A rules file changes nothing. Measured.
Same agent, same thirty tickets: 2.28 broken rules per ticket without a file, 2.27 with one kept by hand.
What each setup broke.
In the file. Broken anyway.
A file is read when the session starts. A rule matters at one line, many turns later.
At the line
Before each edit, the rules for the file being changed.
From your code
Proposed from your code, approved once. No file to maintain.
In every agent
The same rules in all five coding agents.
Even a perfect file stops at 64%.
Pasted word for word, the rules were applied exactly 7 times of 11. Served at the line: 11 of 11.
64%
of the rules applied exactly, perfect rules file
100%
with VibeDefend, the same tasks
Served is not enforced.
Once, a rule served thirteen times was still undone. So every change is also checked before the commit.
In detail
Questions about rules files.
Should I delete my CLAUDE.md or AGENTS.md?
No. Keep what helps the agent work, such as build commands and conventions. The rules that must hold are better served at the line: in the study, the file alone changed nothing measurable.
What does serving rules cost in tokens?
About 18% more tokens per ticket in the study. Against a perfect rules file the cost was the same, for 11 rules applied of 11 instead of 7.
How was it measured?
One agent (Claude Opus 5) worked the same thirty tickets with no tool, with a rules file and with VibeDefend; blind auditors graded every diff. The protocol, the transcripts and the audits are public.
Which agents does it work with?
The installer sets up Claude Code, Cursor, Windsurf, GitHub Copilot and OpenAI Codex. Rules reach all five; the guards run in Claude Code, Cursor and Codex.
Your rules, at the line.
One command. Every agent. No card.
npx -y @cybedefend/vibedefend@latest install
Claude Code
Cursor
Windsurf
GitHub Copilot
OpenAI Codex