Templates control how Autonomy Realms communicates with AI models during analysis and embedding generation. Every template in this system is customizable per realm — your changes override the system defaults without modifying the original files.
When you trigger an analysis or generate an embedding, the system loads prompt templates and fills in placeholders with your data. There are three layers:
{{placeholders}}.These templates structure the user-facing prompt sent to the AI during analysis. They contain placeholders like {{questions}}, {{response_format}}, and {{field_definitions}} which are automatically populated based on which fields you've selected for analysis.
You can customize the framing, instructions, and structure — but the dynamic placeholders ensure the prompt stays in sync with your field selections.
Each analysis pass (surface and structure) has a questions.json file that defines every field the AI can analyze. Each field has three parts:
{
"title": {
"question": "Generate a clear, descriptive title...",
"format": "\"title\": \"...\"",
"definition": "Clear, descriptive title under 10 words."
}
}
{{questions}} placeholder.{{response_format}} placeholder to show the AI exactly what shape its response should take.{{field_definitions}} placeholder to give the AI additional context.When you run an analysis with specific fields selected (e.g., only "title" and "tags"), only those fields' questions, formats, and definitions are included in the prompt. Fields you didn't select are omitted entirely.
When analyzing a cluster, the AI needs to see summaries of the signals inside it. The Signal Template defines how each individual signal is rendered in the cluster's analysis prompt.
These templates use conditional blocks so empty fields are omitted:
{{#signal_energy}}**Energy:** {{signal_energy}}{{/signal_energy}}
If a signal has no signal_energy value, the entire line disappears from the prompt. This keeps prompts clean — a PHOTO signal won't show empty transmission-specific fields.
The same pattern applies to Cluster Templates used in shape analysis.
Embedding prompts define how entity data is formatted before being sent to the embedding model (text-embedding-3-small) for vector generation. The resulting vector enables semantic search and similarity matching.
Each entity type (signal, cluster, shape) has its own embedding prompt. Clusters and shapes also have sub-templates that define how their children (signals or clusters) are rendered within the embedding text.
All templates support conditional blocks using this syntax:
{{#field_name}}Content when field has a value{{/field_name}}
If field_name is empty or null, the entire block (including the line) is removed from the output. If it has a value, the block markers are stripped and any {{field_name}} inside is replaced with the actual value.
Use conditionals to prevent empty labels from cluttering prompts. Use raw {{field_name}} when you always want the field present regardless of value.
Templates auto-save 1.5 seconds after you stop typing. Your customizations are stored as realm settings — they override the default file but don't modify it.
Revert to Default deletes your custom version and restores the system default from the file. This is permanent — your customizations will be lost.
The placeholder tray below the editor shows all available placeholders for the selected template. Placeholders already in use are hidden. Click raw to insert a direct replacement or cond to insert a conditional block.
Placeholders that say "(auto-built)" like {{response_format}} and {{field_definitions}} are dynamically generated from the questions file — you don't fill them in manually, but you control where they appear in the template.