simonw

17 ms·

361.
by simonw 2mo ago
That works for Claude Code but doesn't work for https://claude.ai Claude Chat. From my original comment: > They also don't publish the Claude Code system prompts, which is silly because those are trivial to extract
362.
by simonw 2mo ago
You mean this? draw an svg of a circle
363.
by simonw 2mo ago
It should just about be usable in 32GB.
364.
by simonw 2mo ago
These are the prompts that Anthropic publish on https://platform.claude.com/docs/en/release-notes/system-pro...
365.
by simonw 2mo ago
Was that with the default xhigh reasoning setting? I suggest trying again with reasoning set to low or turned off entirely.
366.
by simonw 2mo ago
Comparing with Muse Glimmer is a good idea. I ran the same exact HTML tool generating prompt against both Glimmer 30B and Qwen 3.8 27B. Results: Qwen: https://gist.github.com/simonw/121ad098860028b2fab603fa12da1... - 1
367.
by simonw 2mo ago
I had to paste this into Claude and ask it to filter out the core idea. Here's the most important paragraph quoted from the article: > The one repository with working CI has a workflow file of about 40 lines that runs a type check,
368.
by simonw 2mo ago
It's an indictment of writing only unit tests. If the system also had integration tests this would likely have been caught, but even more important is the lesson that no amount of automated tests is an excuse not to actually try usin
369.
by simonw 2mo ago
That's not a very common pattern, to be honest. A bunch of people have been experimenting with automatic model routing recently but mainly as a cost optimization, since tokens for the best models have got expensive once you start pipin
370.
by simonw 2mo ago
For following instructions I think the main thing is to work with strong models (Opus/Fable 5, GPT-5.6) and then practice with them and build up intuition for what works. I find they mostly do exactly what I ask them to, but I think th
371.
by simonw 2mo ago
Yeah, Claude chat does show little in-app messages occasionally warning that Opus or Fable will burn through your rates faster.
372.
by simonw 2mo ago
I don't understand. System prompts are part of the software that customers pay to access. Complaining about that is a bit like complaining that your Netflix subscription includes paying to execute the compiled code that Netflix wrote t
373.
by simonw 2mo ago
Pretty much everything in Pi is handled by extensions. There's an extension for customizing the system prompt here: https://pi.dev/packages/pi-custom-system-prompt My guess is that harnesses don't make core s
374.
by simonw 2mo ago
No, because these ones affect the consumer chat products and not the API or Claude Code. (Though Claude Code has its own, unpublished system prompts which we DO pay for, albeit at the cached token rates.)
375.
by simonw 2mo ago
These system prompts don't affect the API, they are for the Claude consumer chat products. We aren't charged extra for them. They're also prefix cached, so the cost to Anthropic and performance hit is greatly reduced.
376.
by simonw 2mo ago
I have a folder where I rebuild these as a git commit history so you can more easily see what has changed: https://github.com/simonw/research/commits/main/extract-syst... For example here's what cha
377.
by simonw 2mo ago
Have you seen that recently? I used to see that happen, but I've not caught it with the more recent (Opus 4.5+, Fable 5, GOT 5.5/5.6) models.
378.
by simonw 2mo ago
Tell it to use red/green TDD and start things off with an already configured test suite, maybe with a single test that asserts 1+1==2. Make sure it know how to run the tests before it starts writing any additional code. Then set it a c
379.
by simonw 2mo ago
You mean they have zero management experience at all?
380.
by simonw 2mo ago
I think of leadership as distinct from management. Importantly, I think you can be in a position of leadership without needing to be in a position of management. Management requires direct reports: you make decisions which directly impact t
381.
by simonw 2mo ago
I think this comparison holds up. I've noticed that a lot of people I know who get really good results from LLMs and agents are people with significant people management experience. Companies like Anthropic seem to understand that too.
382.
by simonw 2mo ago
Anyone seen this show up in any APIs yet? I'd love to try it out faster than my Mac can run it.
383.
by simonw 2mo ago
It did a pretty outstanding job on that: https://tools.simonwillison.net/markdown-svg-renderer.html#u... This took 56 minutes! - but it didn't have exclusive use of LM Studio, I ran a few other test prompts while it wa
384.
by simonw 2mo ago
IMPORTANT UPDATE: This was my fault. Gemini 3.7 Flash produces SVG that works in all three browsers. My vibe-coded tool for rendering those SVGs in a document was stripping some attributes. I've fixed that bug now and https:/&#x2
385.
by simonw 2mo ago
For anyone who followed yesterday's Gemini 3.7 Flash pelican which rendered in Safari but not in Firefox or Chrome... https://news.ycombinator.com/item?id=49289112#49290012 . - that turned out to be my fault, not the m
386.
by simonw 2mo ago
Absolutely the best pelican I've seen from a model that runs on my laptop: https://tools.simonwillison.net/markdown-svg-renderer#url=ht... Bicycle is the right shape. Pelican beak is excellent. Nice background. Most im
387.
by simonw 2mo ago
I've been trying to think of a UK musical act that works as well for "Nationalize Adele" and I've got nothing.
388.
by simonw 2mo ago
I think this entire disagreement is about language, and in particular the implications of the term "boring". The term we are looking for here is something that means: "A piece of technology that has been around for long enoug
389.
by simonw 2mo ago
Discussed here five days ago: https://news.ycombinator.com/item?id=49079995 (Honestly though this is such a cool piece of work it deserves a second conversation, that one only hit 40 comments.)
390.
by simonw 2mo ago
LLMs aren't architected to handle filter-style comprehensive search without setting them up with additional tools. Asking an LLM for a list of every county in the USA for example, or every county with a population of more than 100,000

More ›