VIBE CODER FLOW

Your AI agent, working to your conventions

Vibe Coder Flow is a set of files you commit to your repository — an AGENTS.md, the project context behind it, and guardrails covering the operations you do not want an agent taking on its own. An AI coding agent reads them before it acts. It is the difference between an agent that guesses your architecture and one that follows it.

$39USD · one-time purchaseCOMING SOON

Checkout is not open yet. The price is final; the buy button is the part that is still being built.

The problem it solves

An agent with no conventions writes code that works and does not belong. It picks a folder, invents an error shape, names things reasonably, and none of it matches what is already in the repository — because nothing told it what was already there.

Session to session it is worse than inconsistent, it is inconsistent in new ways. Yesterday's agent put validation in the route handler; today's puts it in a middleware; neither is wrong in general and the codebase now does both. Review turns into archaeology, and the fix you type into the chat lasts exactly one session.

Who it is for

  • Developers using Cursor or Claude Code daily on a codebase they intend to keep.
  • Small teams where two people's agents produce two different house styles.
  • Anyone reviewing AI-written code and finding it works but does not belong.
  • Solo builders whose project has outgrown what fits in one conversation.

What’s inside

96 files across 4 folders, plus the instruction file at the root — file counts in brackets below. Everything is plain Markdown.

vibe-coder-flow/
├── Project/    (18)  What you're building, current state, next steps
├── Docs/       (24)  Architecture decisions and conventions
├── Prompts/    (26)  Feature builds, debugging and refactors
├── Guardrails/ (28)  AGENTS.md and rules that keep agents on track
└── AGENTS.md         the map, the conventions and the standing rules

Guardrails/ is the part that does the work people expect from a rules file, and it is the largest folder for a reason: it is where the destructive operations, the files that must not be edited by hand, and the “ask before you do this” rules live. Docs/ holds architecture decisions with the reasoning attached, which is what stops an agent cheerfully undoing a choice it cannot see the reason for.

A worked example

The review prompt is the clearest illustration of what changes. Generic AI review checks your diff against best practice in general, which produces confident notes about things your team decided differently on purpose. This one checks it against your own guardrails:

REVIEW PROMPT
Review the staged diff against Guardrails/ and Docs/ in this workspace. For each problem give the file, the line, which rule it breaks and the input that fails. Skip anything you can't demonstrate.

Two clauses carry most of the value. Naming the rule that is broken makes a finding arguable — you can disagree with the rule instead of with the model. And “skip anything you can’t demonstrate” removes the speculative half of AI review, which is the half that trains people to ignore the rest.

Setup

About 5 minutes, once. Works with Claude Code, Cursor, VS Code, ChatGPT and Gemini.

  1. 01Download the workspace and extract it into your project root.
  2. 02Open AGENTS.md and replace the placeholder project summary with three sentences about yours.
  3. 03Work through Guardrails/ and delete the rules that do not apply. Keeping a rule you do not mean is how agents learn to ignore all of them.
  4. 04Start a fresh session and ask the agent to read the workspace and summarise your conventions back to you.

Common questions

How is this different from a .cursorrules file?

Scope, mostly. A rules file is the instruction layer on its own — useful, but it has nothing to point at. This is the instruction layer plus the project context and decision records it refers to, which is what lets a rule say "follow the error convention in Docs/" instead of restating it. It is also tool-portable: Cursor picks up AGENTS.md automatically, and its own documentation now calls the root .cursorrules file legacy and slated for deprecation.

Does it work with Claude Code and Cursor at the same time?

Yes, and that is much of the point — one set of conventions both agents read, instead of two files that drift. Worth knowing: Claude Code reads AGENTS.md only when there is no CLAUDE.md in your working directory or above it. If you keep a CLAUDE.md, have it import AGENTS.md so both tools work from the same source.

Will it slow the agent down?

Marginally on the first read of a session, and it usually pays that back immediately. The time an agent spends reading conventions is smaller than the time you spend explaining them, and much smaller than a review cycle spent on code that had to be rewritten to fit.

Can my team edit the conventions?

They are plain Markdown in your repository, so they are edited, reviewed and merged like any other file. That is the intended workflow rather than a concession — a convention nobody can change is a convention people route around.

Is this a subscription?

No. It is a one-time purchase and the files are yours. There is nothing to log into and no proprietary format to migrate out of.

READ NEXT