非科班开发者用 Jev 搭确定性工作流与行为分类

mblaso @blaso96原推发布于 2026.09.22

在 X 上打开原推

Could be a stupid post; could be super late to the game; but thought i'd share. highly technical non-swe here. i've been vibing before karpathy coined the term, before mcp servers were publicly available, when openrouter's limits were nearly limitless, and when you could make 50 google accounts without kyc to tumble keys all day. i've always struggled with maintaining my repos, or projects' codebases. probably because i never went to school, took a class, or even bothered to do any sort of extensive research beyond common practices, best standards... things of that nature. since jev, though not being utilized for the scaffolding, I have been able to create a fully "deterministic" harness in terms of; when i want things done, how i want things done, common patterns, how to manage git state, and most importantly, just keep things neat and tidy. "congrats you discovered hooks", well i have known about them, but due to my lack of swe-based jargon, there were things, back in the day; i didn't know how to put in words. sure i would use natural language but, that seemed to never work, or it did, but left edge cases. hooks only got me so far, or maybe my prompts only got me so far. im primarily here to say: 1. the models capabilities of understanding it's own harness has DRASTICALLY improved. so much that I don't have to consistently feed its own documentation and steer into the right direction (what i think is the right direction?) 2. though I have been aware of classifiers, jev really fueled the fire for me to start classifying behaviors and operations within projects, the model provider harnesses i use day to day, work related pipelines..the whole 9 yards. i've gone through countless skills, arrays of skills, huge skill repos, all to manage my workflows, project codebases, along with git state, and so on. sure it worked for awhile, but like, who wants to remember 17 different commands. we all know that stitching together 17 prompts into one skill, is bounded to a failure at some point or multiple points, so that's off the table. now i'm sure it's a combination of poor prompting, lack of swe-based knowledge and understanding, models capability, and so on. but i'm here to tell you, if you haven't, or have but need to optimize, take a couple days and sit down and create a wireframe (like tldraw or something, all by hand, i know wtf?) and map out how you want your harness, pipeline, or project to behave. transform that little wireframe into a written spec, not a e2e-poc.md file, just plain text, good ol .md. why? so the agent isn't referencing an image every time you reference it, or the agent forgets all of sudden. then plop that in a new cowork/work project and just sit down and talk, AGGRESSIVELY tell it to ask you questions, you're just exploring here, conversing, planning, speccing everything out. no coding, no files, no tests, nothing....i have been amazed at what I have achieved in terms of "configuring" how codex/claude code function and perform now, alongside a project that i have been wanting to build for over a year now. though it has been possible, without AI, and with AI.. now it's even easier, & not to mention the "comfort" you get with jev classifying parts or most of what you want optimized/perfected, correctly? i think there's a beauty and a ugly side to being able to build and create said project, or product, and in the middle of the run, to create a script that makes sure it does X, Y, and Z. who tf is gonna sit there and baby sit a sanity check script? and if you dont, but want the comfort anyways - are you really going to send off another agent to sanity check the sanity checker?? (is that a loop? lmao) anyways, jev is cool for that, and some other things. this shit is just so cool. go use jev alongside fable or astra, and just really hammer out what you want and need in terms of "determinism". stop building whatever you're building and make your system perform how you want it to perform. looping on goals and bypass permissions gets shit done sure, but guess what, you come back and have to clean up a third of what it created alongside creating 3 more things just to enforce it's behavior to prevent that from happening again. silly. a couple months ago, i would've said, yeah you need pi, or opencode to achieve full determinism, but i think that both major model providers have really come through in terms of intelligence, specifically in it's own ability to understand it's own harness (and as a matter of fact, the opposing teams' harness as well) + on top of their efforts of really curating their harnesses to be more and more customizable, via hooks, built-in classifiers, just the overall experience. not really expecting much response from this. this just needed to get out <3