I've noticed the time-of-day variance too. My working theory is it's related to load, not model changes. Same prompt at 6am Sydney time (when US is asleep) consistently gets better results than the same prompt at noon. The "ignoring instructions" behavior usually means it's working from a compressed context where earlier instructions got summarized away.
I've had a similar experience - I basically have to tell it to do everything twice now. It especially loves ignoring instructions to re-read a file (or re-read in FULL) and just uses its (stale) cache/context version. It also wants to 'pattern match' instead of actually reading what's provided which leads to lots of really basic logic errors.
Over the last 6 months I've been creating some truly mindblowing apps, apps that feel like they are entirely out of reach for today's Claude.
I've also noticed that if I do some assisted coding at 4-5am UK time, and then perform the same actions at 3-8pm UK time the results are vastly different:
- It takes much much longer to consider my input and work out a response
- Any response given has to be thoroughly considered (I've had to rollback changes)
- Changes scoped in the plan are stubbed or missing entirely
I've put it down to two things (a) Early enshitification, perhaps they simply don't feel the need to provide a consistent level of service because any observation of performance is highly subjective and (b) oversubscription, they scored massive marketing points by being blacklisted by US DoD (despite being integrated into many systems in a different capacity)
I've noticed the time-of-day variance too. My working theory is it's related to load, not model changes. Same prompt at 6am Sydney time (when US is asleep) consistently gets better results than the same prompt at noon. The "ignoring instructions" behavior usually means it's working from a compressed context where earlier instructions got summarized away.
I've had a similar experience - I basically have to tell it to do everything twice now. It especially loves ignoring instructions to re-read a file (or re-read in FULL) and just uses its (stale) cache/context version. It also wants to 'pattern match' instead of actually reading what's provided which leads to lots of really basic logic errors.
I've noticed this too.
Over the last 6 months I've been creating some truly mindblowing apps, apps that feel like they are entirely out of reach for today's Claude.
I've also noticed that if I do some assisted coding at 4-5am UK time, and then perform the same actions at 3-8pm UK time the results are vastly different:
- It takes much much longer to consider my input and work out a response
- Any response given has to be thoroughly considered (I've had to rollback changes)
- Changes scoped in the plan are stubbed or missing entirely
I've put it down to two things (a) Early enshitification, perhaps they simply don't feel the need to provide a consistent level of service because any observation of performance is highly subjective and (b) oversubscription, they scored massive marketing points by being blacklisted by US DoD (despite being integrated into many systems in a different capacity)