Skip to content

DynaPrompt

Prompt management that grows with your LLM application.

DynaPrompt brings the configuration philosophy of Dynaconf to the world of LLM prompt engineering — zero I/O at import time, environment-aware layering, and Jinja2-powered templates.

Why DynaPrompt?

# Prompts are scattered — mixed into application logic
def analyze_call(transcript: str) -> str:
    prompt = f"""You are an expert call analyzer.
Analyze the following call transcript and extract:
- Key issues raised
- Customer sentiment (1-5)
- Agent performance

Transcript:
{transcript}

Respond in JSON."""
    return call_llm(prompt, model="gpt-4o", temperature=0.1)
  • ❌ Hard to update prompts without touching code
  • ❌ No environment-specific overrides (dev vs. prod model)
  • ❌ No validation or token limit guards
  • ❌ No version tracking or audit logs
prompts/analyze_call.md
---
model: gpt-4o
temperature: 0.1
---
You are an expert call analyzer.
Analyze the following call transcript and extract:
- Key issues raised
- Customer sentiment (1-5)
- Agent performance

Transcript: {{ transcript }}

Respond in JSON.
app.py
from dynaprompt import DynaPrompt

prompts = DynaPrompt(settings_files=["prompts/"])
result = prompts.analyze_call.render(transcript=transcript)
  • ✅ Prompts are versioned and reviewable like code
  • ✅ Switch models per environment (dev → gpt-3.5, prod → gpt-4o)
  • ✅ Token limit validation built-in
  • ✅ Every render produces a SHA-256 hash for audit trails

Installation

uv add dynaprompt

pip install dynaprompt

5-Minute Quick Start

Step 1: Create a Prompt File

prompts/greeting.md
---
model: gpt-4o
temperature: 0.7
max_tokens: 512
---
You are a helpful assistant for {{ app_name }}.

Greet the user named {{ user_name }} warmly and ask how you can help them today.

Step 2: Load and Render

from dynaprompt import DynaPrompt

prompts = DynaPrompt(settings_files=["prompts/"])

rendered = prompts.greeting.render(
    user_name="Ahmed",
    app_name="TechTrax"
)

print(rendered.text)         # The final prompt string
print(rendered.config)       # {'model': 'gpt-4o', 'temperature': 0.7, ...}
print(rendered.prompt_hash)  # SHA-256 for audit logging

Step 3: Inspect What's Loaded

# See all available prompts
print(prompts.keys())  # ['greeting', 'support.chat', ...]

# Rich debug info
prompts.inspect()

Core Concepts

graph LR
    A[📁 Prompt Files<br/>.md / .toml / .py] --> B[DynaPrompt]
    C[🌍 Environment<br/>dev / staging / prod] --> B
    D[📦 Variables<br/>.json / .yaml / dicts] --> B
    B --> E[PromptNode]
    E --> F[🎨 .render&#40;**kwargs&#41;]
    F --> G[RenderedPrompt<br/>text + config + hash]
Concept Description
DynaPrompt The central manager — lazy-loads and caches everything
PromptNode A single prompt template with its configuration
RenderedPrompt The output of .render() — text, config, and a SHA-256 hash
Environment A named context (development, production) that controls which values win

What's Next?