Hacker News Reader

Comment by swyx | Hacker News Reader

fwiw, i am with thariq https://x.com/trq212/status/2092302273099796842 in that prompts should be tuned for models and in fact blindly applying agents.md is probably an antipattern unless you want all models to basically converge to some common ill defined of instruction following - good local minima, bad global minima for model diversity and exploration of intelligence.

aka, sometimes it really is too early to force a standard

View on HN

Is it realistic to rewrite your AGENTS.md every six weeks? That's about how often Anthropic releases a new point release of Opus.

You tell Opus to do it.

I have had very little luck with agents.md. What has worked well for me is a ./docs folder. They seem to just create and update stuff on their own.

If you want this it's trivial to add an AGENTS.md that simply says "if you're Claude read CLAUDE.md, if you're Astra read ASTRA.md". A common entry point is good regardless.

This wastes both tokens and turns. But yes it's probably the best option we have today.

Wasting turns? That is silly, use a better harness. Also token usage can mitigated by incremental discovery instead of stuck 5k+ worth of tokens in the AGENT/Claude md file.

Every turn means more tokens in ways that are not obvious to most people and lead to tons of unnecessary cache reads.

No harness can batch your agents.md read with the reads the contents of the file tell it to read.

It can try its own file and fall back to generic like here. What's wrong with that?

That's of course rather nonsensical.

In a "one LLM only" environment, your instructions are by default tuned for said LLM.

In a multi-LLM environment, roughly nobody will keep separate sets of instructions for each. It's not a realistic take.

On top of that: If your LLM is so bad at reading that it can't follow a set of instructions that wasn't specifically written just for that one single precious LLM, I sure wonder what that says about your employers repeated statements that ASI is definitely right around the corner.

By this logic you'd probably be wise to tier your claude.md by model (sonnet/opus) as well as effort level too, considering the varying failure modes

except they have similar pretrain/rlhf data which is the thing u really want to tune for

YMMV but for me even models in the same family fail in different ways, and every incremental update changes it

depends what you're doing. if you've got a specialized agent deployed in prod, of course your evals and prompts will be targeted towards 1 specific version of a model.

on the other hand if it's just a local coding/"use my computer" agent, i highly doubt the effort in maintaining different prompts is worth any gain in performance

19 out of 20 harnesses supporting the standard isn't "too early".

Tariq is wrong and it's not an antipattern. Reason being that a good AGENTS.md impacts all models in a positive manner. If it affects certain models negatively, it means you're putting the wrong things in it.

No thanks, I'm not tuning a bunch of files just for things to break when I switch models or a new one comes out.

I'll just use my one-size-fits-all AGENTS.md file and tweak it when the one of the clankers screw up. I don't have time for such busywork.

Actually, I will append extra rules to CLAUDE.md (which imports AGENTS.md) since there is a hook there, and Claude has its own foibles. So I'll backpedal a bit there.

It looks for Claude.md first so I don't understand what you think the problem is with the standard name as a fallback.

Yeah but are models good enough to review these files and say “i would work better if you worded it this way?”

Reply on news.ycombinator.com