SYS.BLOG
How to Vibe Code Large Projects (2026)
How to vibe code a large production project in 2026: what I tell the agents, the tools I give them, agentic code review, and why system design beats code reading.
In March 2025 I wrote up my thoughts on vibe coding and Cursor. It was full of Cursor tips: a .cursorrules template, Notepads for saved context, o1 for planning and Claude for the code. None of that is really useful anymore. Cursor removed Notepads in version 2.0, and the current docs don't even mention .cursorrules.
Since then I vibe coded a site to over a million visitors without reading a single line of its code. I left the old post up as a record. This is how I vibe code large projects now.
How to vibe code: it really is about vibes
Vibe coding comes down to three things: telling the agent what you want to do, communicating with it effectively about what you want, and giving it the tools it needs to do it. That's the whole job.
The tools part matters a lot. On Rifle Configurator the repository has 223 package scripts and 19 agent skills that are there for agents to run, plus my own MCP servers. An MCP server (Model Context Protocol) is the standard way to hand an AI assistant new tools.
The one piece of the old advice I still follow is the rules file. It just isn't .cursorrules anymore. It's CLAUDE.md for Claude Code and AGENTS.md for Codex, and Cursor now reads Project Rules from .cursor/rules as well as AGENTS.md. I have real CLAUDE.md examples if you want a starting point.
What I use to vibe code
Codex does the backend and the logic, and Claude Code does the content. My full Claude Code vs Codex vs Cursor comparison has the details.
I do still use Cursor, but not in the form I used to. Cursor now calls itself a coding agent, not an IDE. I have it run scheduled routines (Cursor calls them Automations), cloud jobs on its Cloud Agents, and the occasional task. Grok Bot uses it for coding too.
Then there are the always-on agents. Hermes has its own server and can open PRs while I'm on the go.
Context management mostly takes care of itself
I do try to keep the context tight, but you no longer have to manage it all by hand. I used to tell people to pick files carefully and start a new chat for every topic. Cursor, Claude Code, and Grok Bot now support long-running single agents. Claude Code compacts its own history when it gets near the limit, and it hands work to subagents that each get their own context window.
Codex has this now too, with subagent workflows. The main agent spawns subagents in parallel threads and pulls the results together. It's basically a chief of staff that manages the other coding agents' sessions for you, so you don't strictly have to do that manually anymore.
Planning workflows are for complicated work only
The planning workflow and the plan revision workflow are still fine if you're doing something complicated. I haven't had to use them in a while. The old post had a whole two-model routine, with a reasoning model for the plan and Claude for the code. For the long-running kind of task, I wrote up Codex /goal.
The tests break, the site doesn't: verification went agentic
Verification still stands, but now the agents do it. I used to write “never accept generated code without review” and “write comprehensive tests for all generated code.” We do have a large number of tests. They break randomly, but the site doesn't, so mostly they just slow me down.
What actually has value is the pre-commit and pre-push checks, and we have lots of them. On Rifle Configurator the pre-commit hook alone runs about 40 checks: internal links, structured data, state legality, critical affiliate routes, catalog duplicates, image formats. The pre-push step adds a builder compatibility gate on top.
Agentic code reviews have been around for a while. My rule is simple. If it's a complex change, it goes through the PR workflow and the review agents look at it. If it's a simple enough change, I ship it directly to master. I covered the setup in the PR review bots section of how I structure projects for AI agents.
Did things break this way? Occasionally. Rifle Configurator has had some outages, mainly affiliate tracking. But with Vercel's deploy system it has never had a code red.
Do I babysit the agents?
Occasionally. Oftentimes I just let them run. So many agents, so many worktrees, and more than 6,000 commits on one project.
The skill that matters now is system design, not code reading
Vibe coding is a system design test now, more than a read-the-code test. The old post ended by saying code reading was the skill that matters most. I don't read the code anymore, so I can't keep saying that.
What I do is keep a mental map. I'm a software engineer, and I understand my systems at a high level, so I can see quickly what needs to be done. It helps me prompt a lot better than a truly blind vibe coder would. It's about communicating, and grokking the system at a high level.
None of this is unlearnable if you're not a software engineer. You can simply ask the AI to explain the concepts to you. If you're starting from zero, my Cursor guide for non-technical founders is the on-ramp, and the Rifle Configurator case study shows where this approach ends up.
FREQUENTLY ASKED QUESTIONS
What is the most effective way to vibe code?
The most effective way to vibe code is to tell the agent what you want, communicate it clearly, and give the agent the tools it needs to do the work. A rules file such as CLAUDE.md or AGENTS.md and scripts the agent can run matter more than prompt templates.
What should I use to vibe code?
Aditya Bawankule uses Codex for backend and logic work and Claude Code for content. He also uses Cursor Cloud Agents and Automations for scheduled and cloud jobs, plus always-on agents such as Hermes Agent and Grok Bot.
What skills do you need for vibe coding?
Vibe coding needs system design skill and clear communication more than code reading. You keep a high-level mental map of how the system works so you can tell the agent what needs to be done, and you can ask the AI to explain any concept you don't know.
Is vibe coding still relevant?
Yes. Rifle Configurator, a site with over 1.2 million visitors in its first 11 months, was built entirely by AI coding agents, and its founder has not read its code. Verification now happens through agentic code reviews on pull requests.
