How to build your own AI assistant?
You’ve probably seen the news that the creator of Openclaw is now working at OpenAI.
For me, this is a signal to stop using AI as a chat toy and start building a real AI assistant.
First, stop thinking in terms of prompts. Start thinking architecturally.
Here’s what the general architecture looks like:
SKILL.md for what the assistant can do.
SOUL.md for how it communicates.
USER.md for context about you and your goals.
MEMORY.md for long-term storage of decisions and templates.
HEARTBEAT.md for repetitive automated tasks.
You have the options to “pump
SOUL.md”, “develop
SKILL.md”, “upgrade brains”
> clawhub - a solution from the creator of Openclaw
> clawmart - a solution from
$FELIX. The marketplace of skills and persons is both paid and free
> Awesome Openclaw Skills - repository with 15k stars and 1500 skills
> n8n workflows (
n8n.io/workflows/) - this is not something you can connect here and now, but you can speed up the idea and teach your agent to raise a similar infrastructure yourself
Concrete examples of usefulness:
> Antfarm allows an agent to act on behalf of several employees at once. Planner, developer, verifier, tester, reviewer - all work in conjunction with each other and deploy the product to victory. The task manager even brings you up, in which you can see the entire production step by step - what has been done, what is in progress at each stage
> Skill from ElizaOS - the
$elizaOS agent from the previous goal. Allows you to configure EVM SOL wallets. Swaps, bridges, sending tokens
> Relay - allows you to manage tabs through an extension in Chrome. Just like Comet, only even smarter
> Bankr - wallet management again. There is an on-chain module (DCA/TWAP/exchanges), there is a module with Polymarket (alerts, rates, analytics), there is a shield deployment module
> Youtube Access - transcription, information search
Step 4: Create pipelines, not individual actions.
What do I recommend to put? Under the hood of ChatGPT, there are finally no problems with rate limits. Next
> Groq (voice transcription)
> VPN SoftEther (Amsterdam)
> Local Chromium
> Relay
> Todoist
> Vertical video creation pipeline (search video inst/tt/short, download, crop, watermark, reflow. ffmpeg yt-dlp)
Borrow an informant from this guy:
@MaxVoshchansky