4 ms·

Sent to solve one task, came back with half of it solved and 2 more problems.

by visarga 2mo ago

capnjazz 2mo ago
"One thing worth your attention", "Two things worth knowing", "One thing to eyeball"
- greenchair 2mo ago
  This is driving me crazy. opus 4.8 did not do this to me not (at least during pre-5.0 timeframe). Feels like the new cycle is one step forward, two steps back.
- FridgeSeal 2mo ago
  And one of them is always something just completely out of scope and the other is something obvious it missed. “One thing worth your attention, if you were to detonate a pipe bomb in your house, it would have a negative effect on your living room”.
- ethin 2mo ago
  I don't use Claude code, just Claude web, and I get this all the time. Or (since I have it push me to actually think) it will ask me some question in our back-and-forth, and then right after it'll provide the answer. As a "hint". Like come on
vunderba 2mo ago
What really enrages me is the amount of effort it puts into justifying weaseling out of work. (THAT'S MY JOB!) It will do everything it can to defer or push it off, to the point where I’ve had to add multiple imperative directives to the AGENTS file telling it, in no uncertain terms, not to defer tasks under any circumstances.
- cyanydeez 2mo ago
  sounds like someone needs a local llm.
  - vunderba 2mo ago
    Oh I do. Headless 128GB RAM machine serving llama.cpp with a number of local models that I use on a daily basis. • Qwen3-VL picks up new images in a NAS, auto captions and adds the text descriptions as a hidden EXIF layer into the image, which is used for fast search and organization in conjunction with a Qdrant vector database. • Gemma3:27b is used for personal translation work (mostly English and Chinese). • Some small 8b models (like llama3.1) for sentiment analysis on text. But haven't really tried using local LLMs in conjunction with agentic harnesses yet.
    - cyanydeez 2mo ago
      recommend opencode w/qwen 35B or 27B with MTP. My secret sauce is to use LLAMAcpp's reasoning-budget and reasoning-message that trigger cut off to overthinking with a message that says to either us subagents or compress the context. opencode's dynamic context pruning plugin can get you pretty far into the stratosphere.
      - vunderba 2mo ago
        > My secret sauce is to use LLAMAcpp's reasoning-budget and reasoning-message that trigger cut off to overthinking with a message that says to either us subagents or compress the context Thanks for the tip - I like this a lot. I remember having to do a lot of tweaking to curtail Qwen QwQ-32b when it would go down an endless psychotic recursive reasoning loops as part of its "chain of reasoning."
        cyanydeez 2mo ago
        I tailored the agent to ao both its system prompt and budget-message align. I havent yet tailored the pruning messages, but mostly it works. Reasoning budget can also be set by client, so potentially smarter.

Sent to solve one task, came back with half of it solved and 2 more problems.

by visarga 2mo ago

Sent to solve one task, came back with half of it solved and 2 more problems.

capnjazz 2mo ago

"One thing worth your attention", "Two things worth knowing", "One thing to eyeball"

greenchair 2mo ago
This is driving me crazy. opus 4.8 did not do this to me not (at least during pre-5.0 timeframe). Feels like the new cycle is one step forward, two steps back.
FridgeSeal 2mo ago
And one of them is always something just completely out of scope and the other is something obvious it missed. “One thing worth your attention, if you were to detonate a pipe bomb in your house, it would have a negative effect on your living room”.
ethin 2mo ago
I don't use Claude code, just Claude web, and I get this all the time. Or (since I have it push me to actually think) it will ask me some question in our back-and-forth, and then right after it'll provide the answer. As a "hint". Like come on

vunderba 2mo ago

What really enrages me is the amount of effort it puts into justifying weaseling out of work. (THAT'S MY JOB!) It will do everything it can to defer or push it off, to the point where I’ve had to add multiple imperative directives to the AGENTS file telling it, in no uncertain terms, not to defer tasks under any circumstances.

cyanydeez 2mo ago
sounds like someone needs a local llm.
- vunderba 2mo ago
  Oh I do. Headless 128GB RAM machine serving llama.cpp with a number of local models that I use on a daily basis. • Qwen3-VL picks up new images in a NAS, auto captions and adds the text descriptions as a hidden EXIF layer into the image, which is used for fast search and organization in conjunction with a Qdrant vector database. • Gemma3:27b is used for personal translation work (mostly English and Chinese). • Some small 8b models (like llama3.1) for sentiment analysis on text. But haven't really tried using local LLMs in conjunction with agentic harnesses yet.
  - cyanydeez 2mo ago
    recommend opencode w/qwen 35B or 27B with MTP. My secret sauce is to use LLAMAcpp's reasoning-budget and reasoning-message that trigger cut off to overthinking with a message that says to either us subagents or compress the context. opencode's dynamic context pruning plugin can get you pretty far into the stratosphere.
    - vunderba 2mo ago
      > My secret sauce is to use LLAMAcpp's reasoning-budget and reasoning-message that trigger cut off to overthinking with a message that says to either us subagents or compress the context Thanks for the tip - I like this a lot. I remember having to do a lot of tweaking to curtail Qwen QwQ-32b when it would go down an endless psychotic recursive reasoning loops as part of its "chain of reasoning."
      - cyanydeez 2mo ago
        I tailored the agent to ao both its system prompt and budget-message align. I havent yet tailored the pruning messages, but mostly it works. Reasoning budget can also be set by client, so potentially smarter.