Scenario Identity Library
Overview
A game or brand runs on named identities: characters and props that must stay recognizable in every
image that ships. This skill runs that lifecycle end to end: interview a brief, generate and gate a
hero image, file approved shots in a named collection, assemble a sheet, and pull the library back
out for new images. The per-image mechanics live in siblings and are not repeated here: connection
and the core loop in scenario, holding a look with references in scenario-consistency, pass/fail
verdicts in scenario-quality-gate. If a sibling skill named here is missing from your available
skills, load it with skill_get (Scenario MCP) or ask the user to install it
(https://github.com/scenario-labs/skills#install); unattended, proceed from tool schemas and flag
the gap.
Quick reference
| Stage |
Mechanism |
Rule |
| Brief |
interview, written as the must-not-change enumeration |
before anything generates |
| Anchor |
hero generation, gated to pass |
uploaded art skips generation |
| Library |
collection named character: <name> or props: <name> |
gated images only |
| Sheet |
Grid Maker assembly of approved shots |
rebuild after membership changes |
| Reuse |
collections_list prefix match, members via search, references per scenario-consistency |
any later session |
From interview to anchor
Interview before generating: the answers become the baseline enumeration every later image restates
(scenario-consistency), so a gap here becomes drift later. Ask what it is, silhouette and
proportions, palette by name or hex, two or three signature details that make it recognizable (the
cracked headlamp, the brass buckle), the rendering style, and which views the set needs: a full view
alone gives a reference slot too little detail to hold close-ups, so an identity that will be shot
close also needs a straight-on close view of the face or signature detail and an exact profile.
Non-interactive runs take what the task supplies and flag the rest. A user who already has art skips
generation: their drawing or photo is the uploaded anchor, per scenario-consistency.
Pick one reference-capable model for the whole identity before the first run (recommend, then the
schema gate in scenario-consistency): every later view rides its reference field, so a model
without one disqualifies itself here, not at view three. Generate the hero, a neutral full view on a
plain background (a reference slot isolates identity best from an uncluttered anchor: nothing beyond
the brief in frame, no text), then gate it per scenario-quality-gate with a round cap fixed up
front (its default is three); absent that skill and the gate itself, review against the brief
enumeration, same cap. The hero is the anchor everything else cites.
The library is a named collection
collection_create with the exact name character: <name> or props: <name>, then
collection_add_assets for the hero (catalog tools, write lane; the create call takes the name plus
the scope pair, and no description field). The name is the registry key: collections are not a
search target, so a later session finds the library by paging collections_list (read lane) and
matching the prefix, then pulls members with search (target="assets",
filters={"collection_ids": [...]} plus a tags filter for the views wanted, which keeps filed
sheets out of reference pulls). Admission is the gate: only passing images enter, because every
future generation samples this pool as reference truth, and one off-model member poisons every pull
after it. Build out the views baseline-plus-delta, one model_run per view, gate each, file each
pass tagged with its view (asset_add_tags: hero, back, and so on), so a later session can tell
the members apart. A redesign (new wardrobe, haircut, age) rewrites the brief enumeration and needs
a new hero off the old (scenario-consistency): rebuild views and sheet from it, and
collection_remove_assets the superseded members before the next pull, or one pull feeds both
versions into the reference field.
Sheets are assembly, not generation
Scenario's Grid Maker packs approved shots into one sheet image: its id is the fixed first-party
model_scenario-grid-maker (Scenario's single deterministic tool for the job, so discovery would
only re-derive it); read its schema with model_schema_get. At this writing it takes images (a
file array capped at 100; wrap even a lone asset) plus layout fields: columns, rows computed
from the count when unset, padding, backgroundColor, and a fixed cellRatio list (auto
default). It has no prompt field: order the images array in reading order, because the array is
the layout. The sheet is itself an asset, filed ungated (it assembles already-approved shots, it is
not a generation): tag it (asset_add_tags, say sheet), file it, and rebuild it after membership
changes rather than editing it: pull the view-tagged members only (never a bare member list, or an
old sheet becomes a grid cell), and swap the old sheet out with collection_remove_assets. It
serves humans, and it serves as a one-slot reference packing the whole turnaround when a model is
slot-starved; with slots to spare, prefer the individual approved shots.
