So this is outdated now, right?

[mindplay-dk]

I watched your recent talk about QRSPI and it sounds like this document is now outdated?

https://github.com/humanlayer/advanced-context-engineering-for-coding-agents/blob/main/ace-fca.md

This document was pretty useful for understanding the RPI workflow.

It's very difficult to follow these 20-minute videos and capture anything useful from them - it's all way too compressed and skips way too many details.

Will you be posting something updated? 🙂

[mindplay-dk]

Dex,

welp.

Fable was released and spec-driven development works now - just as you explained it here in the past.

It turns out we just didn't have a model that was good enough to actually implement a spec - and now we do.

My process is something like:

1. deep research with Gemini or Sonnet to gather information, options and references
2. tailored version of "grill me" customized for the task, usually with Opus, weighing all options, asking for clarifications, finalizing all decisions
3. generating the spec (still Opus)
4. gathering references (papers, articles, etc. - by hand, some sources may need conversion from PDF etc.)
5. spec and references goes to Fable for implementation along with opinionated technical requirements

step 5 is where the chain used to fall off - now it just works.

I completed several projects this week, of increasing difficulty and complexity. all implemented flawlessly.

what's really stunning about this model, is it gets things like types and tests correct in the first or second attempt - where other models basically write slop and then trial-and-error their way to "passing tests", often by changing the tests, Fable basically gets 95% of types right the first attempt, makes 1-2 rounds of changes, runs tests, fixes a couple of bugs, done.

of course, the party ends on the 19th when they shut off access to Fable

and no way would I pay $50/M to build these hobby projects

I must have got $300-$400 of inference on my $18 sub 😅

but I feel we're on the way now, finally

if Anthropic can't figure out how to make this fast and cheap, China will figure it out soon enough 😊