I don't understand how having two separate instances of Claude helps here. I can understand using multiple Claude instances to work in parallel but in this case, it seems all this process is linear...

▲ layer8 6 hours ago | parent | next [-]

The point is to get better prompt corrections by not sharing the same context.

▲ renewiltord 5 hours ago | parent | prev [-]

If you look at the code it will be obvious. Imagine I’m the creator of React. When someone does “create new app” I want to put a Claude.md in the dir so that they can get started easily.

I want this Claude.md to be useful. What is the natural solution to me?

▲ olalonde 4 hours ago | parent [-]

I'd probably do it like this: ask Claude to do a task, and when it fails, have it update its Claude.md so it doesn’t repeat the mistake. After a few iterations, once the Claude.md looks good, just copy-paste it into the scaffolding tool.

▲ renewiltord 4 hours ago | parent [-]

Right, so you see the part where you "ask Claude to do a task" and then "copy-paste it into the template"? He was automating that because he has some n tasks he wants it to do without damaging the prior tasks.

▲ olalonde 4 hours ago | parent [-]

You can just clear the context or restart your Claude instance between tasks. e.g.:

  > do task 1
  ...task fails...
  > please update Claude.md so you don't make X mistake
  > /clear
  > do task 2
  ... task fails ...
  > please update Claude.md so you don't make Y mistake
  > /clear
  etc.

If you want a clean state between tasks you can just commit your Claude.md and `git reset --hard`.

I just don't get why you'd need have to a separate Claude that is solely responsible for updating Claude.md. Maybe they didn't want to bother with git?

▲

renewiltord 3 hours ago | parent [-]

Presumably they didn't want to sit there and monitor Claude Code doing this for each of the 14 things they want done. Using a harness around Claude Code (or its SDK) is perfectly sane for this. I do it routinely. You just automate the entire process so that if you change APIs or you change the tasks, the harness can run and ensure that all of your sets are correctly re-done.

Sitting there and manually typing in "do thing 1; oh it failed? make it not fail. okay, now commit" is incredibly tedious.

▲

olalonde 3 hours ago | parent [-]

They said they were copy/pasting back and forth. But regardless, what do you mean by "harness" and "sets"? Are you referring to a specific tool that orchestrates Claude Code instances? This is not terminology I'm familiar with in this context. If you have any link that explains what you are talking about, would be appreciated.

	▲	renewiltord an hour ago \| parent [-]
		Ah, it's unfortunate. I think we just lack a common language. Another time, perhaps. You're correct that his "pasting the error back in Claude A" does sort of make the whole thing pointless. I might have assumed more competence on his side than is warranted. That makes the whole comment thread on my side unlikely to be correct.