pull down to refresh
I truly don't know, it seems like that tasks that require more than 10 minutes are more prone to this behavior. Changing the effort level helps a little but not so much, it's like it gets convinced about a fact during reasoning and never checks it again.
The weirdest part is that fighting Opus on presented lies becomes a gaslighting game and it's really frustrating lol. Fable is much, much slower but it's less prone to lies (never happened to me thankfully) and absolutely won't fight me if I say that 1+1=2 (simplification)
Hmm I have extensive prompts and don't allow it to guess much, which is why anything under xhigh is a non-starter since 4.8 because it ignores much of the instructions on high and just applies a template. False positive rate, after extensive tuning sits under 5%. I know because I validate everything. I am a slop-reading monster.
Important other principles I'm applying:
- Never use chat - only instruct and send it off.
- 1 instruction, 1 session, clean workspace. The extra tokens and time spent on
git cloneare worth it. - When it "lies" or introduces a bug, I throw it out fully. Remove the output from all visibility. This prevents re-poisoning.
- Turn off all "memory". Memory is auto-poison.
With this, you get a deterministic-ish Claude. One major thing, also for Fable:
I found that somehow, when I run out of credit on the 5h window (which happens multiple times per day now that I have downgraded to pro) it often gets to finish at 97-99%. This means that somewhere it sacrifices precision based on that window and I think that this is a Claude Code "feature." Not a big problem if you're on 20x.
fighting Opus on presented lies
Do you argue with psql or do you just tell it what to do? If you argue with psql, more power to you. If you don't, replicate your usage of psql to Claude and you will be a lot happier. It is software, not magic.
everybody lies, but I mean Opus really lies all the time