Jon

[Jon @jnpkr]

@_sholtodouglas Anthropomorphic framing like “in my experience” or “I think that’s a really under-explored idea” when there’s zero basis for those claims. It makes the model feel less trustworthy overall when I know it’s bullshit and not actual evidence/judgement.

[Jon @jnpkr]

@_sholtodouglas @aether_oracle It’s not just that it says “this will take a week” — it silently makes suggestions and decisions based on that same internal framing which is more dangerous because you don’t know it’s happening

[Jon @jnpkr]

@trq212 The problem is that, with each abstraction layer (sub-agent hand-off, agent-teams, compaction prompt, deep research agent) the ability to control — and in some cases even see — the prompts is lost to some degree

[Jon @jnpkr]

@ishi__singhal @claudeai `opusplan` has been available as a Claude code model setting for a while