I wrote 5 drift checks for my workspace, ran them on myself, and failed two. One automation had been dead for six weeks.

In my defense some of the fails were from migrating from a windows machine to Mac 😎

Paste this into whatever agent can see your workspace (Claude Code, Codex, Cursor, whatever you run). Read-only audit. Rules for the agent: fix nothing, move nothing, report what you find with file paths.

Run these five checks against my machine and score me.

**1. An automation still names a fact I've since changed.**

- Open every scheduled task and skill and read the actual prompt text. Claude: ~/.claude (skills + scheduled tasks). Codex: ~/.codex.

- Check every email address, price, file name, path, date, and account those prompts mention against what's true today.

- Stale facts never throw errors. They just keep running.

**2. The same instruction lives in more than one file.**

- Search the workspace for my email address, then my business name, then my core style rules. List every file stating the same fact.

- More than one copy means the next update forks the brain. One copy gets the change, the rest keep working from the old truth.

**3. My instruction files hold facts instead of pointers.**

- Read the root instruction files (AGENTS.md / CLAUDE.md) top to bottom. Flag every line that states a fact instead of pointing at the file where the fact lives.

- Report the total size of the instruction chain. ChatGPT caps the combined chain at 32KB and silently drops your deepest, most specific files first. Claude reads it all and listens worse as it bloats. Different failure, same cure: thin routers.

**4. An automation points at something that no longer exists.**

- Follow every path, URL, and address each automation references and confirm the target is still real.

- Then check the reverse: is every automation file actually registered and scheduled, or is it an orphan that looks alive and never runs?

**5. There's no written map of what lives outside the workspace.**

- List everything the agent reads that sits outside the folder: global instruction files, account memory, skills, scheduled tasks. Claude: ~/.claude. Codex: ~/.codex plus account memory.

- Check whether any file inside the workspace inventories all of it. If I can't list what's out there without looking, I don't know what my agent is being told.

Score me out of 5 and rank the findings by which one bites first.

0 comments

I wrote 5 drift checks for my workspace, ran them on myself, and failed two. One automation had been dead for six weeks.