pull down to refresh

I tried Claude Opus 4.8 High Effort to try and debug an unexpected data download error, and it ended up eating 90% of my usage limit while providing me with pretty useless thoughts that I'm pretty sure is wrong. I'm pretty sure we're not close to AGI

Here's what you want to spec in your prompt to Claude Opus High, in order:

  1. Objective. 1 sentence. this keeps the whole thing aligned.
  2. Context. List containing every relevant link / file / docs / repos and what it contains. This limits search scope.
  3. Constraints. List or small paragraphs describing the specific actions you need done. This limits "creative non-answers" because it literally tells it what to do
  4. Output format. Just tell it if you want a table, prose, bullet points and if you want markdown in a file or just a reply. I generally ask for markdown "report". If I'm looking for something specific, I spec a table.
  5. Acceptance criteria. Reitererate output format, constraints, objective in a condensed form. I often add "leave no stone unturned", but this is expensive as it easily takes 30-40 minutes on big jobs and will eat all your credits
  6. Subtask decomposition. If it's a big job, specify how to break it down (i.e. what does the main task do, what do subtasks do, how to consolidate results back into main task) Leave this out for simple jobs.
reply

Dang. That doesn't sound much like intelligence to me, though.

reply

That's because it is autocorrect++?

It's not intelligent, it just simulates/fakes it. So you have to parametrize the simulation.

reply

I want you as spokesperson for all these AI companies.

reply

That would be fun. Probably some VC bros will put a bounty on my head tho. They'll lose trillions.

reply

remember that openAI guy who was mysteriously found dead?

reply

Yes. So we don't have to wonder whether it will happen.

reply