Prompt Template Editing

This guide covers editing the LLM prompts used by the adaptation pipeline and prompt enhancement system.

Overview

All LLM prompts are stored in the PromptTemplate model — database-only, with no filesystem fallback. This enables live editing via the Django admin without code deployment.

Templates use Jinja2 syntax, support versioning (one active version per slug), and include usage analytics (render count and last-used timestamp).

Editing a Template

1. Navigate to Core > Prompt Templates

2. Select the template to edit

3. Modify the Template field (Jinja2 syntax)

4. Save — this automatically:

   • Creates a new version (auto-incremented)

   • Deactivates the previous version

   • Invalidates the cache (changes take effect within 5 minutes)

5. No deployment needed — changes are live immediately

Template Categories

Nine templates organized into four categories:

Slug	Category	Purpose
prompt-enhancer-system	Enhancement	System prompt for HF/Anthropic prompt enhancers
prompt-enhancer-user	Enhancement	User prompt for enhancement requests
adaptation	Adaptation	Main TV spot cultural adaptation (script writing)
concept-extraction	Concept	Analyzes original script for core concept
cultural-research	Concept	Produces cultural brief for target market
eval-format	Evaluation	Evaluates language compliance
eval-cultural	Evaluation	Evaluates cultural appropriateness
eval-concept	Evaluation	Evaluates concept fidelity
eval-brand	Evaluation	Evaluates brand consistency

Jinja2 Syntax

Templates use Jinja2 with trim_blocks and lstrip_blocks enabled:

You are adapting a TV commercial for {{ target_market_name }}.

## Original Script
{{ original_json }}

{% if brand_guidelines %}
## Brand Guidelines
{{ brand_guidelines }}
{% endif %}

## Instructions
Adapt the script for the target market while preserving the core concept.

Variables are passed as keyword arguments to render_prompt():

from cw.lib.prompts import render_prompt

prompt = render_prompt(
    "adaptation",
    target_market_name="Germany",
    original_json=script_data,
    brand_guidelines=brand.guidelines,
)

Jinja2 syntax is validated on save — if the template contains invalid syntax, a ValidationError is raised and the save is rejected.

Template Variables

Each template declares its expected variables in the expected_variables JSON field:

{
  "target_market_name": {
    "type": "string",
    "required": true,
    "description": "Name of the target market (e.g., 'Germany')"
  },
  "original_json": {
    "type": "string",
    "required": true,
    "description": "Original script as JSON string"
  }
}

This field is informational — it documents what variables the template expects but does not enforce them at render time.

Versioning

Every edit creates a new version:

• Version number — auto-incremented per slug

• Active flag — only one version per slug can be active at a time

• Previous versions — retained for audit trail and rollback

To roll back to a previous version:

1. Find the previous version in the template list (filter by slug)

2. Set its is_active flag to True

3. Save — the current version is automatically deactivated

Cache Invalidation

Rendered templates are cached with a 5-minute TTL:

• Cache key: prompt_template:{slug}

• Backend: Django cache (Redis or memory)

• TTL: 300 seconds

After editing a template, changes take effect within 5 minutes as cached entries expire. For immediate effect, clear the cache manually or restart the worker.

Usage Analytics

Each template tracks:

• usage_count — number of times rendered (incremented atomically via F() expression)

• last_used_at — timestamp of the most recent render

These fields are visible on the template detail page and help identify which templates are actively used.

Import & Export

# Import templates from data/prompt_templates.json
uv run manage.py import_prompt_templates

# Preview without importing
uv run manage.py import_prompt_templates --dry-run

# Export active templates to data/prompt_templates.json
uv run manage.py export_prompt_templates

# Export to a custom directory
uv run manage.py export_prompt_templates --dir custom/

The data/prompt_templates.json file is the source of truth for bootstrapping new environments.