DESIGN.md:

> Each rule below is enforced mechanically by the skill, not left to vibes.

> R1. Repo docs are the memory; not in HANDOFF.md = didn't happen

SKILL.md:

> Not in docs/HANDOFF.md = didn't happen. Refuse to judge results that exist only in conversation or builder chat output.

"Mechnical enforcement" just means "prompting the LLM a bit extra" these days? It (still) amazes me how much effort and tokens we expend on what could and should be a two line script...

Agents are in a wacky state, which makes projects like this fall into a weird spot. Eg I vaguely expect my agent to do two disparate things: manage dependency injection for tools, prompt modifications, etc, but also be the sort of “brain trust” that controls the flow of execution (can we stop now, do we keep going, etc).

This project is meant to be the latter, but there’s not a clean way to integrate that into Claude Code or Codex because they expect to do both.

Pi can do it, but then your users can’t use their Claude subscriptions, so you have to cludgily try to do the same thing via LLM prompts.

I actually just started doing this by having Fable roleplay as Jeff Dean and to use Codex as Sanjay driving the implementation and have them go back and forth. Works really well and it’s cool to see AI pair program

ANNNNNND it's gone. Guys, I found a way to reduce Fable token usage 100%. You can find it here: github.com/USGov/idiotic-overreach.

yes I'm using Fable to inspect, generate plan and architectural docs then using Gemini to implement then have Fable review, find bugs. saving lots of usage.

I don’t know why you’re getting downvoted. It’s true. Averaged across a wide variety of benchmarks Fable is the only Anthropic model that performs better than GPT 5.5 xhigh.

The problem is that there are a bunch of benchmarks, the model providers often don't even use the same benchmarks, a bunch of them have known problems, and it's expensive to do your own benchmarks.

I am a GPT 5.x booster since to me it just feels smarter, and I generally felt like the benchmarks backed me up, but it's not every benchmark, so sadly we're mostly arguing about vibes.

SWEBench-Pro was a big one, though apparently Claude was reading solutions out of the .git folder it wasn't meant to have access to among other problems.

I find it fascinating that every time this kind of discussion comes up, people talk about night and day experiences between Claude and Codex, in both directions. I’m really wondering what people are doing to get such different outcomes.

I’m currently working on two projects/clients one using Claude, one using Codex. I have a strong preference for the latter, but not because I think it is much more intelligent or writes much better code. It is simply because I find the way of interacting with it more pleasant: more literal, mechanical, makes fewer assumption and or double checks, and is less proactive in my experience. At least until some updates over the last few weeks.

I think I like Codex for the same reason tbh. I think it's just general misanthropy or autism or something lol. Most people seem to prefer Claude.

For me, I think Codex was visibly smarter than Claude until 4.8 came out, it would regularly do better debugging and IMO write better code. 4.8 I think is close.

I think Claude is widely regarded to have a big lead in front-end, which I do not work on.

Claude's Ultrathink is pretty cool, though it eats up tokens like nothing else obviously.

/advisor has been really good experience for me especially with having only a Pro plan.

I exclusively use sonnet and advisor is basically “hey opus chime in on my approach”. been working great as far as i can tell.

Reduce Fable tokens by 80%, simply by not using it!

> I am fairly convinced this is the shape serious agent work keeps converging toward.

"this" being "plan with expensive model, implement with cheap model".

Anyone who follows HN would be hard-pressed to disagree; this architecture is re-invented twice monthly.

https://www.facebook.com/groups/vibecodinglife/posts/1946207... https://github.com/openai/codex/discussions/10628 https://build5nines.com/stop-burning-premium-requests-how-to...

> Not because it is aesthetically pleasing. Because every other shape eventually runs into the same boring failures: context rot, self-grading, goalpost drift, and merge chaos.

Actual failure isn't boring. But struggling through a generated software project that celebrates its own genius and doesn't have a single self-critical or genuinely reflective thing to say...at least watching paint dry I might get giddy off the fumes.

I'm not interested in critiquing the project itself, either, you'll just run that through a model, too.

I don't disagree with any of this. It is generated software, and it's not a novel idea. I didn't mean for it to come off like that. It's just solving an itch that I couldn't find a solution to and I'm getting a lot of personal utility out of it. I do have a lot of experience with agentic memory, multi-agent systems and harnesses and wasn't super impressed by the workflow of Fable calling opus subagents so I figured I'd apply best practices to what already exists to make it a teensy bit better and easier to use.

Reducing token usage is this year's "one weird trick". It doesn't make sense on the face of it.

Even if one discovered something that millions (billions?) of dollars of AI compute and the best statisticians in the world was not able to find via exhaustive research, domain search and training... what do you think are the chances this won't be folded into the next update of every model, making the rigmarole moot?

Extraordinary claims require extraordinary evidence and technology-shattering innovations in AI are not know to come from a markdown.

Last night I switched back to Codex for a minute having burned through my tokens for the week with Fable and oh boy I had a terrible experience. Running in circles over simple problems (which I ended up solving myself, like a peasant) and running "terraform apply" several times despite several instructions all over the place to never do that. The performance difference was stark.

I had a similar experience. So far Fable has been a game changer, at least for the work I used it for. Having said that, I think its writing is definitely worse than GPT 5.5. Ethan Mollick also observed the same. He called it more "Claudy." It generates worse academic prose than other frontier models.

Could you provide some details, if possible, like what model & thinking effort, what kinds of tasks? I used to swap between Claude Code and Codex often, and these days use Codex more because of the usage limits. Wondering if I should go to Claude for a month, I get a strange FOMO when I read vague comments like this.

The one major difference I noticed is that the GPT models are more analytical (e.g. better at mathematical analysis, code review) vs Claude models tend to write more straight forward code. Besides that I don't really see any significant differences.

There are a few gotchas with swapping, like being careful with AGENTS.md/CLAUDE.md naming (Claude Code only recognizes CLAUDE.md, and I think Codex only works with AGENTS.md), and updating skill files to match the tool.

I just symlink AGENTS.md and CLAUDE.md

I was using gpt-5.5 high. Writing terraform code for GCP, debugging app launch and Dockerfile issues, that sort of thing. It was going in loops hallucinating features of GCP, looking things up in strange ways, running terraform apply after being explicitly told in the last interaction not to, and overall not solving problems. These were very straightforward tasks and it couldn't be trusted for five minutes. It's the difference in what I would trust an early senior engineer to do vs what I would trust an unreliable high school intern to do.