Intranets Today

The Best AI for Vibe Coding (From My Desk, Tea in Hand)

Hi, I’m Kayla. I write code late at night with lofi on, cat asleep, and a mug of mint tea. I want tools that feel like a chill pair-programmer. Not a drill sergeant. So I’ve tried a bunch of AI tools for what I call “vibe coding.” You know—flow, small nudges, fewer tabs, more focus.

Here’s what actually worked for me, with real stuff I built.

For a quick reference, I’ve parked my running cheatsheet of these tools and prompt tricks on Intranets Today. If you’re just after the shortlist of what’s on my daily roster, I keep an up-to-date roundup of the vibe-coding tools I actually rely on right here.

My Setup, So You Know Where I’m Coming From

  • Machines: ThinkPad X1 (Linux), MacBook Air (M2)
  • Editors: Cursor, VS Code, and a little Neovim
  • Stacks I use a lot: Next.js/React, Python (FastAPI), Node, a tiny bit of Go
  • Work mix: side projects, small client apps, and weird little UI toys

I care about speed, context, and tone. If the AI sounds like a robot manager, I’m out.

Cursor: The Best “Flow State” Editor

If you want smooth, this is it. Cursor feels like VS Code with a brain and good manners. The inline “edit this part” tools are the star.

Real example:

  • I built a small Next.js weather widget in 40 minutes. I pasted my rough HTML with Tailwind classes and said, “Make this mobile-first, keep the glassy look, reuse my colors.” Cursor rewrote the layout, fixed my janky spacing, and added a neat loading shimmer. I clicked “Accept,” then “Oops, too bright—tone it down.” It adjusted again. No tab dance.

What I love:

  • Context is sticky. It “remembers” the file vibe.
  • Multi-file edits feel safe. It shows diffs I can scan fast.
  • Good at Tailwind, React, and small refactors.

Where it trips:

  • It will sometimes loop on long changes and edit the same block twice.
  • If you’re offline, you feel it.
  • It can be too confident on libs it “thinks” I use.

Vibe score: 9/10. It stays out of my way and keeps me in flow.

GitHub Copilot Chat: Great Autocomplete, Solid Fixes

Copilot feels like a quiet teammate who finishes your sentences and explains errors if you ask.

Real example:

  • I wrote a tiny Go CLI to call the Stripe API. Copilot filled flags, error wraps, and even the basic request code. When I hit a 403, I pasted the log into Chat. It said, “You used the test key; the route needs live or a mock.” Facepalm. Fixed in two minutes.

What I love:

  • The inline ghost text is fast and usually right for small stuff.
  • Good at test stubs and boring glue code.
  • Chat is helpful for error messages and quick docs reads.

Limits:

  • Can be generic on tricky logic.
  • Rate limits hit at odd times for me.
  • Sometimes it repeats advice I already tried.

Vibe score: 8/10. Autocomplete king. Less artsy, more practical.

Windsurf (by Codeium): Big Plans, Big Changes

Windsurf is bold. It proposes a plan, then touches many files at once. Great when I want structure. (If you want another perspective on a similarly “all-in” editor, I spent a weekend with Bolt and wrote a hands-on review over here.)

Real example:

  • I had a Bash script that cleaned CSVs. I asked, “Turn this into a Node tool with a config file and tests.” Windsurf made a plan, built a CLI with yargs, set up Jest, and added a sample config. I did one pass and shipped it the same night.

What I love:

  • Project-wide awareness. It treats the repo like a whole story.
  • The plan view calms me. I see the path before changes land.

Limits:

  • On my MacBook Air, fans spin. It can feel heavy.
  • It sometimes over-engineers. Not every script needs a factory pattern.
  • On Windows, I had some path weirdness with test runs.

Vibe score: 7.5/10. Best when I want “adult supervision” on a messy repo.

Claude 3.5 Sonnet: Calm Brain for Tough Stuff

I use Claude in the browser and with Aider in the terminal for diff-based edits. It reads long files like a champ and keeps a gentle tone I like. (I kept a full week-in-the-trenches diary with Claude that you can read here.)

Real examples:

  • I pasted a 400-line React component and said, “Make it smaller and keep button logic the same.” It split it into three clean parts and added nice comments.
  • I also asked for a spicy headline style for a landing page. It kept my brand voice, but less cringe. That balance is hard.

What I love:

  • Long context, clear reasoning, readable refactors.
  • Great at naming and error phrasing. Like a kind editor.

Limits:

  • Slower on giant codebase chats.
  • It won’t run my code, so I still test everything.

Vibe score: 8.5/10. My “thinky” buddy. Pairs well with Cursor.

Replit Ghostwriter: Fast Browser Builds

When I want to make a toy and share it fast, Ghostwriter is simple. (My deeper hands-on take is here.)

Real example:

  • I built a little CSS glitch art page for my portfolio. I said, “Give me a neon glitch hero with click to distort.” In five minutes, I had a live link. I tweaked colors in the same tab. Sent it to a friend right away.

What I love:

  • No setup, quick deploy, fun for experiments.

Limits:

  • You live in Replit’s world. Feels tight for bigger work.
  • Not my pick for private client code.

Vibe score: 7/10. Great for playful bursts.

Continue + Local Models: Quiet, Private, Good Enough

Continue.dev in VS Code lets me point at local or hosted models. I use a lightweight Llama model when I’m on a train with spotty Wi-Fi. (I also tried Base44 for a week—warts and all—and wrote up that adventure here.)

Real example:

  • I set the style to “friendly, short answers, code-first.” I asked for a Jest test for a file upload helper. It wrote a decent test. Not genius, but good enough—and offline.

What I love:

  • Privacy and control. It feels calm.
  • You can swap models by mood or hardware.

Limits:

  • Local models miss hard logic.
  • You need to tune prompts more.

Vibe score: 6.5/10. I like it for quiet days and travel.

A Quick Day in the Life: Three Wins, One Mess

  • Morning: Cursor helped me fix a flaky Jest test with a race condition. I typed, “Make this wait for the mock server to settle.” It rewrote the await logic. Test went green.
  • Lunch break: Claude rewrote a gnarly regex into two small helpers with names I’m not ashamed of. I could breathe again.
  • Evening: Copilot autocomplete filled 80% of a FastAPI endpoint and the pydantic model. I wrote the edge case by hand. Flow felt smooth.
  • And the mess: Windsurf made a plan that added a layer I did not need. I rolled it back and asked for “simpler, keep files flat.” The second plan was perfect. So yeah, be firm with it.

If you’re curious what happens when you hand the steering wheel to an AI entirely—as a so-called “AI CEO” for a week—I chronicled that chaotic experiment here.

Settings and Prompts That Keep the Vibe

  • In Cursor: I keep suggestions short and turn on “preview diffs.” I often say, “Small change, no new deps.” That phrase saves me.
  • With Claude: I set the tone—“Be concise, keep naming plain, explain once.” It listens.
  • With Copilot: I accept small chunks, then test. Nudge, test, repeat. That rhythm works.

Little prompt I use a lot:

  • “Keep style, reduce lines by 20%, no behavior change.”
  • “Explain the bug in one sentence, then show the patch.”

Before I wrap, a quick side note on energy: marathon coding sessions can tank your focus if your body’s running on fumes. I’ve been poking around resources on how hormones play into sustained mental performance, and one rabbit hole led me to this thorough [

zielkeguptill