| license | Apache-2.0 |
| name | prompt-template-manager |
| description | Version-control, parameterize, and A/B test LLM prompt templates with Git-native workflows. Activate on: prompt versioning, prompt templates, A/B test prompts, manage prompts, prompt registry. NOT for: writing prompts from scratch (prompt-engineer), fine-tuning data (fine-tuning-dataset-curator). |
| allowed-tools | Read,Write,Edit,Bash(python:*,pip:*,npm:*,npx:*) |
| category | AI & Machine Learning |
| tags | ["prompt-management","versioning","a-b-testing","templates","llm-ops"] |
| pairs-with | [{"skill":"prompt-engineer","reason":"Prompt engineer writes content; this skill manages lifecycle"},{"skill":"llm-evaluation-harness","reason":"A/B tests need automated evaluation to pick winners"},{"skill":"llm-cost-optimizer","reason":"Template compression and model routing reduce per-prompt cost"}] |
Prompt Template Manager
Version-control, parameterize, and A/B test LLM prompt templates using Git-native workflows and structured registries.
Activation Triggers
Activate on: "prompt versioning", "prompt template", "A/B test prompts", "prompt registry", "manage prompt variants", "prompt as code", "parameterized prompts", "prompt lifecycle"
NOT for: Writing or optimizing individual prompts (prompt-engineer), fine-tuning dataset preparation (fine-tuning-dataset-curator), or LLM application architecture (ai-engineer)
Quick Start
- Structure prompts as files — Store each prompt template in
prompts/{domain}/{name}.yaml with metadata, variables, and versioned content.
- Parameterize with variables — Use
{{variable}} placeholders for dynamic content; separate static instruction from dynamic context.
- Version with Git — Each prompt change is a commit. Tags mark production versions. Branches for experiments.
- A/B test — Deploy multiple variants simultaneously, route traffic by percentage, collect metrics per variant.
- Promote winners — Evaluate via automated harness or human review, promote winning variant to production tag.
Core Capabilities
| Domain | Technologies | Notes |
|---|
| Template Format | YAML + Jinja2, Handlebars, Mustache | YAML frontmatter for metadata, body for template |
| Version Control | Git tags, branches, semantic versioning | v1.2.0 tags for production, branches for experiments |
| A/B Testing | Feature flags (LaunchDarkly, Unleash, custom) | Percentage-based routing with metrics collection |
| Registry | File-based, PostgreSQL, Redis | Central lookup for template resolution at runtime |
| Evaluation | LLM-as-judge, human eval, RAGAS | Automated comparison of variant outputs |
| Rendering | Jinja2 (Python), Handlebars (JS), custom | Variable interpolation with type validation |
Architecture Patterns
Pattern 1: File-Based Prompt Registry
