The instruction your agent follows is production code. It decides what ships, what gets refunded, what gets escalated. Yet most of these instructions live as a paste in a chat window, copied between three tools, with no history and no owner.
When a prompt is just a paste, you cannot answer the questions you would never accept being unanswerable about code: Which version is running? Who changed it, and why? Can I roll back the change that started the weird behavior on Tuesday?
Treat instructions as artifacts
The fix is not exotic — it is the discipline you already apply to code:
- History — every change is a version, with a diff and a reason.
- Provenance — you can trace the running instruction back to the review that approved it.
- One source of truth — the approved version syncs into each project instead of drifting across copies.
How it fits your workflow
Improve an instruction where you already work — in your editor, over MCP — then sync the approved version into any repo with one command. The same artifact lands in Claude, Codex, or Cursor, and the version history travels with it.
The payoff is boring in the best way: when something changes, you can see what and revert it. An instruction you cannot version is an instruction you cannot trust to hold.