Source image candidates for a brand-new WorldSeed scene, let an AI visually review them, and generate a local review HTML for human selection. Use when this capability is needed.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Source image candidates for a brand-new WorldSeed scene, let an AI visually review them, and generate a local review HTML for human selection. Use when this capability is needed.
metadata
{"author":"AIScientists-Dev"}
Asset Sourcing
Use this skill when a user wants image candidates for a new or existing WorldSeed scene and the goal is to build a reviewable asset board, not auto-pick final assets.
This skill is standalone. It stops at a reviewable bundle and selected picks. Import into a scene should happen later as a separate step.
Supported Sources
Default sources:
openverse
aic
cleveland
vam
wellcome
ycba
met
Optional sources:
wikimedia
walters
nasa
Read before using a source you have not used recently.
references/source-catalog.md
Execution Rules
Use supported sources only.
Treat sourcing as entity-grounded, not exact-illustration-only. exact, adjacent, and vibe can all be usable.
Do not stop after one keyword search.
Let the AI choose keywords and source order per entity.
Search only the sources that are justified for the current entity.
Search different sources in parallel when useful.
Start with likely fast sources and widen only when needed.
Treat met as a later pass, not a universal first call.
Only use optional sources when they are justified:
wikimedia for known motifs, named figures, or coverage gaps
walters for watches, keys, portraits, and manuscript-adjacent objects
nasa for cosmic or scientific entities only
Preserve source URLs, image URLs, and rights text in the manifest.
Search first. Download only shortlisted candidates by default.
Treat metadata as recall only. Final retention and top picks must be image-verified.
Reject obvious homonym and name-collision matches when the image subject is wrong, even if the title contains the entity token.
Prefer a visually coherent set across entities. When multiple candidates are equally valid, favor the ones that fit the scene's overall medium, period, and mood.
Avoid obvious horror, gore, medical shock, body horror, or otherwise unsettling imagery unless the user explicitly wants that tone.
Do not auto-pick final assets. The default output is a candidate board for agent/human review in HTML.
The review page is the main product. It should let a human quickly scan the strongest reviewed picks, open more options only when needed, and choose final assets.
The current picker behavior is:
one long scroll page
one section per entity
reviewed picks appear first with Top 1/2/3 only when the manifest contains trusted visual-review recommendations
if there are no reviewed picks yet, the first few candidates still appear in the same slot, but without Top badges
no metadata-fit language should drive the unreviewed UI
click one card per entity to select it
export a minimal picks object with:
title
source
8. Apply visual review recommendations
If an agent or human has visually reviewed the downloaded images and chosen ranked picks, write them back into the manifest before the final render.
Review file shape:
{"reviewer":"agent-name","picks":{"raven":[{"source":"wikimedia","title":"Kalila and Dimna - The Owls and the raven","reason":"Only candidate that clearly reads as a raven scene."}]}}
