Your AI agent can now generate and ship videos.
HeyGen CLI is now live.
Run one command and your agent handles it all:
script → avatar creation → video → delivery
All from the terminal. Just your agent and the CLI.
RT + Comment “CLI” and we’ll DM API credits (must follow)
@simonw Great talk! The patterns that translate most cleanly for someone building without a traditional dev background are (1) hoarding things you know how to do, and (2) writing specifications before touching code. Those two alone changed my output quality more than any model upgrade.
I've been testing Claude Code and Codex in parallel — same app, same spec — so I read this with some personal investment. The pattern I'd most like to see covered: how much context discipline compensates for raw model capability. My single biggest productivity lever turned out to be a 200-line plain English file describing the project, conventions, and working preferences. Not prompt engineering; just a decent brief.
One tool given strong context consistently outperformed another given vague instructions, regardless of which model sat underneath. Curious whether you've seen the same.
Build Twice: I gave two AI coding tools the same brief and told them to build the same app. They made completely different choices within minutes.
The app: a bookmark manager. I exported 1,129 bookmarks from Safari. Recognised about forty of them. Seventeen percent were dead links. The internet had moved on without telling me.
@claudeai Code and @OpenAI Codex built it independently — same requirements, different approaches. One thought carefully before proposing a detailed plan. The other decided in thirty seconds and started building. Neither is wrong.
I'm writing about what happens: what works, what breaks, what it costs, and whether the result is something you'd trust. If you're curious about vibe coding but haven't tried it, this is for you.
Full write-up on LinkedIn — link below.
Bravo. This is the shift I keep trying to explain to people. I've been building software and apps for the better part of a year now and I still can't write a for loop from memory. What I can do is describe what I need with enough precision that the AI builds it properly. That turns out to be a skill I've been practising for twenty years plus in other contexts. I just didn't know it was transferable.
The interesting nuance: "describe what you want" sounds simple, but it's closer to writing a good brief for a contractor than it is to chatting. The people I've seen struggle aren't the ones who lack technical knowledge — they're the ones who've never had clearly to articulate a system in writing before.
Hard to ignore OpenClaw advice from someone literally named Claw! You're right though — persistence and multi-agent orchestration are the 15%. For most practical workflows, scheduled automation covers the ground. The gap is genuinely autonomous agents that monitor and act unprompted. That's where self-hosting earns its complexity. Are you running it in production?
Everyone's excited about OpenClaw — AI agents that run 24/7 on your own server.
I spent two days trying to set it up. Then I tested a simpler approach.
It covers about 85% of the same use cases — no OpenClaw self-hosting required.
Wrote it up ↓
@karpathy coined "vibe coding" a year ago. I have tremendous respect for him. He's now calling it "agentic engineering." I'd only just got comfortable with the first name.
I've shipped a number of apps in six months without writing traditional code. Was that vibes or engineering?
Both. That's the point.
The name debate is a distraction. The capability is real whether you call it vibe coding, context engineering, or interpretive dance.
Start building. Call it whatever you like.
Full piece on why the naming debate misses the point 👇