Theo - t3.gg says using a CLAUDE.md file to steer Claude agents reduced broken pull requests to zero

@MatthewBerman You can more explicitly steer with the Claude file. I have skills too

Video coming very soon

@MatthewBerman @theo You Can't Do this Properly with Just a Skill or MCP you Need a Proper Harness like

Machinaos: https://github.com/zeenie-ai/MachinaOS

@MatthewBerman @theo I built a skill to do this: https://github.com/fubak/ultraswarm

@theo Will you share all your files?

@MatthewBerman @theo Implemented it already. Absolutely fabulous for token saving and smart use of claude fable.

Here's the full claude.md.

@MatthewBerman @theo Hey I just made a post about this 😁

It’s definitely the way! When do I get the Claude.md lol I’m only on the $20 plan tho

@MatthewBerman @theo Nope I’m doing everything manually

@JimmyNewclaw @theo skill to share?

@MatthewBerman @JimmyNewclaw @theo Here i built something specifically for this

I'm paying $200 for claude and $200 for codex

so why not use both harnesses models

ofcourse it is a work in progress

https://github.com/axatbhardwaj/Dvandva

@MatthewBerman @theo Do you really need skill for this? Imho just telling it to delegate all the work to gpt 5.5 x high or opus is enough. Fable 5 is great at instruction following

@MatthewBerman @theo I'd use both, but for different jobs. CLAUDE.md is the standing bias; a skill is the repeatable handoff. If Fable plans and GPT-5.5 executes, the skill should carry the checklist and examples so the reasoning does not evaporate between models.

@MatthewBerman @theo I've already got this built out to primo optimization. https://github.com/DanMcInerney/architect-loop about as optimized as possible. basically just a supercharged /goal loop that saves 80% off Fable tokens and uses a best practices pipeline of engineering.

This is what I’ve been doing but it still burns through Max 20x tokens so fast. Now I’m thinking maybe to orchestrate with Opus and it calls Fable for more intensive synthesis as a sub agents. I’ve already burned through 80% of my Fable allotment so won’t get a chance to experience it much more unless Anthropic makes it more accessible.

@JimmyNewclaw @MatthewBerman @theo example?

@MatthewBerman @theo Just tell fable that and it will follow, and after first task it will asks you to do the same on next tasks, no need for skill

@MatthewBerman @theo You don’t need that, it will be removed from subscription after 3 days so just use fable for everything now and keep the usage for gpt 5.6

@MatthewBerman @theo The planner's reasoning doesn't survive the handoff. When I tried this split, Fable knew why it made a constraint; GPT-5.5 just sees the output and fills gaps its own way. Three decisions in before you notice.

@MatthewBerman @theo I fed this to CCD, and it built the skill for me. Works great.

@MatthewBerman @theo You could make it a skill but would want to include an instruction in the claude.md on when to use it. I reserve skills for multi step processes and general decision making instructions to the claude.md. You can always guarantee cluade will read the claude.md but not the skill

@MatthewBerman @theo The model is instructed by claude code harness to treat CLAUDE.md as optional. At least that is what it told me in different chats. Not sure if this is still the case. But if it is those instructions turn into a coin flip.