Your prompts drive real product behavior. Why shouldn't they get the same treatment as your code? Prompt management as code means storing prompts in Git, reviewing changes in pull requests, and deploying by push - with prompt git as the source of truth.
Why prompts should live in Git
Editing a prompt in a web dashboard and hitting save gives you no history, no review, no rollback. Storing prompts in prompt repositories gives you everything you already rely on for code:
- Versioning - every save is backed by a git commit.
- Review - peer review prompt changes before they go live.
- Rollback - revert any release in one click.
- Audit - exactly who changed what, and why.
Prompt Markup Language (PML)
In Infere, prompts are authored in PML - a file format with YAML frontmatter, tagged sections, and mustache variables. It reads like a config file and validates on every push:
---
name: customer_support_v2
model: auto (balanced)
---
you are a helpful support agent.
Reply using the user's language: {{language}}
context:
customer_tier: {{tier}}The flow: clone, push, deploy
Prompt git follows the workflow you already know:
- Clone the repository over HTTPS or SSH (e.g.
git clone https://KEY:[email protected]/git/acme/acme-prompts). - Edit the PML file in your IDE.
- Push to a branch and open a pull request.
- Review, approve, and merge - evaluation gates run automatically.
- Merge to
mainand the edge updates in seconds.
Branch protection, required peer reviews, and auto-deploy-on-merge give you the same guardrails as a well-run codebase.
Repositories and sharing
Prompt repositories can be scoped per organization and shared across workspaces. One team's production prompts live in one repo; shared repositories let you reuse evaluated prompts across teams without losing centralized governance.
The same discipline you apply to code - review, version, deploy - applied to the logic your AI runs on.
Next steps
Infere is an AI prompt management platform built on prompt git: repositories, branch protection, pull requests, and deploy-by-push, with evaluators stored alongside your prompts. Treat your prompts like code and you'll stop shipping prompt changes blind.
Prompt changes directly affect token consumption and cost - our true cost of agentic coding breakdown shows how output length and model choice drive the bill.