メニュー
Perform gradient adjustments on the Trend Engine opportunity-scoring weights to align suggested opportunities with actual views and retention growth. Use when DS-Star runs its nightly tuning pass to close the analytics→backlog feedback loop.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Query the keyless academic and public-domain registries (OpenAlex, Crossref, Semantic Scholar, Project Gutenberg, Internet Archive, Stack Exchange) for a topic and get back normalized, copyright-safe reference metadata (title, link, provenance). Use when gather-citations or verify-claims needs primary-literature grounding instead of the LLM's training memory.
Query Wikipedia, Wikidata, Stack Exchange, public Discourse forums, academic APIs, and public-domain archives to synthesize facts, human struggles, and philosophical debates to build the narrative arc for a topic. Store only facts/events/dates/relationships, dialogic tension points, and reference metadata — never full articles (copyright-safe pipeline, spec §2.5).
Query Metacritic, GameSpot, and TVGuide (resolved and read via Bright Data) for a personality's existing game/film/TV adaptation, extracting critic/user scores, narrative themes critics highlighted, and reception debates. Use when generate-psych-profile or backlog-ranking want a critical-reception signal on top of gather-citations' facts and fandom-lore's fan interpretation.
Compose the per-beat music score plan (Lyria cues, deliberate silence segments, optional sound-effect cues) for a content run by calling Hermes's music-director judgment agent. Use when the Content Factory agent must produce the music_director_output that Hermes's content-run pipeline needs for music generation, beat-synced editing, and the SFX layer.
Decide the whole-video visual grade, typography, and per-shot treatments (transitions, camera moves, grade deviations, light leaks) for a content run by calling Hermes's visual-director judgment agent. Use when the Content Factory agent must produce the visual_director_output that Hermes's content-run pipeline needs for shot styling in Video Assembly.
Trigger Hermes's deterministic 11-agent video-production pipeline for an editorially-approved run, and poll it to completion. Use when submit-editorial-review records a human PASS and the approved asset bundle is ready to actually produce.
| name | optimize-scoring-weights |
| description | Perform gradient adjustments on the Trend Engine opportunity-scoring weights to align suggested opportunities with actual views and retention growth. Use when DS-Star runs its nightly tuning pass to close the analytics→backlog feedback loop. |
| metadata | {"agent":"ds-star","source":"Project Atlas Agent Skills Manifest §7","layer":"executive-strategy","production_host":"railway"} |
Tune the opportunity-scoring weights from realized performance.
def optimize_scoring_weights(actual_vs_predicted: list) -> dict:
"""Gradient-adjust Trend Engine scoring weights to match actual views/retention."""
actual_vs_predicted (list of {topic, predicted_score, actual_views, actual_retention}),
if available -- revenue_actuals/story_scores (Railway) are still stubbed pending the
Postgres/AGE backbone, so this often won't be populated yet.{generated_at, optimized_weights: {search, discussion, video, evergreen, emotion, competition, monetization, regional}, learning_rate, status, used_stub_data} (see
hermes/agents/optimize-weights/schema.json).Apply a revenue_multiplier from trailing-90-day revenue_actuals per archetype on top of the
engagement score, so archetypes that convert (not just engage) get prioritized.
Until revenue_actuals/story_scores are wired, the underlying agent checks vidiq_user_channels
for a connected channel and uses vidiq_channel_analytics against it for real actual_vs_predicted
figures when one exists. If no channel is connected yet, it proceeds on whatever
actual_vs_predicted was passed in this skill's own input and marks used_stub_data: true in
the output -- check that field before trusting a run's weights as grounded in real data.
Call the hermes-bridge MCP tool run_judgment_agent(insight_type="optimize-weights", query=<ask>, context={"actual_vs_predicted": actual_vs_predicted}). Hits Hermes's POST /v1/agents/optimize-weights (hermes/agents/optimize-weights/) — note the insight_type is optimize-weights, not this skill's own name. This is the standard LLM judgment-loop path (execute_self_improving_hermes_agent), not a separate pure-Python implementation -- no such implementation exists in this repo today. The TS ds-star-scientists Temporal activities (server/src/modules/temporal/activities/ds-star-scientists.activities.ts) don't currently call an optimize-weights scientist at all, so this skill is the only path that reaches it.
backlog_rankings, revenue_actuals, story_scores (Railway) feed Hermes's own data pipeline, not this skill directly -- still stubbed; VidIQ own-channel data is the stopgap until then (see above).Runs as a standard Hermes judgment-loop agent (see Implementation) -- not zero-LLM despite earlier documentation here claiming otherwise.