Scenario Formats skill

Scenario Formats

Overview

Format adaptation is derivation, not regeneration: re-prompting the concept per placement makes five cousins, not five formats of one approved master. The failure modes are picking the wrong operation (a resize where canvas needed inventing, a crop that beheads the subject) and deriving in the wrong order. The canvas tools' exact contracts live in scenario-image-editing; this skill is the decision ladder and the order of operations around them, with per-placement ratios, pixel targets, and safe zones in references/placement-specs.md. A video master climbs the same ladder: video resize versus generative reframe lives in scenario-video-editing, and far ratios recompose per scenario-video. Connection and the core loop: see the scenario skill. Lettering per format: scenario-text-overlay. 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.

The decision ladder

Per target format, take the first rung that fits. Each rung is a model_run on a model found with search (target="models", public=true), never a dedicated MCP tool of its own:

The target needs	Operation
Same ratio, other pixels	Resize: exact, scales pixels, invents nothing
A tighter ratio, edges expendable	Resize with fit: "cover": exact, center-crops the overflow; only when the subject clears the cut
More canvas, composition intact	Expand or uncrop: exact pixels, fills outward
Another ratio, recomposed around the subject	Generative reframe; the text-protecting variant when lettering or marks are on the plate
A far ratio or a new visual hierarchy	Recompose natively: a fresh run per scenario-image with the master as reference (scenario-consistency), composed for that ratio

The last rung applies more often than it looks: 16:9 key art seldom survives to 9:16 by any amount of outpainting, which is why the director siblings (scenario-video-ads) design each delivery ratio as its own composition.

Order of operations

1. Approve one master first, at the largest clean size available, on a text-free plate (unattended, the task's named master or the brief's own description stands in for the approval).
2. Derive every format from that master, never from another derivative: derivative-of-derivative compounds re-rendering artifacts.
3. Reframe and expand before grading and grain: generative canvas work re-renders the image, so finishing passes applied first come back partly reinterpreted (scenario-image-editing holds the pipeline order). A video reframe also precedes any upscale: it renders at its own resolution tier whatever the input carries (1080p at most at authoring time), so an upscale ahead of it is discarded; upscale the reframed clip to delivery size.
4. Verify by measurement before finishing: reframe ratio enums are approximate (a 4:5 request came back 29:36 at authoring time), so check output dimensions, land exact with a resize, and batch-check that subjects survived with asset_analyze (scenario-asset-analysis).
5. Letter last: re-overlay lettering per format with scenario-text-overlay, sized to each final canvas, since type scaled by a resize turns soft and each placement has its own safe area. Keep critical content and text away from the edges platform UI covers: each placement's target and safe zone is in references/placement-specs.md as an authoring-time value, so confirm contractual specs with the user (unattended, task instructions win over the reference, and the delivery report states which was used).

Worked example: key art to story, feed, and thumbnail

1. Master: 16:9 text-free key art, approved.
2. 1:1 feed: generative reframe with a prompt naming what must stay ("the knight centered, both banners visible"), then a resize pass to the exact deliverable pixels.
3. 9:16 story: a far ratio, so recompose natively: one run per scenario-image with the master as reference and a 9:16 composition clause; the master holds palette and subject on-look.
4. Thumbnail: readability rules the crop, so reframe tight on the face or product, land exact with a resize, then asset_display the result and judge it small: thumbnails are seen at a tenth of their size.
5. Verify dimensions against the deliverable list and batch-check that subjects survived with one asset_analyze pass, then overlay the title card per format onto each final canvas and file the set in a collection.

Common mistakes

• Re-prompting per format: five generations of one prompt are five different images; derive from the master.
• Stretching into a new ratio with a resize (fit: "stretch"): it distorts. A ratio change is fit: "cover" when the edges are expendable, else reframe, expand, or recomposition.
• Grading and grain before reframing: the order is reshape, then grade, then texture.
• Deriving from the lettered master: generative operations re-render type; keep a text-free plate and re-overlay per format.
• Trusting a ratio enum: measure the output and resize to exact.
• Center-cropping toward 9:16 because it is cheap: it is also how subjects lose their heads; the ladder exists to be climbed.
• Skipping dry_run on reframes: generative canvas work is the expensive step in the chain, tens of times an effect's price (scenario-image-editing).