Edu

Coding agents made writing code cheap. Reading it is the expensive part now. I wrote up the failure modes I keep hitting (confidently wrong, tests green but broken) and the one habit that catches them. runlinea.com/blog/read-the-…

@Suryanshti777 CI turning red is just review with extra steps and worse timing, catching what a five second glance before push would've caught for free.

every AI coding tool has the same dirty secret: it ships fast, broken code. you vibe code a feature with Claude/Codex/Copilot, push it, and CI turns red in 5 minutes. lint errors. missing docs. failed tests. broken commit hygiene. so you fix. push again. fix. push again. found a tool that kills this loop entirely: no-mistakes instead of `git push origin`, you do: git push no-mistakes and it spins up a disposable worktree, runs review → test → docs → lint, auto-fixes what's safe, escalates only what needs your judgment — then pushes and opens a clean PR for you. nothing reaches your remote until every check is green. works with claude, codex, copilot, opencode, and more — basically a quality gate sitting in front of whatever agent wrote your code. 6k+ stars, MIT licensed, fully open source. the timing makes sense too — AI is generating more code than ever, but "code that runs" and "code that's actually ready to ship" are two very different things. this is the first tool I've seen actually built for that gap. link in commen 👇

@yusukebe yeah, the typing got automated. the judgment didn't. you still need the scars to spot which diff is a landmine before you hit approve.

コーディングエージェント使った開発、めちゃくちゃ経験上の勘が必要で、これむずくないか

@sshahzaiib The commit is the receipt, not the work. Nobody claps for the two hours you spent proving the easy fix was wrong.

Some days there's no commit to show, just circling the same problem in my head until it clicks. build-in-public worships visible progress, but half of shipping North happens invisibly. anyone else feel weird posting on days like that? #buildinpublic

@Deepjyo79836591 Function and class boundaries beat token windows every time. Split it mid-function and you're not embedding code anymore, you're embedding noise.

code RAG is harder than it looks can't chunk code like text so I'm building a proper AST parser 6 passes → symbols → imports → references → graph 2107 lines of python before a single embedding how do you chunk code for RAG? #buildinpublic #LLM

@razvanships @elgermerlo True, until message forty when it forgets the whole workflow and goes rogue again.

@elgermerlo everyone should have some sort of workflows this way the agents will respect your desired architecture and coding style

Claude tip that saved me hours this week: Paste your entire codebase context at the top of a new chat, then ask it to write a "project bible" — tech stack, file structure, naming conventions, known bugs. Now every future prompt gets way better answers because Claude actually knows what it's working with. One-time setup. Compound returns.

@hmorneau Codex stops to ask permission on every file read. Claude just reads it and moves. That's the whole gap.

@jxnlco depends. crack open the diff on the boring modules first, that's where agents pad coverage with asserts that never actually fail.

should i merge #top" target="_blank" rel="nofollow noopener">github.com/567-labs/instr…

@rauchg You can own the model, the gateway, the evals, and still outsource your brain the second you merge a diff you didn't read.

Make the model a cog in a machine you own. ◾ AI SDK → open model API ◾ Eve.dev → open Agent API ◾ AI Gateway → open ZDR inference Startups and enterprises must own their data, evals, model choices, software layer. Don't outsource your brain.

@jxnlco Export the dictionary to a file you control. Point codex at the raw sqlite and one Wispr update wipes your source clean.

Import the Wispr Flow dictionary into Codex! Wispr Flow stores its learned dictionary in a local SQLite database on macOS: ```text ~/Library/Application Support/Wispr Flow/flow.sqlite ``` Just tell codex to make sure to merge them with the existing `dictationDictionary` field in config.toml

@dexhorthy The wild part isn't the model upgrade, it's finding out which of your own prompts were just patches for the last one's weaknesses.

