Viewing profile — pcwelder

The most I enjoy working with AI is my special workflow. I ask it to plan the feature in a separate worktree. In parallel I start coding without being biased by AI and vice versa. …

Could be json or non json. Instead of using tools in API, you ask model to share structured output in text. You parse the string to get the JSON. Gives much more control over thing…

https://news.ycombinator.com/item?id=48496895 Found the time traveller.

Great! Finally follows custom tool call format (k2.6 couldn't). It's a good indicator of instructions following and agentic behaviour. UIs it's generating is pretty good, not witho…

None of the deepseek models are multimodal. How are you guys able to use it in daily work without image input? For example it's just so natural to share screenshots in a chat.

Opus 4.7 selects such palette and motifs by default. Might even be first iteration of claude design.

Opus is not the correct tier to compare this flash model with. On my tasks it has not been as good as even Sonnet 4.6 so far. Instruction following over long context feels worse. I…

It's worth noting that Claude Code itself doesn't use the `insert` tool. (It also uses custom edit tool not the suite's predefined str_replace) Also as a person developing agentic …

To sonnet 4.6 if you tell it first that "You're being tested for intelligence." It answers correctly 100% of the times. My hypothesis is that some models err towards assuming human…

Great work, but concurrency is lost. With search-replace you could work on separate part of a file independently with the LLM. Not to mention with each edit all lines below are shi…

Cool! Please share your work if possible! I couldn't decide on folding and reducing noise so I'm stuck on that front. I believe there is some elegant solution that I'm missing, hop…

All anthropic models. Gemini 2.5 pro and above. Gemini 3 flash is very good too. GPT models can follow tool format correctly but don't keep on going. Grok-4+ are decent but with is…

I had added z-ai in allow list explicitly and verified that it's the one being used.

It's live on openrouter now. In my personal benchmark it's bad. So far the benchmark has been a really good indicator of instruction following and agentic behaviour in general. To …

Each repost is worth it. This, along with John Ousterhout's talk [1] on deep interfaces was transformational for me. And this is coming from a guy who codes in python, so lots of t…

> These are sending all files it can access TBF, Cursor's code indexing works the same way, it has to send all workspace files to their servers. Auto-completion systems need previo…

Sonnet has the same behavior: drops thinking on user message. Curiously in the latest Opus they have removed this behavior and all thinking tokens are preserved.

``` from anthropic.types import MessageParam data: list[MessageParam] = [{"role": "user", "content": [{"type": "text", "text": ""}]}] ``` This for example works both in mypy and py…

Displaying inferred types inline is a killer feature (inspired from rust lang server?). It was a pleasant surprise! It's fast too as promised. However, it doesn't work well with Ty…

To those who are not deterred and feel yolo mode is worth the risk, there are two patterns that should perk your ears up. - Cleanup or deletion tasks. Be ready to hit ctrl c anytim…