The VibeDrift blog.
Field notes and guides on drift in AI-written code — what it is, how to measure it, and how to keep your codebase coherent as agents write more of it.
AI writes code faster than ever, and a new kind of technical debt piles up just as fast: drift. What AI technical debt is, why it's invisible, and how to measure and pay it down.
Claude Code is brilliant in a fresh repo and drifts in a big one, because every session is stateless. The practices that help, the ones that don't, and the MCP fix that makes it check before it writes.
An MCP server lets an AI coding agent call tools while it works. What the Model Context Protocol is, how a server works, how it differs from a plugin or API, and how to add one to Claude Code or Cursor.
AI code compiles, passes tests, passes the linter, and passes review, then quietly breaks your codebase anyway. Why review misses behavioral drift, and what actually catches it.
Cursor rules nudge the model up front but can't guarantee it follows them. MCP checks the real codebase mid-task. Why config files aren't enough, and why you want both.
How to write a CLAUDE.md your agent actually follows, and the honest part nobody mentions: where it quietly fails, and what to pair it with so your conventions don't erode.
Five MCP tools that let your AI agent ask the codebase how it already works before it writes, so it stops contradicting your conventions and reinventing code that already exists.
A practical playbook for keeping a large codebase consistent when AI writes most of it: declare conventions, check them in-loop, measure drift, and gate CI so regressions can't merge.
Make coherence a merge requirement, not a hope. How to add a VibeDrift drift gate to GitHub Actions, pick a --fail-on-score threshold, and ratchet it up over time.
What code coherence means, and how the Vibe Drift Score measures whether your codebase agrees with itself. How to read the 0–100 score, the five detectors, and your scan-over-scan delta.
The full controlled eval: with and without VibeDrift, graded by an independent pass. The signal cut introduced drift where the agent couldn't infer the convention, and did nothing where it could.
AI moved the bottleneck from writing code to keeping a codebase coherent. VibeDrift now runs as an MCP server so your agent checks the repo's conventions before it writes, and we measured the impact.
AI coding tools don't just introduce bugs, they introduce behavioral deviation. Your codebase slowly stops agreeing with itself, and no existing tool measures it.