
Biff 2.0.4: LLM Skills
This release has a hodge-podge of tweaks and fixes based on my own
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
I've just released Biff v2.0.4. This release has a hodge-podge of tweaks and fixes based on my own usage. There is also one big addition: a set of LLM skills. From the starter project's README:
After upgrading your biff.tasks dependency to 2.0.4, these skills will be synced to your local project whenever you run clj -M:run dev.
The most important skill here is biff-code-review. It contains a list of rules for writing code that conforms to Biff conventions. The rules are the result of my own painstaking review of reams of crappy LLM-generated Biff code. From testing them out on a fresh project, I'm fairly pleased with the result. The code I've gotten is definitely still a little weird in places, but it's crossed the "good enough" threshold.
My general flow for making Biff apps is now:
Sometimes I'll use a fancy model to review specs/code and a dumber model to do the implementation.
I had some trouble figuring out a structure for getting LLMs to actually follow my rules. I've had success with this flow:
I try to phrase each rule so it's very explicit/unambiguous/black and white. Either the rule is violated or it's not.
biff-code-review instructs the LLM to copy a checklist for all the rules into your project. The LLM goes through your code and adds a bullet point (with filename and line numbers) to the checklist file whenever it finds a violation. After the code review, any checklist items without violations are checked off.
The LLM then goes through the checklist and applies fixes.
I structured this in a way so that if needed you can do a human-in-the-loop thing: you can edit the checklist before step 3 starts to add any violations the LLM may have missed. During step 3, the LLM doesn't edit the checklist file, so you can manually mark off/not mark off items and then redo step 3.
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。