Prompt Template Manager (v2.0)

Prompt Template Manager is a definitive Python toolkit for engineering complex, versioned, and reusable LLM prompts. It transforms prompts from static strings into managed assets that can be composed, versioned for specific models, and hot-reloaded in production.

---

📖 Table of Contents

• Quick Start
• Installation
• 1. Configuration Architecture
  • The YAML Schema
  • Prompt Versioning
  • Chat/Messages Format
  • Metadata (_meta)
• 2. Template Logic & Substitution
  • Variable Resolution ({var})
  • Recursive Includes ({include})
  • Environment Variables ({env:VAR})
  • Global Context
• 3. API Reference: PromptManager
• 4. API Reference: Prompt
• 5. Advanced Patterns (Few-Shot, etc.)
• 6. CLI Documentation
• 7. Performance & Best Practices

---

⚡ Quick Start

1. Define prompts.yaml

system_persona:
  _default: helpful
  helpful: "You are a helpful assistant."

analyze_text:
  _default: v1
  v1: |
    {include system_persona}
    Summarize this: {text}

2. Run in Python

from prompt_template_manager import PromptManager

# Load and fetch
manager = PromptManager("prompts.yaml")
prompt = manager.get("analyze_text")

# Format and print
print(prompt.format({"text": "LLMs are evolving."}))

---

💾 Installation

# Core library (PyYAML only)
pip install prompt-template-manager

# All features (Watchdog, Tiktoken, Anthropic SDK)
pip install "prompt-template-manager[all]"

---

1. Configuration Architecture

The YAML Schema

The core of the library is the YAML configuration. A file can contain multiple prompts, each identified by a unique key.

Simple (Static) Prompt

Best for snippets that never change.

footer: "End of prompt. Response required."

Versioned Prompt (Standard)

Required for prompts that evolve or require metadata.

task_generator:
  _default: v2
  _meta:
    description: "Generates code tasks"
    author: "Dave"
  v1: "Task: {task}"
  v2: "Revised Task: {task}. Be concise."

Prompt Versioning

You can use standard version keys (v1, v2) or Model Tags to optimize for specific LLMs.

summarizer:
  _default: gpt-4
  gpt-4: "Summarize: {text}"
  claude-3: "Please summarize: {text}"
  llama-3: "Summ: {text}"

Chat/Messages Format

Prompts can be defined as a list of role-content pairs.

code_review:
  _default: v1
  v1:
    - role: system
      content: "You are a senior dev."
    - role: user
      content: "Check this {language} code:\n{code}"

Metadata (_meta)

Metadata is a key-value store (Strings only) for tracking prompt provenance.

my_prompt:
  _default: v1
  _meta:
    max_tokens: "500"
    temperature: "0.7"
  v1: "..."

---

2. Template Logic & Substitution

Variable Resolution ({var})

Any text in curly braces that isn't a reserved tag (include or env) is a variable.

Mode	Usage	Behavior on Missing Variable
Strict	p.format(data, strict=True)	Raises ValueError listing missing keys.
Non-Strict	p.format(data, strict=False)	Leaves the placeholder as-is (e.g., {missing_var}).

Recursive Includes ({include})

Inject one prompt into another. If the included prompt is versioned, its _default version is used.

persona: "You are an AI."
task: "{include persona} Solve: {problem}"

Note: Circular references will raise an ImportError.

Environment Variables ({env:VAR})

Resolved at load time (when PromptManager is initialized).

• Success: Replaces tag with os.environ[VAR].
• Failure: Raises ValueError at load time if the variable is not set.

system_warning: "Running in {env:ENVIRONMENT} mode."

Global Context

Inject variables into every prompt managed by a specific PromptManager.

manager = PromptManager("prompts.yaml", context={"app_id": "PTM-001"})
# Every prompt in this manager now knows {app_id}

---

3. API Reference: PromptManager

PromptManager is thread-safe for reading (.get()) after initialization.

Initialization & Loading

def __init__(
    self, 
    source_path: str | Path, 
    context: dict[str, Any] | None = None, 
    watch: bool = False
) -> None

• source_path: File or directory. If a directory, all .yaml/.yml files are loaded.
• Collision Rule: If multiple files define the same prompt name, PromptManager raises a ValueError during initialization to prevent ambiguous resolution.

Hot Reloading (Watch Mode)

Requires watchdog. Automatically reloads the library when source files change.

manager = PromptManager("prompts.yaml", watch=True)

Retrieving Prompts (.get)

def get(
    self, 
    name: str, 
    version: int | str | None = None, 
    model: str | None = None
) -> Prompt

• Raises KeyError if not found, providing Levenshtein-based suggestions (e.g., "Did you mean 'summarize'?").

Cataloging (.list)

def list(self) -> list[dict[str, Any]]

Returns a list of dictionaries describing every loaded prompt, its versions, and required variables.

---

4. API Reference: Prompt

Formatting & Substitution

def format(
    self, 
    data: dict[str, Any], 
    strict: bool = True
) -> str | list[dict[str, str]]

Returns a string for text prompts, or a list of messages for chat prompts.

Partial Application

def partial(self, data: dict[str, Any]) -> Prompt

Fill some variables now, returning a new Prompt object for the remaining variables.

Token Estimation

def estimate_tokens(
    self, 
    provider: str = "openai", 
    model: str = "gpt-4"
) -> int

Requires tiktoken (OpenAI) or anthropic SDK. Raises ImportError if missing.

SDK Compatibility

def to_openai_messages(self, data: dict[str, Any]) -> list[dict[str, str]]
def to_anthropic_messages(self, data: dict[str, Any]) -> list[dict[str, str]]

---

5. Advanced Patterns

Few-Shot Prompting

Use includes to manage your example bank separately from your task logic.

YAML:

# examples.yaml
sentiment_examples: |
  Text: "I love this!" | Sentiment: Positive
  Text: "This is bad." | Sentiment: Negative

# classifier.yaml
classify_task: |
  {include sentiment_examples}
  Text: "{input_text}" | Sentiment:

Model Migration / A/B Testing

# Easily switch between model-optimized versions
p_gpt = manager.get("task", model="gpt-4")
p_claude = manager.get("task", model="claude-3")

---

6. CLI Documentation

validate

Recursively checks all YAML files in a directory for syntax, type errors, and circular includes.

ptm validate ./prompts

list

Prints a human-readable table of prompts, versions, and variables.

ptm list prompts.yaml

get

Renders a prompt to STDOUT.

ptm get prompts.yaml my_prompt --version v1

---

7. Performance & Best Practices

1. Include Caching: PromptManager uses an internal LRU cache for resolved includes. Deeply nested templates resolve in constant time after the first access.
2. Directory Loading: We recommend a "One Domain per File" approach:
   • prompts/auth_flow.yaml
   • prompts/data_extraction.yaml
3. Environment Isolation: Use {env:VAR} for instructions that vary by deployment (e.g., {env:LLM_SAFETY_LEVEL}).

---

Created by Dave Manufor — 2026