@BraaiEngineer Maybe it's a sign to move away from claude code? It's a very useful feature to have this.
English
Senior AI Micromanager
2K posts
Senior AI Micromanager
@chrisczynski
Devops/tech consulting and available for projects! Haskell/Nix ideally. Running some projects at -$20 MMR.
UTC 0 Katılım Ağustos 2013
1.3K Takip Edilen244 Takipçiler
Why doesn't Claude Code have Steer like Codex yet? It's so annoying to kick off a long-running task and you have to completely stop it to steer it.
English
@rickasaurus "deploy the app"? Thats oldschool. We're vibe coding live on production now. Don't get left behind.
English
Love putting my agent to old code. Pi restored one of my first ever Open Source web applications.
English
@escartin Strongly agree. My side project is something like this. And I'm having phenomenal results and cost effectiveness even with deepseek models alone. The oh-my-pi harness is also doing interesting things along this line (having an "advisor" model run in parallel.
English
This business thesis has been in my mind lately:
- There is a generation of cheap AI models that can take on a lot more business tasks when properly managed than out of the box.
- New models keep getting more capable so they can oneshot more of those tasks by the day, but that is coming with a steep price increase. Fable and other SOTAs are an example of this.
- Not all those tasks are worth the spending and even if ROI+ nobody wants to waste money anyway
- So there are moats in doing stuff that SOTA models could do with thousands but with cheaper models + smart orchestration, decision trees, guardrails and engineered prompts.
Think of an "enhanced wrappers" category that are beyond UI/UX.
English
"If you want next work" ?? It's hard to know if Grok is genuinely trying to condense information or it's a weird quirk.
Senior AI Micromanager@chrisczynski
Grok 4.5 seems kinda bad at communication. It's like I'm speaking to HR or something. Uses horrendous wording.
English
@escartin It's not only limited just to builder/entrepreneur. I think it genuinely accelerates things and one's dopamine reward as well, a few friends of mine report doing late nights, compulsive "one more turn" with LLMs. I expect we'll see a lot more of this until we all adjust/balance.
English
Nothing personal against these guys, but if you're the type of person that can end up in a hospital because of AI model access, deadlines, the market, and anything business/money related really, you'll be better with a job you don't hate and finding meaning elsewhere.
The builder/entrepreneur life is only healthy if you can detach from projects and outcomes a bit, like relativize what your tiny app and professional success mean in the big scheme of things.
Personally, I find my mental escape in family-friends-civilization.
Any business problem gets a "so what?" against those pillars and immediately gets demoted to where it really belongs.
It's one of the best ways of life. Don't let it kill you.
English
@thdxr I also rewind the context often if too much bloat gets picked up.
English
@thdxr Given how new and immature LLM tech is, micromanaging agents can be quite useful I'd argue. I'm starting to use trees of models to accomplish things, and knowing how each layer or different model operates is quite useful.
English
Wait... So no one is doing literate programming documentation now? Even though this is literally the easiest thing ever for LLMs to generate? And ensures you have great docs that can't easily get out of phase from source code?
English
@valyala Yes I meant after it's pushed though :). If git natively supported decoupled git messages that'd be really cool.
English
@chrisczynski You can freely amend commit messages until the local commits are pushed to the upstream git repository. See stackoverflow.com/a/1186549 .
I frequently change the local commit title and description with `git commit --amend` before pushing it to the upstream repo.
English
I feel the same. Histrically software engineers weren't good in writing good commit titles and descriptions. So typical commit titles were "WIP", "fix", "refactor", "chore", etc. Now I see detailed commit titles and very long commit messages with many points, which describe all the changes made by the commit. The problem is that these new titles and messages are mostly useless (and even harmful) because of the reasons outlined by Kenton below.
Make an effort and write good concise commit title and description by hand. If you cannot do this, then it is better to use "WIP" title instead of AI-generated misleading slop.
Kenton Varda@KentonVarda
I just declared a moratorium against AI-written change descriptions (e.g. PR and commit messages, also issues/tickets) from my team. AI was writing change descriptions that were worse than useless to me as I tried to review PRs: outlining details of the code that could easily be seen by looking at the code, but omitting the higher-level framing needed to understand broadly what the code is doing. I think people like having AI write these things because the output looks structured and thorough, which makes it feel professional in a way. But this isn't actually valuable. Concise, high-level descriptions are better for everyone. If I need to use my own AI to interpret what your AI wrote then something is wrong. Let AI write code, sure, but for the description, I'd rather see your prompt than your output. We could maybe have extended agents.md with guidelines on writing descriptions, but this seemed a bit pointless since a good, concise change description only takes a few minutes to write -- not a significant time savings to delegate to AI. At least, it doesn't take long if you understand the code -- and if you don't understand the code, then I'm definitely not merging it.
English
background agents + worktrees are so powerful.
Slate manages my worktrees for me. all i need to do is talk to slate and it gives me updates on:
the progress of tasks, what decisions need to be made, and what's ready to review
it's non-blocking, so i can keep feeding tasks to slate
English
Also keeps assuming that subagents will have it's same full context?? Even though I specifically highlight in the prompt that they don't... Maybe this subagent flow is kinda out of it's distribution :(
It seem phenomenal at coding though.
English
Grok 4.5 seems kinda bad at communication. It's like I'm speaking to HR or something. Uses horrendous wording.
English
And also, it seems to have a strong habit of verifying anything constantly. I used to get annoyed at that, but now I think it's really useful even if it slightly burns more tokens.
English
I am so fond of DS pro 4. The fact that the model isn't strong is actually good, because it'll likely be confused by the same things you are. Unlike stronger models which don't seem to convey that at all.
Senior AI Micromanager@chrisczynski
Grok 4.5 seems kinda bad at communication. It's like I'm speaking to HR or something. Uses horrendous wording.
English
@kerckhove_ts x.com/mitchellh/stat…
maybe relevant?
Mitchell Hashimoto@mitchellh
Ghostty is optimized for long running human operation so the startup includes a bunch of preallocation and registration (dbus) for that idea. If you’re on Linux and want instant launch then read our systemd integration page. That’ll do same process single instance launch and it’s faster than pretty much anything.
English
AI compaction is an anti pattern. Everything should fit into the context window.
English
@DavidOndrej1 github.com/davidondrej/sk…
blank skill?/ :( it had a promising title
English
im releasing all my agent skills to the public
this is hundreds of hours of trial & error
every single global .agents skill i have
go grab it. it's free.
English
@mitsuhiko The orchestrator/coordinator workflow? Seems to be the thing happening right now, I'm seeing very good results with it.
English
We're doing another edition of State of Agentic Coding. Anything we should cover that we might not be thinking about?
English