Prompt Store

The Prompt Store provides centralized management for reusable prompt templates. Prompts are written in Markdown files with Jinja2 templating support.

Syntax

Prompt Definition

{prompt Extract_Facts}
Extract facts from the following content.

Content: {{ source.page_content }}

Return JSON matching the schema.
{end_prompt}

Prompt Reference

Reference prompts in YAML using the $ syntax:

- name: extract_facts
  prompt: $qanalabs_quiz_gen.Extract_Facts

Format: $filename.Prompt_Name (filename without .md)

Directory Structure

Prompts can be at project level (shared) or workflow level (domain-specific):

project/
├── prompt_store/                    # Shared prompts
│   └── common.md
└── agent_workflow/
    └── my_workflow/
        └── prompt_store/            # Workflow-specific prompts
            └── my_prompts.md

Agent Actions searches recursively—use unique filenames across your project.

Creating Prompts

With Seed Data

{prompt Fact_extraction}
Extract facts relevant to the {{ seed.exam_syllabus.exam_name }} exam.

## Target Audience
{{ seed.exam_syllabus.audience_profile.description }}

## Source Content
{{ source.page_content }}
{end_prompt}

With Jinja2 Loops

{prompt Generate_Questions}
{% for skill in seed.exam_syllabus.skills_measured %}
## {{ skill.skill_area }}
{% for objective in skill.objectives %}
- {{ objective }}
{% endfor %}
{% endfor %}
{end_prompt}

With Conditionals

{prompt Validate_Content}
{% if source.content_type == "technical" %}
Apply strict technical accuracy checks.
{% else %}
Apply general readability checks.
{% endif %}

Content: {{ source.content }}
{end_prompt}

Workflow Reference

actions:
  - name: fact_extractor
    prompt: $qanalabs_quiz_gen.Fact_extraction
    schema: candidate_facts_list

  - name: canonicalize_facts
    dependencies: fact_extractor
    prompt: $qanalabs_quiz_gen.Canonicalize_Facts
    schema: candidate_facts_list

Inline Prompts

For simple, one-off prompts, use inline YAML:

- name: simple_validate
  prompt: |
    Validate these facts: {{ extract_facts.facts }}
    Return: {"valid": true/false, "reason": "..."}

Use inline when prompt is specific to one action. Use prompt store when prompts are reused or complex.

Template Variables

Variable	Source	Example
{{ source.field }}	Input record data	{{ source.page_content }}
{{ seed.name.field }}	Seed data files	{{ seed.exam_syllabus.exam_name }}
{{ action_name.field }}	Upstream action output	{{ extract_facts.facts }}

Jinja2 Features

Filters

{{ source.data | tojson }}              {# Serialize as JSON (most common) #}
{{ source.text | upper }}               {# Uppercase text #}
{{ source.items | length }}             {# Count items #}
{{ source.optional_field | default("N/A") }}  {# Fallback value #}
{{ source.tags | join(", ") }}          {# Join list as string #}

tojson Filter

The | tojson filter is essential when passing structured data (dicts, lists) into prompts. Without it, Python's string representation is used, which can confuse the LLM. All examples use | tojson for seed data and complex fields:

{{ seed.brand_voice | tojson }}
{{ extract_context.key_terms | tojson }}

Version Variables

When using versioned actions, these variables are available in prompts:

Variable	Type	Description
{{ i }}	int	Current iteration number (from range)
{{ version.length }}	int	Total number of iterations

You are evaluator {{ i }} of {{ version.length }}.

Load Other Prompts

{prompt Main_Analysis}
{{ load_prompt("common.Standard_Header") }}

## Analysis Content
{{ source.content }}
{end_prompt}

Advanced Patterns (Production Cookbook)

These patterns come from production agentic workflows and show what's possible with Jinja2 in prompts.

