Points
Upload the OpenClaw documentation into a Claude project to create a dedicated troubleshooting baseline — it solves roughly 99% of setup issues.
Use the OAuth method (your existing $20 ChatGPT or Anthropic subscription) to avoid expensive API costs, and always configure backup models.
Memory problems are almost always caused by memory never being saved in the first place; add an auto-save instruction to the heartbeat file so it logs every 30 minutes.
Organize your OpenClaw conversations into separate Telegram groups and topics with group-specific system prompts to avoid context bleed.
Stronger models are meaningfully more resistant to prompt injection; pair that with least-access principles and agent-owned accounts for a solid security posture.
Custom skills are the path to real automation — whenever you do something repeatedly, tell your OpenClaw to turn it into a skill.
The #1 tool to find startup ideas/trends - https://www.ideabrowser.com
LCA helps Fortune 500s and fast-growing startups build their future - from Warner Music to Fortnite to Dropbox. We turn 'what if' into reality with AI, apps, and next-gen products https://latecheckout.agency/
The Vibe Marketer - Resources for people into vibe marketing/marketing with AI: https://www.thevibemarketer.com/
FIND ME ON SOCIAL
X/Twitter: https://twitter.com/gregisenberg
Instagram: https://instagram.com/gregisenberg/
LinkedIn: https://www.linkedin.com/in/gisenberg/
FIND MORITZ ON SOCIAL
Youtube: https://www.youtube.com/@promptwarrior/videos
Instagram: https://www.youtube.com/@promptwarrior/
The AI-powered Podcast Player
