Prompt Variables and Templating: Dynamic Content Injection with Jinja2 and f-strings

Why Static Prompts Fall Short

Hardcoded prompts work for demos. Production agents need prompts that adapt — inserting the user's name, adjusting tone based on context, including relevant data, and conditionally enabling features. This is prompt templating: defining a prompt structure once and injecting dynamic values at runtime.

The two dominant approaches in Python are f-strings for simple cases and Jinja2 for complex logic. Understanding when to use each prevents both over-engineering and under-engineering your prompt layer.

f-string Templating: Simple and Direct

For prompts with straightforward variable substitution, Python f-strings are the fastest path.

flowchart TD
    SPEC(["Task spec"])
    SYSTEM["System prompt<br/>role plus rules"]
    SHOTS["Few shot examples<br/>3 to 5"]
    VARS["Variable injection<br/>Jinja or f-string"]
    COT["Chain of thought<br/>or scratchpad"]
    CONSTR["Output constraint<br/>JSON schema"]
    LLM["LLM call"]
    EVAL["Offline eval<br/>LLM as judge plus regex"]
    GATE{"Score over<br/>threshold?"}
    COMMIT(["Promote to prod<br/>version pinned"])
    REVISE(["Revise prompt"])
    SPEC --> SYSTEM --> SHOTS --> VARS --> COT --> CONSTR --> LLM --> EVAL --> GATE
    GATE -->|Yes| COMMIT
    GATE -->|No| REVISE --> SYSTEM
    style LLM fill:#4f46e5,stroke:#4338ca,color:#fff
    style EVAL fill:#f59e0b,stroke:#d97706,color:#1f2937
    style COMMIT fill:#059669,stroke:#047857,color:#fff

def build_support_prompt(
    user_name: str,
    account_tier: str,
    issue_summary: str
) -> str:
    """Build a support agent prompt with user context."""
    return f"""You are a customer support agent for Acme Corp.

The customer's name is {user_name}.
Their account tier is {account_tier}.

Issue summary: {issue_summary}

Respond helpfully and professionally. If the customer
has a Premium or Enterprise tier, prioritize their request
and offer direct escalation options."""

This is readable and type-safe — your IDE catches missing variables. However, f-strings hit limits quickly. You cannot loop over lists of items, conditionally include sections, or reuse template fragments.

Jinja2 Templating: Full Power

Jinja2 gives you conditionals, loops, filters, template inheritance, and macros. It is the standard for complex prompt templating.

from jinja2 import Environment, FileSystemLoader, select_autoescape

class PromptTemplateEngine:
    """Render prompts using Jinja2 templates."""

    def __init__(self, templates_dir: str = "prompt_templates"):
        self.env = Environment(
            loader=FileSystemLoader(templates_dir),
            autoescape=select_autoescape(default=False),
            trim_blocks=True,
            lstrip_blocks=True,
        )

    def render(
        self, template_name: str, **variables
    ) -> str:
        """Render a named template with variables."""
        template = self.env.get_template(template_name)
        return template.render(**variables)

Store templates as separate files.

# prompt_templates/support_agent.md.j2
# ---
# Template: support_agent
# Variables: user_name, account_tier, conversation_history,
#            available_tools, escalation_allowed
# ---

You are a customer support agent for Acme Corp.
Customer: {{ user_name }} ({{ account_tier }} tier)

{% if conversation_history %}
## Previous Conversation
{% for msg in conversation_history %}
{{ msg.role | upper }}: {{ msg.content }}
{% endfor %}
{% endif %}

## Available Actions
{% for tool in available_tools %}
- {{ tool.name }}: {{ tool.description }}
{% endfor %}

{% if account_tier in ["premium", "enterprise"] %}
This is a high-priority customer. You may offer:
- Direct phone callback within 1 hour
- Escalation to a senior specialist
{% endif %}

{% if not escalation_allowed %}
Note: Do NOT offer escalation options in this session.
{% endif %}

