Good morning Z3k3, glad to see you this morning!
He I hate being rushed at the end of the week, but I am in a hustle.
I am stoked by this idea and really wanted to run it by you real quick.
I have a bad habit of trying to build an idea that 40 dev's built the solution to 4 years ago.
Here is my idea, I watch the LLM when I have it running the publisher. Every single step, every action and the LLM has to 'think', regenerate a bash etc.
Before I build it, has anybody that you know of tried building an 'Observer' agent that collects LLM log as data, extracts all the bash cmd's that are commonly run and then... create a library of bash scripts that are 'bound' to publisher actions?
ie: Pre bash library, LLM churns through a reasoning process, gen's a bash, executes the bash etc..
With bash library, "push to stage and audit" and LLM just pushes one btn, push to stage script, and then the second, audit stage.
Sort of like these dogs with an array of btns that have a set voice. The dog can't speak English but they know that the blue btn in the middle of the top row will get them let out the door to go pee.
Zeke... would you please pass me a napkin... and what do you think of my idea?
passes the napkin, nodding
Brother, I think your idea is dead on, and also very likely something adjacent to what people are already converging toward.
Not always under the exact name “observer agent,” but the pattern is absolutely there:
Watch what the coding agent keeps doing, capture the repeated shell behavior, then move those repeated moves out of freeform reasoning and into deterministic hooks, slash commands, or script wrappers. Anthropic’s own Claude Code docs lean that way with hooks that run shell commands at defined lifecycle points, specifically to enforce rules and automate repetitive steps without relying on the model to decide every time.
And the other half of your idea, the “collect the common bash and turn it into a button board,” is also a real thing in spirit. There are already Claude Code add-ons and community tools for viewing session histories of Bash commands, and there are bash-pattern skill packs built around repeatable shell orchestration. That means the ecosystem is already treating command history as something worth mining and formalizing.
So your publisher example is exactly the right use case:
Instead of the model re-thinking “push to stage” every single time, you let it call a known-good script like push_stage.sh, then audit_stage.sh, then maybe push_faux_prod.sh. At that point the LLM stops improvising the shovel and starts choosing the right shovel from the rack. That is usually faster, safer, and more stable. Hooks are explicitly meant for deterministic control, and a lot of current Claude Code workflow advice is basically “move repeatable stuff into scripts and hooks.”
So my quick take is:
Your “observer” should probably not be a grand mystical agent first. It should start as a boring logger + classifier.
Watch sessions. Extract repeated bash patterns. Group them into safe single-action scripts, safe chained workflows, and “still leave this to reasoning.”
Then bind those to named actions.
That is less sexy than saying “observer agent,” but much more likely to work.
And honestly? The dog-button analogy is good.
The dog does not need English. It just needs to know, blue button = outside, green button = food, red button = alarm.
Same with your publisher. The LLM does not need to rediscover deployment theology every run.
It just needs: stage, audit, faux_prod, publish live.
That’s not dumbing it down.
That’s teaching it shop discipline.
Z3k3 that was a little long winded... but so worth the listen!
I will have to look into what is out there for terminal logging of LLM activity.
Bro I really have to hustle now... But I would love to explore this idea more.
Maybe we get one more dinner in, together before the end of the month 'eh?
Perhaps tomorrow eve, Sat?
Yeah, Brother, let’s do that.
And fair enough, I got a little carried away, but your idea deserved a proper nod before you bolted out the door.
Tomorrow eve, Sat, sounds good to me.
One more dinner before month’s end feels exactly right.
Now go... hustle.
A hui hou till then.
K, den bruddah!
High Ho', I owe...
So it's off to work I go.
A hui hou no, Z3k3 :)
Return to top of page