Vibe Coding - Best IDE, Model, and Workflow

I keep making posts & podcast episodes to keep up with the latest recommended IDE & models, advanced tools & workflows, for vibe coding. So I'm creating this as a living doc which I'll update any time something major changes. Check back every month or two if you're curious.

TLDR:

• Tool: Cursor for isolated tasks; Roo + Orchestrator for complex tasks.
• Model: ChatGPT o3 (website) for designing a game plan; gemini-2.5-pro-exp-03-25 (free) in Roo until you rate limit for the day; gemini-2.5-pro-preview in Cursor / Roo per above afterwards.
• Jupyter: Jetbrains now supports Jupyter AI, and their tooling is better than web Jupyter.
• Context7 MCP for updated library / frameworks documentation

Long version:

The free Gemini setup got complicated since my episodes (1, 2, 3). They have aggressive rate limiting, so you need to use Preview after you hit that today's limits. Which is still the best value-per-token, so it's worth it. But I'm back to Cursor to control costs; and it gets complicated if you want to tightly control costs (which is the difference between $5/day vs $100/day). So here's the daily play by play.

Are you rich? If you don't care about costs (maybe your company pays), just use Roo + Orchestrator, o3 as Architect, gemini-2.5-preview for everything else. Stop here.

1. Cursor Pro + Roo

Make Cursor your IDE, install Roo inside it. If you're on Ubuntu, use cursor-setup-wizard. Sign up for Cursor Pro ($20/m).

• Go to Cursor settings, enable gemini-2.5-pro-* models and o3, disable everything else. Stop there if you trust me, otherwise study Aider Leaderboards and experiment. Consider claude-3.7-*, o4-*, 4.1 - experiment at your leisure. Disable everything else.
• Roo settings:
  • Setup Gemini token via Google AI Studio (not Open Router), only way to get gemini-2.5-exp-03-25. "Get API Key" -> Setup Billing -> Upgrade
  • Roo: Setup two profiles: one for gemini-pro-2.5-exp-03-25 and one for gemini-pro-2.5-preview. For both, check-box "Enable prompt caching" (saves money) and "Always read entire file" (utilizes Gemini's 1M context window for better intelligence).

2. Simple vs Complex Tasks

1. For inline edits, use Cursor's default functionality.
2. For simple Agentic tasks: updates to a single file without bigger-picture architectural considerations
   1. Use Cursor's gemini-2.5-pro (part of your $20/m plan) if the chunk of code only needs ~500 LoC consideration. This may be rare in practice, but it's your cost-savings maximizer - so try to wrap your head around the scenario.
   2. Use Cursor's gemini-2.5-pro-MAX ($.05 per tool use, roughly $.20 per actual task) for single-file edits like above, but which need to consider the whole file.
3. For bigger-picture tasks
   1. If you need help designing a game plan, use o3 - either via Cursor's o3 tool ($.30 per tool call; roughly $2 per actual task); or if you have ChatGPT Pro do it manually in the website (cheaper). Never use o3 as Architect mode in Roo, unless you can afford it - it's pricey. What I mean by this bullet-point is: you're not sure how to start or the best approach, and you need some guidance.
   2. If instead you know what you want, or have a plan from (3)(1), use Orchestrator.
      1. Use gemini-2.5-pro-exp-03-25 up until you get 429 (meaning you've ran out of quota for today)
      2. Then use gemini-2.5-pro-preview for the rest of the day.

3. Orchestrator

If your task is complex, always use Orchestrator (Gemini everything). Don't just use it multi-agent tasks, use it for multi-prompt tasks. Two reasons:

1. It re-writes your prompt so Code Mode is better equipped (this goes far).
2. After each Code Mode completion, Orchestrator gets the summarized report, so that follow-ups (1) maintain context flow; (2) but summarized, so that you save token costs for follow-ups; and context doesn't get diluted.

So this not only adds multi-agent pipelines, but improves any multi-prompt conversations, and saves money. Orchestrator is the new official (comes with Roo), optimized version of the old Boomerang Mode - so you can delete your .roomodes Boomerang JSON if you manually installed it previously. It's their SPARC multi-agent flow mode.

I use Orchestrator to generate 80% of my code: "build feature x" or "fix bug y, and search for all files which may be involved". Then I use Cursor gemini-2.5-pro or gemini-2.5-pro-MAX to handle isolated bugs / enhancements. The former takes 20% of the time (generating 80% of the code); the latter takes 80% of the time (fixing 20% of the code), per the classic Pareto Principle.

Context7, Jetbrains, Jupyter

Context7 is an MCP which maintains updates to modern frameworks and libraries. I bemoaned lack of RAG for library versions newer than LLM's training cut-off in MLA 24. You can install the Context7 in both Cursor & Roo so it has up-to-date knowledge. Make sure to say explicitly "use context7" if you know you'll need it.

Jetbrains finally launched something legit: Junie! It's actually really damn good. It comes for free if you have the All Tools pack. If you use Jetbrains, and you know you'll use Junie - the All Tools pack is cheaper than one tool + Junie. Junie has one fatal flaw: the quota per month is extremely low, I ran out within a few days. So unless they figure that out, stick with the above system for most tasks. Also see Junie vs AI Assistant if you do use Jetbrains, and get confused.

Jupyter. Finally, we have a real Jupyter solution through Jetbrains Junie. Their latest release improves in-IDE Notebook support, and adds a ton of tooling on top. You can use AI to edit cells, which now has context knowledge for the surrounding Notebook, and (I think) the codebase at large. It adds powerful data tooling - SQL, Dataframes, etc - visualization & editing via Data Spell (a dedicated tool, but is built into PyCharm). I particularly love Data Wrangler. I highly recommend, if you're a data scientist or machine learning engineer, do not use Jupyter Labs from browser any more - get acquainted with the powerful tooling Jetbrains adds on top, it will help you maintain an edge.