world-builder

World Builder

A trained custom model locks one character's likeness. It does not, by itself, tell you what that character's world looks like — the palette its packaging uses, whether it has a village, what its landmarks are, whether "family" scenes read as a community or as a crowd of clones. This skill is the staged pipeline for answering that, developed over several real sessions building out the world around PorchPing's mascot ("Ding-Bot") and generalized here so it applies to any already-trained character.

The whole pipeline is sequenced around one constraint: every image that uses the custom model spends training-specific budget and risks compounding whatever quirks the model has (see references/character-batch-discipline.md), so cheaper no-model passes come first, cheaper batches come before expensive ones, and nothing gets called "done" without a human visually reviewing it.

Prerequisite — do not start without this

A custom model must already exist for the character, trained via the custom-model-training skill (dataset → train_model → get_model polling → custom_model_uri). If the user hasn't mentioned a custom_model_uri, ask for it or run mcp__ideogram__list_models to find it before doing anything else. Don't start generating "world" images against a character description alone — the entire point of this pipeline is testing how a trained model behaves across many contexts, not describing the character freshly in each prompt (which character-model-sheet already does, and which drifts exactly the way custom-model-training's existence is meant to prevent).

Cross-cutting discipline — applies to every step below, no exceptions

Read these two reference files once, before step 1, and re-apply them on every single prompt for the rest of the pipeline:

• references/palette-lock.md — the locked palette (paper/dominant, primary, secondary, ink/trim, and at most one reserved accent used in exactly one place) must be defined and quoted, verbatim, in every prompt from step 1 onward. A world built on a drifting palette isn't a world, it's six unrelated images.
• references/anti-slop-discipline.md — the reusable, brand-agnostic ban list (glowing orbs, neural-network nodes, circuit-board textures, gradient washes, stock-photo people, glossy mirror-shine, plastic-toy uncanny valley, and the rest). Run the pre-generation gate in this file before every generate_image / generate_images_bulk call, the same discipline character-model-sheet and brand-identity-sheet already apply to their own single-image gates, scaled up to a multi-batch pipeline.