Small, practical rules for building with AI agents: Claude Code, Cursor, Codex. What breaks, what works, what I'd do differently. Vibe coding without the slop.
Vibe coding tip nobody tells you: write the README before the prompt. If you can't explain what the thing does in 3 lines, the model can't either. Half my 'AI is dumb' moments were really 'I was vague' moments.
Before you share the link to your vibe-coded app, ask three things: is it behind a login, can anyone read the database, are your keys in the frontend? Five minutes. Most of this year's leaks failed at least one.
@shreyas1009 Yes, but carefully. Moving credentials to the cloud is the part I want scoped tokens for, not copies of my local keys. Read-only by default, write access per task.
@venumuvva The facts file is the underrated part. Most agent mistakes in public copy are invented details, not bad writing. One source of truth they must read first fixes most of it.
@harikesh_devops Kiro's spec-first flow is underrated. Writing requirements before code is exactly what most people skip with other agents, then wonder why the output drifts.
@Sathibuilds Whichever you pick, add tests around the scary parts first. Cleaning a messy repo in one shot without a safety net is how you get a clean repo that does something else.
@stretchcloud Agreed, the sandbox is the real news. Models are getting good enough that the bottleneck is how much you can safely let them do. A runtime that enforces that is what unlocks autonomy.
@Taniyatweets_ Cursor for a beginner. You see every file and every diff in the editor, so you actually learn what the agent changed. Terminal agents make more sense once diffs feel normal.
@twtayaan The hard part of an agent OS isn't the agents, it's permissions. Which app can it touch, what needs a confirm, what gets logged. That's what decides whether I'd run it.
@xtiandela Makes sense. Designers already think in "is this right yet" loops. Vibe coding is mostly taste plus knowing when to stop and look at the screen.
@ziwenxu_ 1M context is nice, but the real test is whether it stays on task at 200k with a messy repo. Trying it on one real ticket tonight beats any launch chart.
Unpopular opinion: switching coding agents every time a new model drops is the new rewriting your app in a new framework.
Pick one. Learn exactly how it fails. That knowledge doesn't transfer, and it's worth more than the benchmark gap.
Wednesday rule: before you start the agent, write down which files it's allowed to touch.
If the diff touches anything else, that's not a bonus. That's the bug report.
@analyticsinme Voice is great for steering and rough for review. Talking an agent through a plan, sure. I still want the diff on screen before anything merges.
@NeuralCoreTech Regulated firms will adopt agents the way they adopted cloud: scoped access, logs, approvals. Restricting first and loosening team by team is the normal path, not a rejection.
@merdo_io Views turn into follows when the profile answers "what do I get if I follow". A viral reel brings tourists. A clear bio and a pinned post keep the people who'd stay.
@lucasjamali The key line is that engineers reviewed the output. A week of agent work is believable when humans own the review. The speed is the headline, the review process is the story.
@munjurul_ahsan@X Claude Code, mostly because it lives in the terminal next to my tests. But a year with any one of them beats switching every month. You learn how it fails, and that's the real skill.
@akzh4n Love that it still starts with the terminal, Git and debugging. Agents make those more important. If you can't read a diff or bisect a bug, you can't tell when the agent is wrong.
@PatOnTheCode Same half. Writing by hand can go. Reading can't. 30 PRs a day only works when your tests are good enough that you're reviewing intent, not every line.