This kind of stuff should be done by harness, not by the model. There is no context window to fit your project in whole. Each request to the model is fire-and-get-response-and-forget.
My point is the good harness makes DeepSeek2 better than Fable.
mudasobwa
Creator of Cure
dimitarvp
Looking forward to people giving you feedback then. I too feel very powerless lately and started coming to terms with the fact that even a world-class LLM like Fable does forget stuff and its own practice to make memories helps but rarely and only somewhat; regularly forgets the bigger picture, even if it’s really good at tactics.
I am not as far ahead on that journey. Curious what you and other people will bring forward to improve the human+LLM workflows.
0
webofbits
webofbits
webofbits
i have added this issue though feat(skills): add post-beta intent-based selection · Issue #136 · ccarvalho-eng/draught · GitHub
0
dimitarvp
Funny. I just now dropped out of a call with one of my colleagues where I proposed something like this: a tree-like workflow LLM process where the harness itself slaps the agent to verify its work after each step. Claude Code has plan mode but it was so tiring to go through it that I just gave up on it.
My top #1 problem with Fable is this, exactly: it goes in the woods and does everything by itself (or delegates to Opus) and then launches a self-review fleet… and it finds 5 blocking bugs with data and money loss. 
(And let’s not even mention the 17 MEDIUMs and 36 LOWs. 
) I’ll begin trying to use some of the Elixir skills you posted (and gather those from other people) and begin experimenting. Not much time but honestly, I’ll make the time because the amount of slop that’s being produced is just mind-boggling.
1
webofbits
The codex repo was inspired by the claude plugin which should have way more skills by now claude-elixir-phoenix/plugins/elixir-phoenix/skills at main · oliver-kriska/claude-elixir-phoenix · GitHub
1
webofbits
Quick update: Draught’s escript supports the interactive editor correctly, but Burrito 1.6.0 native binaries fall back to line-based input. Burrito currently pipes the BEAM’s stdout on Unix even when attached to a terminal, preventing OTP from enabling raw terminal mode. This affects cursor editing, command completion, multiline shortcuts, and bracketed paste.
The problem is tracked in Burrito #234, with a green fix in PR #235. The fix inherits stdout for interactive terminals while preserving Burrito’s pipe handling for redirected output.
Until that fix is released, Draught’s escript distribution provides the complete interactive experience.
1
webofbits