Worked example: Nima, courier robot
- The interview yields the brief: rounded silhouette, copper shell, one cracked headlamp, canvas
satchel, cel shading; views front, back, left, three-quarter.
recommend with the brief's own words; model_schema_get confirms a true reference field (else
the next candidate); model_run the hero, jobs_wait whenever model_run returns a job_id
without assets, gate per scenario-quality-gate, iterate to pass.
collection_create "character: Nima"; collection_add_assets the hero, tagged hero.
- Three more views, baseline-plus-delta with the hero in the reference field, gated, filed on pass
with their view tags.
- Run Grid Maker on the view-tagged members (this session already holds the ids; a just-filed asset
can trail the
search index) with images as [front, back, left, three-quarter] and
columns: 4; tag the output sheet and file it.
- Weeks later, "Nima in a rainy alley":
collections_list finds character: Nima, members come
from search, the hero-tagged member and two on-model shots ride the reference field, prompt
per scenario-consistency.
Common mistakes
- Generating before interviewing: a vague brief cannot be enumerated, and the library inherits the
drift.
- Admitting ungated images: a failed image iterates per
scenario-quality-gate or never enters.
- Prompting Grid Maker or expecting it to generate: it arranges its
images input, nothing more;
generation belongs to the view runs.
- Passing the sheet as an img2img
image: composition lock (scenario-consistency); a sheet rides
a true reference slot only.
- Free-form collection names:
Nima stuff is unfindable by convention; the character: and
props: prefixes are the lookup contract.
- A different model per view: reference behavior differs per family; one schema-checked model
carries the whole identity.
1---2name: scenario-identity-library3description: "Use when a named character or prop must become a reusable identity in Scenario: creating a character or prop from a brief or interview, building its approved image collection (character: <name> or props: <name>), assembling a character sheet or turnaround, or generating new images of an established one. Triggers include create a new character, design a prop, character sheet, turnaround, more images of this character. Keywords: character design, identity, library, collection, grid."4license: MIT5metadata:6 source: https://skillmd.com/skills/scenario-labs/scenario-identity-library7---89# Scenario Identity Library1011## Overview1213A game or brand runs on named identities: characters and props that must stay recognizable in every14image that ships. This skill runs that lifecycle end to end: interview a brief, generate and gate a15hero image, file approved shots in a named collection, assemble a sheet, and pull the library back16out for new images. The per-image mechanics live in siblings and are not repeated here: connection17and the core loop in `scenario`, holding a look with references in `scenario-consistency`, pass/fail18verdicts in `scenario-quality-gate`. If a sibling skill named here is missing from your available19skills, load it with `skill_get` (Scenario MCP) or ask the user to install it20(https://github.com/scenario-labs/skills#install); unattended, proceed from tool schemas and flag21the gap.2223## Quick reference2425| Stage | Mechanism | Rule |26| ------- | -------------------------------------------------------------------------------------------- | -------------------------------- |27| Brief | interview, written as the must-not-change enumeration | before anything generates |28| Anchor | hero generation, gated to pass | uploaded art skips generation |29| Library | collection named `character: <name>` or `props: <name>` | gated images only |30| Sheet | Grid Maker assembly of approved shots | rebuild after membership changes |31| Reuse | `collections_list` prefix match, members via `search`, references per `scenario-consistency` | any later session |3233## From interview to anchor3435Interview before generating: the answers become the baseline enumeration every later image restates36(`scenario-consistency`), so a gap here becomes drift later. Ask what it is, silhouette and37proportions, palette by name or hex, two or three signature details that make it recognizable (the38cracked headlamp, the brass buckle), the rendering style, and which views the set needs: a full view39alone gives a reference slot too little detail to hold close-ups, so an identity that will be shot40close also needs a straight-on close view of the face or signature detail and an exact profile.41Non-interactive runs take what the task supplies and flag the rest. A user who already has art skips42generation: their drawing or photo is the uploaded anchor, per `scenario-consistency`.4344Pick one reference-capable model for the whole identity before the first run (`recommend`, then the45schema gate in `scenario-consistency`): every later view rides its reference