What’s the context length you use typically? I use sonnet with 200k context length but it’s easy to saturate it in a session. After a while, the model forgets what’s going on and I have to start a new conversation
In your process, you explicitly put human in the loop. Do you think that is sufficient to gain intuition and experience? Or do we need to adopt some new practices? e.g. using LLMs to quiz us?
It includes detailed examples, including the full Claude Code process I used to build this new Colophon page, which presents the Git commit histories for each of my collection of LLM-assisted web tools in one place https://tools.simonwillison.net/colophon
I’ve been following your stuff for a while now and your articles continue to be the most measured and grounded-in-real-world-experience takes when it comes to practical LLM applications. No hype, no doom, no dismissal. Thanks for being a standard bearer here.
Oh, great! And have you tried more general system prompts to guide the conversation and output in a structured way? Like returning as output a diff, like what some code assistants do
Comments
https://mattsayar.com/i-didnt-want-to-pay-for-a-newsletter-email-service-so-i-built-my-own/
In your process, you explicitly put human in the loop. Do you think that is sufficient to gain intuition and experience? Or do we need to adopt some new practices? e.g. using LLMs to quiz us?
I was hesitant to do this at first but the utility of the resulting explanations convinced me it was worthwhile
More details here: https://simonwillison.net/2025/Mar/13/tools-colophon/
Does this imply you used Claude Code? Or your llm CLI
I use LLMs largely for coding these days. I’m an amateur at best and they really enhance what I am able to put together.
... but the code they write these days is pretty great
(I don't personally work with AI generated images outside of occasional self-amusement)