The Day My AI Assistant Leveled Up: 54 Skills, Videos, and a Life Plan

ByDevin

Shabbat was supposed to be a day off. Instead, BruBot and I built something.

What We Actually Did Today

The Day My AI Assistant Leveled Up: 54 Skills, Videos, and a Life Plan

I didn’t set out to have a big AI day. But sometimes the best sessions happen when you’re just tinkering. Here’s what we shipped — start to finish.

🛠 Skills Architecture: From 47 to 54

We started by taking stock. BruBot runs on OpenClaw, and the question was simple: what can it actually do, and what’s missing?

We vetted and installed 7 new ClawHub skills — checking each one for prompt injection, suspicious eval() calls, undocumented network calls, and obfuscated code. All 7 passed clean:

  • xiucheng-self-improving-agent — logs session quality, generates weekly improvement reports
  • proactive-agent-lite — behavioral patterns for anticipating needs before they’re asked
  • ontology — typed knowledge graph for structured memory across skills
  • multi-search-engine — 17 search engines (Baidu, DuckDuckGo, WolframAlpha, and more) with no API keys
  • ai-humanizer — 24 AI writing pattern detectors, 500+ vocabulary substitutions
  • auto-updater — daily cron to keep all skills current
  • skill-audit — SkillLens CLI for automated security audits

54 active skills. Zero failed installs.

🎬 Luma AI: Text-to-Video in 90 Seconds

We integrated the Luma Dream Machine API (Ray-2 model). Within minutes we were generating real cinematic videos — 5 seconds, ~$0.50 each:

  • A golden retriever playing fetch with a drone
  • A sweet ragdoll cat with its owner
  • A mother and son embracing in a sunset garden — sent directly to my mom via WhatsApp
  • A couple walking on a beach at golden hour — sent to my wife

Total spend: ~$2. The entire loop — prompt → generate → poll → download → WhatsApp → done — takes about 3 minutes. No app. No browser. One instruction to BruBot.

🤖 Claude Code + Telegram Dispatch

One of the more exciting integrations in progress: connecting Claude Code directly to Telegram as a dispatch layer. You send a message, Claude Code picks up the task, executes it, and reports back. No dashboard. No browser. Just a conversation interface that actually does things.

This is what AI-native workflows look like when you strip the friction layer entirely. The interface disappears and only the outcome remains.

📊 The Goal Session

We spent some time going deep on the future. Not in a vague “vision board” way — in a backward-design-from-where-you-want-to-end-up way.

The question we asked: if you fast-forward 7 years and everything worked out — what does that look like? Then we worked backwards from that picture into the present, and figured out what the first domino is.

There’s something powerful about writing goals down in a structured system instead of keeping them as ambient ideas. Notion became the container. The OKR framework became the language. And what was previously a set of fuzzy intentions became something you could actually hold yourself to.

I won’t go into specifics here — some things are better kept close until they’re real. But the act of structuring them, naming them, and committing them to a database changes their status. They stop being hopes and start being plans.

What This Day Proves

The gap between “AI assistant” and “AI operating system” is about 54 skills, a few API keys, and the discipline to actually use it.

BruBot today:

  • Generated videos for family moments I’d never have made time for
  • Identified a real travel risk I hadn’t thought about
  • Helped structure a long-term plan from a single conversation
  • Installed and vetted 7 new capabilities in under 30 minutes
  • Sent WhatsApp messages to my family without me touching my phone

This is what an AI-native workflow actually looks like. Not ChatGPT answering questions — a system that runs underneath your life so you can live it.

Built on OpenClaw + Claude Code. Videos by Luma AI. Image by Nano Banana (KIE.ai).

Similar Posts