| name | generate-seo-metadata |
| description | Draft a YouTube-optimized title, description, and tag list for a content run by calling Hermes's seo-metadata judgment agent, grounded in real keyword data via vidiq. Use when the Content Factory agent must produce the publish_metadata.title/description/tags that Hermes's content-run pipeline needs before dispatching to video-publish. |
| metadata | {"agent":"content-factory","source":"Project Atlas Requirements §3.5/10.2","layer":"L5 Content Factory","host":"railway"} |
generate-seo-metadata
Draft a search-optimized title, description, and tag list for a content run.
When to use
- Alongside
plan-assets/generate-thumbnails, before dispatching to Hermes's content-run pipeline.
Inputs / Sources
Script + hook + archetype context. Use the vidiq MCP server's vidiq_keyword_research tool
(mode: research) to ground the title's primary keyword and the tag list in real search-volume/
competition data, rather than guessing keywords from judgment alone. vidiq is the prioritized
subscription; fall back to nexlev only if vidiq can't answer the lookup.
Output
{title, description, tags} (see hermes/agents/seo-metadata/schema.json) — this shape is
exactly what ContentRunRequest.publish_metadata expects for its title/description/tags
fields; dispatch-hermes-content-run assembles them in alongside privacy_status (which stays
chief-editor's own call, unrelated to SEO).
Function signature (manifest contract)
def generate_seo_metadata(script_text: str, hook: str, archetype_context: dict) -> dict:
"""Returns {"title": str, "description": str, "tags": [str, ...]}."""
Implementation
Call the hermes-bridge MCP tool run_judgment_agent(insight_type="seo-metadata", query=<ask>, context={"script_text": script_text, "hook": hook, **archetype_context}). Hits POST /v1/agents/seo-metadata — same self-correcting loop / mock-fallback pattern, no new backend.
Model
deepseek-direct/deepseek-chat — Cheap high-volume tier (spec §11.3).