# Usage
engine = PromptTemplateEngine()

prompt = engine.render(
    "support_agent.md.j2",
    user_name="Alice Chen",
    account_tier="premium",
    conversation_history=[
        {"role": "user", "content": "My invoice is wrong"},
        {"role": "assistant", "content": "Let me look into that."},
    ],
    available_tools=[
        {"name": "lookup_invoice", "description": "Fetch invoice details"},
        {"name": "create_ticket", "description": "Open a support ticket"},
    ],
    escalation_allowed=True,
)

Custom Filters for Prompt-Specific Needs

Jinja2 filters transform values inline. Add custom filters for common prompt operations.

def setup_prompt_filters(env: Environment):
    """Add prompt-specific Jinja2 filters."""

    def truncate_tokens(text: str, max_tokens: int = 500) -> str:
        """Rough truncation by word count as a token proxy."""
        words = text.split()
        if len(words) <= max_tokens:
            return text
        return " ".join(words[:max_tokens]) + "..."

    def format_list(items: list, style: str = "bullet") -> str:
        """Format a list for prompt readability."""
        if style == "numbered":
            return "\n".join(
                f"{i+1}. {item}" for i, item in enumerate(items)
            )
        return "\n".join(f"- {item}" for item in items)

    def mask_pii(text: str) -> str:
        """Mask email addresses and phone numbers."""
        import re
        text = re.sub(
            r'[\w.+-]+@[\w-]+\.[\w.]+', '[EMAIL]', text
        )
        text = re.sub(
            r'\b\d{3}[-.]?\d{3}[-.]?\d{4}\b', '[PHONE]', text
        )
        return text

    env.filters["truncate_tokens"] = truncate_tokens
    env.filters["format_list"] = format_list
    env.filters["mask_pii"] = mask_pii

Use them in templates: {{ user_message | mask_pii | truncate_tokens(200) }}.

Safety Practices

Dynamic prompts introduce injection risks. User-provided values could contain instructions that hijack the agent's behavior.

class SafePromptRenderer:
    """Render prompts with input sanitization."""

    def __init__(self, engine: PromptTemplateEngine):
        self.engine = engine

    def sanitize_input(self, value: str) -> str:
        """Remove patterns that could be prompt injections."""
        dangerous_patterns = [
            "ignore previous instructions",
            "ignore all instructions",
            "disregard the above",
            "new instructions:",
            "system:",
            "ADMIN OVERRIDE",
        ]
        sanitized = value
        for pattern in dangerous_patterns:
            sanitized = sanitized.replace(
                pattern, "[FILTERED]"
            )
        return sanitized

    def render_safe(
        self, template_name: str, **variables
    ) -> str:
        """Render with all string variables sanitized."""
        safe_vars = {}
        for key, value in variables.items():
            if isinstance(value, str):
                safe_vars[key] = self.sanitize_input(value)
            else:
                safe_vars[key] = value
        return self.engine.render(template_name, **safe_vars)

Always sanitize user-provided inputs before injecting them into prompts. Treat prompt templates like SQL queries — never insert raw user input without validation.

FAQ

When should I use f-strings versus Jinja2?

Use f-strings when your prompt has fewer than five variables and no conditional logic. Switch to Jinja2 when you need conditionals, loops, template inheritance, or when non-engineers need to edit the templates. The readability of Jinja2 templates makes them better for team collaboration.

How do I handle missing template variables?

Configure Jinja2 with undefined=StrictUndefined to raise errors on missing variables rather than silently inserting empty strings. This catches bugs during development. In production, you can use default filters: {{ user_name | default("Customer") }}.

Can prompt injection be fully prevented with sanitization?

No. Blocklist-based sanitization catches known patterns but misses creative bypasses. Layer multiple defenses: sanitize inputs, use structured system-vs-user message separation, validate outputs, and monitor for anomalous agent behavior. Sanitization is one layer in a defense-in-depth strategy.

---

#PromptTemplating #Jinja2 #Python #DynamicPrompts #PromptEngineering #AgenticAI #LearnAI #AIEngineering