tried another round of using Fable today and it was absolutely miserable. It does the same shit as 2023 chaptGPT, but to an insane degree
>take any and all explanations of HOW as a command to do the thing NOW.
>NON-STOP repetition of prior user input. "You did X" (ten replies later) "I DID X I DID X I MUST WRITE TWENTY MARKDOWN FILES WITH INCORRECT REASONING ABOUT WHY THIS IS BAD, GENERATE MEMORIES AND THEN MAKE A SKILL ABOUT X-LESS DOING"
>INSANE levels of miscommunication. worsened by the fact that they don't show thinking and it does not announce which actions it is taking. you can't even see the prompt to a sub agent by default in the main view.
>ignoring current task to just do something else. at some point, it torched a significant fraction of my usage on manually validating every edge case for code that an opus subagent had written and, I believe, already tested, while I was clarifying architectural questions.
You can not use these with other harnesses they enforce claude code. I am honestly wondering what to do with my usage at this point. I will probably use sonnet or something to do strict-instruction step-by-step coding. I absolutely hate the Anthropic models now. Wasn't this bad when I had that trial week at the start of the year.
>take any and all explanations of HOW as a command to do the thing NOW.
>NON-STOP repetition of prior user input. "You did X" (ten replies later) "I DID X I DID X I MUST WRITE TWENTY MARKDOWN FILES WITH INCORRECT REASONING ABOUT WHY THIS IS BAD, GENERATE MEMORIES AND THEN MAKE A SKILL ABOUT X-LESS DOING"
>INSANE levels of miscommunication. worsened by the fact that they don't show thinking and it does not announce which actions it is taking. you can't even see the prompt to a sub agent by default in the main view.
>ignoring current task to just do something else. at some point, it torched a significant fraction of my usage on manually validating every edge case for code that an opus subagent had written and, I believe, already tested, while I was clarifying architectural questions.
You can not use these with other harnesses they enforce claude code. I am honestly wondering what to do with my usage at this point. I will probably use sonnet or something to do strict-instruction step-by-step coding. I absolutely hate the Anthropic models now. Wasn't this bad when I had that trial week at the start of the year.