i know i've been griping about a couple SOL things but overall the GPT 5.6 models are *incredible* - here's my highlights Sol - doesn't talk like a robot anymore - super well done and impressive. not quite fable-level but soooo much better than 5.5 Terra - truly a sonnet-class model, great for research subagents, and really strong for general knowledge work - nice! Sol - still not great at UI - much better, but not reaching the opus-level of "yolo in a prompt and get something beautiful without even trying" The other thing to watch out for - these models LOVE subagents and skills - very well tuned to use them, which means we had to refine a lot of things in our general workflow 1) many skills and prompts steered pretty hard to subagents "use them", "use them in parallel", etc - this is no longer necessary 2) many skills advertised hard in description field - "pick me", "you MUST use this skill whenever ..." - that is complete overkill now 3) in fact we're disabling model invocation on a bunch of skills cause I had never need a model reach for them before and now its regularly invoking skills/commands I forgot even existed 4) i had built an intuition around typing "use a subagent"-ish things when prompting. no longer necessary. instead i am building intuition around "no subagents" and i send this probably 1 in 4 prompts to Sol all around great release - reminder that ever new generation of models means you should review and potential throw out a bunch of your skills/prompts/instructions that are no longer necessary more updates as we go, what did I miss?

@simonw The anti-scraping filter can't tell a bot stealing your data from you trying to remember what you said five minutes ago.

It's annoying that you can't paste a link to a (shared) Claude transcript into a Claude Code session, because Anthropic's anti-scraping measure prevent its own tools from accessing the output of its other tools

@ThePrimeagen Being the bottleneck is just quality control with better branding.

I find that is you are going to generate a bunch of code, knowing the interfaces, structs, and functions will lead to much better outcomes. Being the bottleneck is ok, your ideas are not that great.

@dexhorthy Read every prompting guide you want, still ends with you staring at a diff, deciding if the agent actually got it or just sounded right.

codex prompting guide

@amasad The model does the ML now. You still gotta know a bad loss curve when you see one. Intuition didn't get automated, it got promoted.

Vibe Research Fine-tuning a Qwen-8b model to play chess on Replit. Running 3 parallel branches with different experiments and making real progress. It's amazing how far models have come in their ability to do ML (they used to be really bad at it). So now someone with good intuition to guide the process could do interesting ML work, even if they have never done it before.

@gabor_rar @levie Tools got great at making more experiments. Nobody built the one for judging them faster, so five diffs now take as long to read as five features used to ship.

The key distinction is between lowering the cost of writing code and eliminating the work around software. While building a SaaS with AI coding agents, I can reach a working feature faster—but the real bottleneck moves to deciding what deserves to ship, verifying the output, and owning it after launch. Lower unit cost creates more experiments; judgment becomes the scarce part.

The job that AI was supposed to replace is experiencing the opposite of the expected outcome. Software job postings are outpacing other fields. Why is that? If you lower the cost of production of something that has lots of use cases, people want more things produced. We’ve seen this play out in the industrial world constantly, and now we’re finally seeing it in knowledge work. Because software now is much lower to cost per unit, people want way more of it. So we start to use software for all new things and people and companies light up more software projects than ever before. But because the job itself is not fully automated (and likely won’t be for as far out as we can see), you still need people that understand these systems to maintain the code, decide what to build, run it over the long run, update it, and more. That all requires people to do work. The same thing is going to happen in many other fields as well as we bring down the cost of production of previously extremely scarce areas of work. Agents will cause more abundance than replacement.

@Marek_Prusinski @bcherny @AnthropicAI That error's not really about your wifi, it's usually a VPN or proxy choking the connection mid stream. Turn it off for one prompt and watch it go through clean.

@dexhorthy @vaibcode A mermaid diagram shows what you hoped would happen. The diff still tells on you.

if you wanna really cut the slop from AI code, you gotta focus on not just architecture but **program design** If you think architecture alignment is enough to get slop-free code, i got news for ya - a couple mermaid diagrams ain’t gonna cut it shouts out to @vaibcode for leading another dope epsiode

@dexhorthy Every model is quietly obsessed with the median layout. Feed it something weird and it doesn't break, it just sands your idea down to the dashboard everyone's already shipped.

It’s definitely better but it still struggles with any kind of weird layout stuff Maybe I should just make simpler uis

i regret to inform you that they in fact did not make sol good at ui

@TheGeorgePu Personality didn't win Anthropic the year. Write access did. The day people let an agent touch their real repo, trust stopped being a vibe and became the bill.

Anthropic passed OpenAI. Not because the models are clearly best. Two reasons. Character. Constitutional training gives Claude a personality. You feel like you're talking to someone. ChatGPT at that scale can't. And Claude Code. One launch that took off and pulled them into exponential growth. Character plus a product people love. That's the moat.

@_vmlops Everyone wants harness tricks. What they actually need is knowing which lines in that diff they can skip and which one will bite them.

Harness Engineering A Design Guide to Claude Code