
Once you move past ad-hoc copy-paste prompts and start running AI tasks in production—content outlines, email subject-line variants, metadata generation—you need somewhere to keep the templates. Three common patterns emerge: file system, database, or hard-coded strings in your application code. Each trades off versioning ease, deployment friction, and team access differently.
File system: Git history as your version log
Storing prompts as plain-text files—.txt, .md, or .json—in a /prompts directory gives you Git history for free. Every edit, rollback, and branch lives in version control alongside your code. If a new template tanks your output quality, you diff the commit and revert.
This works well for solo operators and small teams who already live in a code editor. You can structure files by task (subject-line.txt, outline.txt) or by model (claude-rewrite.txt). Your application reads the file at runtime or caches it in memory. Deployment is a git pull and restart.
The downside: non-technical team members can’t edit prompts without opening a pull request or SSH-ing into a server. If you want a marketing person to tweak tone without touching code, file storage adds friction.
Database: runtime editing and A/B tests
Storing prompts in a database table—id, name, version, body, created_at—lets you update templates through an admin UI without redeploying. You can run A/B tests by flagging two versions active and routing requests randomly, then compare output quality or conversion rates.
Versioning requires deliberate schema design. A simple approach: never update rows; instead insert a new row with an incremented version integer and a is_active boolean. Your application queries WHERE name = ? AND is_active = TRUE. To roll back, toggle the boolean on the previous version.
Database storage fits teams where non-engineers need prompt access, or where you iterate fast and deploy infrequently. The trade-off: you lose Git’s diff and blame tools unless you export snapshots to version control periodically. If your database backup schedule lags, you risk losing prompt history.
Hard-coded: when simplicity beats flexibility
Embedding prompts directly in code as string constants—const OUTLINE_PROMPT = "..."—removes external dependencies. No file I/O, no database query. The prompt is compiled into your binary or bundled into your script. If the task is stable and the template hasn’t changed in months, hard-coding reduces moving parts.
This pattern works for narrowly scoped tools: a WordPress plugin that generates alt text, a CLI that summarizes RSS feeds. The moment you need variants or frequent iteration, hard-coded strings become a deployment bottleneck. Every tweak requires a code change, test, and push.
One non-obvious use: hard-code your system prompts (the instructions that set model behavior globally) and store task-specific user prompts externally. System prompts rarely change; user prompts evolve with your content strategy. Splitting them keeps stable config in code and flexible templates in files or database.
Hybrid: configuration meets content
Many production setups combine approaches. Store system prompts and model parameters (temperature, max tokens) in environment variables or a config file; keep user-facing templates in the database with an admin UI. Version-control a nightly export of active prompts as a backup and audit trail.
If you use Claude for multiple tasks—say, newsletter intro rewrites and social-post generation—namespace your prompts by feature (newsletter_intro_v3, social_post_v2) and log which version ran for each job. When output quality shifts, you can trace it to a prompt change rather than guessing.
For solo operators just starting: begin with files in Git. Upgrade to database storage when a second person needs editing access or when you’re testing three variants of the same template in parallel. Reserve hard-coding for the smallest, most stable utilities.
Got a question about tooling, infrastructure, or online-business workflows? Reply to this email—we answer reader questions every Sunday.
Heads up — some links in this article are affiliate links. If you sign up through them, we may earn a small commission at no extra cost to you. We only recommend tools we use ourselves.
