fourthark

Presumably https://en.wikipedia.org/wiki/S-expression

Strongly agree about the deterministic part. Even more important than a good design, the plan must not show any doubt, whether it's in the form of open questions or weasel words. 95% of the time those vague words mean I didn't think something through, and it will do something hideous in order to make the plan work

This is exactly what the article is about. The tradeoff is that you have to throughly review the plans and iterate on them, which is tiring. But the LLM will write good code faster than you, if you tell it what good code is.

> it is weakest beneath the frozen continent of Antarctica *after accounting for Earth's rotation*.

Why do you find it useless for legacy code? I find I have to give it plenty of context but it does pretty well on legacy code.

And Ask DeepWiki is a great shortcut for finding the right context… Granted this is open source and DW is free.

Is it the specific nature of your work?

What is your argument? There are a lot of bots, therefore humans are no longer in charge?

But only if there is a competent compiler engineer running the AI, reviewing specs, and providing decent design goals.

Yes it will be far easier than if they did it without AI, but should we really call it “produced by AI” at that point?

“This is key!”

Can't tell what this demo is streaming, looks like a static line but it's working hard on something. It can't seem to decide whether to display the top number in red or green either.

Third possibility negates the first two.

Also, did VB6 put anyone out of work?

Maybe you have JS disabled? I see it flash from Jan 22 to Jan 21. :-)

You can have it write the specs and tests, too, and review and refine them much faster than you could write them.

Maybe not quite as simple but you can save and /resume lots of sessions in CC and switch between them quickly.

The end product is... just the website?

I feel like I'm missing something.

Thankfully there is a Typst port of this package!

https://typst.app/universe/package/fleck/

The policy in TFA is a nicer way of saying that.

It's not that hard to fix the area problem with sunburst charts, by decreasing the radius for outer rings.

E.g.

https://github.com/vasturiano/sunburst-chart

Yes as lostdog says, it’s a new feature that writes plans in plan mode to ~/.claude/plans. And it thinks it needs to continue the same plan that it started.

So you either need to be very explicit about starting a NEW plan if you want to do more than one plan in a session, or close and start a new session between plans.

Hopefully this new feature will get less buggy. Previously the plan was only in context and not written to disk.

But PowerShell 5.1 is still the one that ships with Windows.

It will get pattern matching when JS does. Not certain yet but in progress.

https://github.com/tc39/proposal-pattern-matching

True, you have to add guardrails outside the LLM.

Very tricky, though. I’d be curious to hear your response to simonw’s opinion on this.

Yes that’s the point, you can’t protect against that, so you shouldn’t construct the “lethal trifecta”

https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/

We probably wouldn’t be here without it!

Weirdly, I’ve found that when that happens I can close Claude and then run `claude --continue` and now it has room to compact. Makes no sense.

But I have no idea what state it will be in after compact, so it’s better to ask it to write a complete and thorough report including what source files to read. Lot more work but better than going off the rails.

Most people don’t take this opportunity, though.

> It’s a lot easier than working with actual people because you don’t have to worry about offending or discouraging them

But you have to worry much more about confusing LLMs by introducing contradictory ideas or talking about too many things at once.

They will not call you out and they will not recover.

You don’t need to reboot your coworkers.

Yeah if it’s obscure stuff you might have to guide it, show it just the relevant code / context. Outline the design for it.

They seem autonomous but often aren’t.

Argh

I’m curious what you mean. It’s a language with an open source implementation.

Others can and do implement parts of the language - e.g. Pandoc implements a parser and evaluates Typst.

Seems like training would be a better match, where you need tons of compute but don’t care about latency.