field, so a model46without one disqualifies itself here, not at view three. Generate the hero, a neutral full view on a47plain background (a reference slot isolates identity best from an uncluttered anchor: nothing beyond48the brief in frame, no text), then gate it per `scenario-quality-gate` with a round cap fixed up49front (its default is three); absent that skill and the gate itself, review against the brief50enumeration, same cap. The hero is the anchor everything else cites.5152## The library is a named collection5354`collection_create` with the exact name `character: <name>` or `props: <name>`, then55`collection_add_assets` for the hero (catalog tools, write lane; the create call takes the name plus56the scope pair, and no description field). The name is the registry key: collections are not a57`search` target, so a later session finds the library by paging `collections_list` (read lane) and58matching the prefix, then pulls members with `search` (`target="assets"`,59`filters={"collection_ids": [...]}` plus a `tags` filter for the views wanted, which keeps filed60sheets out of reference pulls). Admission is the gate: only passing images enter, because every61future generation samples this pool as reference truth, and one off-model member poisons every pull62after it. Build out the views baseline-plus-delta, one `model_run` per view, gate each, file each63pass tagged with its view (`asset_add_tags`: `hero`, `back`, and so on), so a later session can tell64the members apart. A redesign (new wardrobe, haircut, age) rewrites the brief enumeration and needs65a new hero off the old (`scenario-consistency`): rebuild views and sheet from it, and66`collection_remove_assets` the superseded members before the next pull, or one pull feeds both67versions into the reference field.6869## Sheets are assembly, not generation7071Scenario's Grid Maker packs approved shots into one sheet image: its id is the fixed first-party72`model_scenario-grid-maker` (Scenario's single deterministic tool for the job, so discovery would73only re-derive it); read its schema with `model_schema_get`. At this writing it takes `images` (a74file array capped at 100; wrap even a lone asset) plus layout fields: `columns`, `rows` computed75from the count when unset, `padding`, `backgroundColor`, and a fixed `cellRatio` list (`auto`76default). It has no prompt field: order the `images` array in reading order, because the array is77the layout. The sheet is itself an asset, filed ungated (it assembles already-approved shots, it is78not a generation): tag it (`asset_add_tags`, say `sheet`), file it, and rebuild it after membership79changes rather than editing it: pull the view-tagged members only (never a bare member list, or an80old sheet becomes a grid cell), and swap the old sheet out with `collection_remove_assets`. It81serves humans, and it serves as a one-slot reference packing the whole turnaround when a model is82slot-starved; with slots to spare, prefer the individual approved shots.8384## Worked example: Nima, courier robot85861. The interview yields the brief: rounded silhouette, copper shell, one cracked headlamp, canvas87 satchel, cel shading; views front, back, left, three-quarter.882. `recommend` with the brief's own words; `model_schema_get` confirms a true reference field (else89 the next candidate); `model_run` the hero, `jobs_wait` whenever `model_run` returns a `job_id`90 without assets, gate per `scenario-quality-gate`, iterate to pass.913. `collection_create` `"character: Nima"`; `collection_add_assets` the hero, tagged `hero`.924. Three more views, baseline-plus-delta with the hero in the reference field, gated, filed on pass93 with their view tags.945. Run Grid Maker on the view-tagged members (this session already holds the ids; a just-filed asset95 can trail the `search` index) with `images` as [front, back, left, three-quarter] and96 `columns: 4`; tag the output `sheet` and file it.976. Weeks later, "Nima in a rainy alley": `collections_list` finds `character: Nima`, members come98 from `search`, the `hero`-tagged member and two on-model shots ride the reference field, prompt99 per `scenario-consistency`.100101## Common mistakes102103- Generating before interviewing: a vague brief cannot be enumerated, and the library inherits the104 drift.105- Admitting ungated images: a failed image iterates per `scenario-quality-gate` or never enters.106- Prompting Grid Maker or expecting it to generate: it arranges its `images` input, nothing more;107 generation belongs to the view runs.108- Passing the sheet as an img2img `image`: composition lock (`scenario-consistency`); a sheet rides109 a true reference slot only.110- Free-form collection names: `Nima stuff` is unfindable by convention; the `character: ` and111 `props: ` prefixes are the lookup contract.112- A different model per view: reference behavior differs per family; one schema-checked model113 carries the whole identity.