Zach Cox

🏔️🤖

[Zach Cox @zcox]

opus for show, sonnet for a pro

[Zach Cox @zcox]

i'll buy an @AnthropicAI eng their fav drink to RL "add target/ to .gitignore in new Rust project before first git commit" into claude

[Zach Cox @zcox]

@TheVixhal progressive disclosure ftw! very cool

[Zach Cox @zcox]

The last GA (non-preview) release of an LLM from Google was gemini-2.5 in June 2025. 9 months ago.

Since then we've seen Claude 4.5/4.6 and GPT 5/5.1/5.2/5.3/5.4.

Why isn't Google releasing any more models in GA?

[Zach Cox @zcox]

coding agents won't kill open source

which is faster?
1. fully research, spec, plan, generate, and verify your own bespoke standards-compliant http stack and web application framework in Rust
2. use axum

repeat for json, sql, s3, kafka, OS, containers, k8s, etc

[Zach Cox @zcox]

gm

[Zach Cox @zcox]

I prefer "verification" over "backpressure" for coding agents

backpressure implies resistance. but we want faster loops, not slower

verification provides signal back to the generate step. fewer, faster iterations to convergence

[Zach Cox @zcox]

example: https://t.co/ugJbqlc2bw

[Zach Cox @zcox]

a single IMPLEMENTATION_PLAN .md grows unbounded and wastes context, or requires constant lossy maintenance

instead, decompose tasks within specs so agents can explore via progressive disclosure

and track status via emojis:
🔲 Not started
🔄 In progress
✅ Complete

[Zach Cox @zcox]

dir structure looks like:

[Zach Cox @zcox]

@headinthebox while you are still experimenting with a loop I extended mine to a spiral and added rotation in a second orthogonal plane to form a clifford torus. my claude Code now optimizes skills across the multiverse.

[Zach Cox @zcox]

OG surely has better magic spells than me

but I've made it pretty far reading posts, watching videos, putting in reps, incrementing weight, burning tokens

the techniques are effective on 5-10yr old brownfield enterprise Scala services. but always hungry to learn more!

[geoff @GeoffreyHuntley]

as soon as i post it

corps will rebundle it and sell it

where’s the upside for us et al topics

heck i even had a frontier lab in my DMs

after posting the above tweet

i was seriously considering moving to SF

to join them but they said no

but now they are in my DMs asking for the solve lol 💀

[Zach Cox @zcox]

And then CC responds to that with a super dope haiku about penguins.

[Zach Cox @zcox]

I assumed Claude Code put the skill content in a tool result, but it's actually in a synthetic user message sent after a tool call.

Using @LangChain LangSmith's Claude Code hook you can see the tool call & user message.

I'm curious why the user message? Drives the model forward more effectively than a tool result?

[Zach Cox @zcox]

You can see that CC called the Skill tool with argument "skill: write-dope-haiku" and the result is only "Launching skill: write-dope-haiku".

The following user message (which I didn't send) contains:

Base directory for this skill: /path/to/skill

<contents of SKILL .md file>

[Zach Cox @zcox]

1. Ralph correctly
2. optimize loop

fewer faster iterations

common slow parts:
- finding next task
- starting/stopping servers, dbs, etc
- git commit after task completion
- Claude's affinity for giant summarization of accomplishments

ruthlessly minimize or convert to scripts

[Zach Cox @zcox]

I really like the progressive disclosure structure of Loom's specs https://t.co/wSR5xGWdnQ

Seems like a good idea for plans/tasks too. Single IMPLEMENTATION_PLAN .md gets too big.

[Zach Cox @zcox]

agent memory in action

[Zach Cox @zcox]

key component of GH's Ralph: self-updating status checklist

plan .md contains:
# Status
- [x] Phase 1
- [ ] Phase 2
- [ ] Phase 3

prompt .md contains:
study @ plan .md complete the next task only then update the status checklist

cat prompt .md | claude