Our prompt sprawl is out of control. We've got prompts in Slack threads, Google Docs, and random `.txt` files. No history, no rollback, no consistency. It's breaking our Claude.ai workflows.
Treat prompts like code. Here's our setup:
* **One Git repo** for all prompts. Structure by project/use-case.
* **Store prompts in plain text files** (`.md` or `.txt`). Use descriptive names.
* **Template with variables** to avoid duplication. Example:
```text
You are a code reviewer for {{language}}. Focus on:
- Security flaws ({{security_checks}})
- Logic errors
- Suggest optimizations
```
* **Commit messages** that explain the *why* behind prompt changes.
* **CI pipeline** to test key prompts if possible (e.g., validate output structure with a script).
This gives us version history, peer review via PRs, and a single source of truth.
—cp
—cp