| name | generate-thumbnails |
| description | Draft thumbnail concepts (prompt, focal emotion, text overlay) for a content run by calling Hermes's thumbnail judgment agent. Use when the Content Factory agent must produce the thumbnail_concepts that Hermes's content-run pipeline needs to render actual thumbnail images via thumbnail-generation. |
| metadata | {"agent":"content-factory","source":"Project Atlas Requirements §3.5/10.2","layer":"L5 Content Factory","host":"railway"} |
generate-thumbnails
Draft thumbnail concepts (text descriptions, not images) for a content run.
When to use
- Alongside
plan-assets, before dispatching to Hermes's content-run pipeline.
Inputs / Sources
Script + Vibe + archetype context.
Output
{concepts: [{prompt, focal_emotion, text_overlay}]} (see hermes/agents/thumbnail/schema.json) —
the concepts list is exactly the shape ContentRunRequest.thumbnail_concepts expects;
Hermes's thumbnail-generation production agent turns each concept into an actual Higgsfield
image once the content run starts (see hermes/agents/thumbnail-generation/executor.py).
Function signature (manifest contract)
def generate_thumbnail_concepts(script_text: str, vibe: dict, archetype_context: dict) -> dict:
"""Returns {"concepts": [{"prompt": str, "focal_emotion": str, "text_overlay": str}, ...]}."""
Implementation
Call the hermes-bridge MCP tool run_judgment_agent(insight_type="thumbnail", query=<ask>, context={"script_text": script_text, "vibe": vibe, **archetype_context}). Hits POST /v1/agents/thumbnail — same self-correcting loop / mock-fallback pattern, no new backend. The actual image generation (Higgsfield) happens later, deterministically, inside Hermes's thumbnail-generation production agent.
Model
deepseek-direct/deepseek-chat — Cheap high-volume tier (spec §11.3).
