The title is the worst part of the essay. Which is super interesting, to wit: fable's a very capable model when it comes to transforming concepts in and out of structural descriptions, and you can use it (along with a test suite I presume) to transpile a codebase. I wouldn't have thought to do this, but I think it makes sense, and I like it - it's using model intelligence at a few different steps for sensible things. Thanks for the writeup. De clickbait your title though or you'll keep getting clickbait rage responses :)

▲ aka-rider 4 days ago | parent | next [-]

I'm glad you liked it. I was trying to keep both, the title and the content as straightforward as possible.

Another point is, Fable is reasonably cheap if you don't allow it to read or write.

▲ vessenes 4 days ago | parent [-]

Thank you. Yes I would like to hear more about that - I’d be interested to read details about limiting some of its native tool use.

▲ aka-rider 3 days ago | parent [-]

Fable is good at instruction following, so just adding the snippet I mentioned in the post is good enough for the most cases.

There is a way to control tool invocations at the harness level when writing skills or agents.

Example: ~/.claude/agents/critic.md

    ---
    name: critic
    description: >
      Plan Critic. Reviews an implementation plan. Use before the implementation.
    tools: Read, Agent
    ---

For Claude skills, the frontmatter is different, the keys are:

    allowed-tools: Read Grep
    disallowed-tools: WebSearch Glob

https://code.claude.com/docs/en/tools-reference

▲

vessenes 3 days ago | parent [-]

thank you!

	▲	aka-rider 3 days ago \| parent [-]
		I'm glad to help

▲ dilyevsky 3 days ago | parent | prev [-]

Ime it remains the best model for prototyping, sims, visualization etc throwaway scenarios. Just don’t look at the generated code if you just let it rip on the problem for hours ;) but that advice applies to any model unfortunately

	▲	aka-rider 3 days ago \| parent [-]
		> if you just let it rip on the problem for hours In my experience, you can get good results if Fable does't write code itself, only spawn subagents. I can run Fable for 10 hours, and it would output 50k tokens and read 300k (30% of the context window). The resulting code is okay-ish. I would rarely merge LLM-produced code first try without an adversary review.