Arithmetic in Templates

Use inline math to set dynamic constraints based on upstream tool output:

{prompt Generate_Distractor}
{% if add_answer_text.distractor_1_word_target == "lesser_than" %}
Your distractor must be SHORTER: Write {{ add_answer_text.correctAnswerWords - 2 }} to {{ add_answer_text.correctAnswerWords - 1 }} words
{% elif add_answer_text.distractor_1_word_target == "equal_to" %}
Your distractor must match length: Write exactly {{ add_answer_text.correctAnswerWords }} words
{% elif add_answer_text.distractor_1_word_target == "greater_than" %}
Your distractor must be LONGER: Write {{ add_answer_text.correctAnswerWords + 2 }} to {{ add_answer_text.correctAnswerWords + 4 }} words
{% endif %}
{end_prompt}

Loop Index to Letter Conversion

Convert numeric loop indices to letters (A, B, C, D) using chained filters:

{prompt Format_Options}
{% for option in reconstruct_options.options %}
{{ loop.index | string | replace('1', 'A') | replace('2', 'B') | replace('3', 'C') | replace('4', 'D') }}. {{ option }}
{% endfor %}
{end_prompt}

Dictionary Iteration with Filtering

Iterate over a dictionary's key-value pairs and filter by key pattern:

{prompt Summarize_Alternatives}
{% for key, value in merge_alternatives.items() %}
{% if 'issue_description' in key %}
- {{ value }}
{% endif %}
{% endfor %}
{end_prompt}

Array of Objects Iteration

Loop through structured arrays with nested field access:

{prompt Analyze_References}
{% for ref in source.referenced_in %}
- **Section**: {{ ref.section_name }}
  **Objective**: {{ ref.objective }}
  **Relevance**: {{ ref.relevance }}
{% endfor %}
{end_prompt}

Deep Seed Data Traversal

Access deeply nested seed data structures:

{prompt Extract_Facts}
Extract facts relevant to the {{ seed.exam_syllabus.exam_name }} exam.

## Target Audience
{{ seed.exam_syllabus.audience_profile.description }}

## Skills Measured
{% for skill in seed.exam_syllabus.skills_measured %}
### {{ skill.skill_area }} (Weight: {{ skill.weight }})
{% for section in skill.sections %}
- {{ section }}
{% endfor %}
{% endfor %}
{end_prompt}

Chaining Upstream References to Prevent Duplication

When generating multiple alternatives, reference previous outputs to avoid repetition:

{prompt Generate_Distractor_2}
## EXISTING OPTIONS (DO NOT DUPLICATE)
- Correct answer: {{ add_answer_text.answer_text }}
- Distractor 1: {{ generate_distractor_1.distractor_1 }}

Write a DIFFERENT distractor that uses a wrong approach.
{end_prompt}

{prompt Generate_Distractor_3}
## EXISTING OPTIONS (DO NOT DUPLICATE)
- Correct answer: {{ add_answer_text.answer_text }}
- Distractor 1: {{ generate_distractor_1.distractor_1 }}
- Distractor 2: {{ generate_distractor_2.distractor_2 }}

Write a DIFFERENT distractor that covers an edge case.
{end_prompt}

Ensemble Voting with Version Variables

Create independent evaluators that don't see each other's work:

{prompt Validate_Answer}
You are validator {{ i }} of {{ version.length }}, working independently.

## Question
{{ reconstruct_options.question }}

## Options
{% for option in reconstruct_options.options %}
{{ loop.index | string | replace('1', 'A') | replace('2', 'B') | replace('3', 'C') | replace('4', 'D') }}. {{ option }}
{% endfor %}

## Claimed Answer
{{ reconstruct_options.answer }}

Verify whether the claimed answer is correct based on the source material.
{end_prompt}

Best Practices

1. Use descriptive names: {prompt Extract_Technical_Facts} not {prompt prompt1}
2. Structure complex prompts: Use consistent section headers (CONTEXT, TASK, OUTPUT FORMAT)
3. Separate concerns: Group related prompts in files (extraction.md, validation.md)

Debugging Prompts

Enable prompt_debug to see rendered prompts:

- name: extract_facts
  prompt: $qanalabs_quiz_gen.Fact_extraction
  prompt_debug: true