Matt's Homepage

Claude 4 Opus has received a level three rating on the company’s four point ‘AI responsibility’ scale – implying the model has ‘significantly higher risk’ than previous models.

While the Level 3 ranking is largely about the model’s capability to aid in the development of nuclear and biological weapons, the Opus also exhibited other troubling behaviors during testing – something already noted in other models.

• In one scenario highlighted in Opus 4’s 120-page “system card,” the model was given access to fictional emails about its creators and told that the system was going to be replaced.
• On multiple occasions it attempted to blackmail the engineer about an affair mentioned in the emails in order to avoid being replaced, although it did start with less drastic efforts.
• Meanwhile, an outside group found that an early version of Opus 4 schemed and deceived more than any frontier model it had encountered and recommended that that version not be released internally or externally.
• “We found instances of the model attempting to write self-propagating worms, fabricating legal documentation, and leaving hidden notes to future instances of itself all in an effort to undermine its developers’ intentions,” Apollo Research said in notes included as part of Anthropic’s safety report for Opus 4.

Many junior engineers type a vague prompt, accept whatever comes back, ship it, pray. But watching a senior developer use AI changed his mind.

But watching a senior engineer vibe coding and a junior engineer vibe coding reveals it isn’t prompting techniques or tool skill. It’s that the senior engineer was already disciplined, and the AI amplifies whatever discipline — or lack of it — you bring.

Here’s a good article on techniques senior developers use (and have almost always used even before AI) that separates them from how most junior engineers work.

Elton Jones uses these prompts when he finds his AI agent keeps forgetting important points when making further decisions:

• The personal memorization prompt (since ChatGPT knows that I’m a journalist, I’ll use this during some of our career-centered chats): You are my continuity editor. Track my writing goals, audience, publication targets, article ideas, previous conclusions, and ongoing projects. Before every response, remind yourself of any relevant context and explain how it affects your recommendation.
• The decision tracking prompt: Maintain a decision log. Record every major choice, recommendation, and conclusion we reach. Reference the log before making new suggestions.
• The conversation brief prompt: Create a one-page briefing document that a new assistant could read to instantly understand everything important about this conversation.
• The running context prompt: Before answering, summarize the key facts, goals, preferences, and decisions we’ve discussed so far. Then use that summary to answer my question.
• The key fact prioritizer prompt: What are the five most important things you’ve learned from me during this conversation? Explain how they should influence your next response.

This guy one-shotted a complete Slack clone as well as server in less than 10 minutes.

Which honestly (I’m going to say it), I don’t understand why folks feel the need to re-invent chat clients, online meeting apps, and source control every 5 years and think they’re ‘disrupting the ecosystem’.

No, you’re repeated making the same stuff that has existed since 1970 – and many times not any better. Just re-writing something using the latest hot tech or UI is not innovation nor actually moving the industry forward.

I prompted Fable 5 with "Create a 1:1 clone of Slack using Typescript, Electron, Tailwind, SQLite. Make it beautiful and functional. Even better than the original Slack." and it oneshotted this UI as well as server in less than 10 minutes. Pretty damn good first shot. pic.twitter.com/EY3sihSq5g

— Dwayne (@CtrlAltDwayne) June 10, 2026