splicedeck: Skill for Claude Code

Cutting vertical clips

A vertical deliverable is cut from the same analysis as the master, not from a second pass. The whole design is one analysis per source shared by every deliverable, so a run producing a master and three clips spawns no more children than one producing the master alone.

Framing is owned by docs/framing.md. This file is the path through the verbs.

The order

(a finished master sheet)  →  carve  →  watch  →  caption  →  verify  →  deliver

• carve --sheet c1 --profile tall-1080 --in_ticks 0 --out_ticks 2700000 writes a second sheet from the first: body and caption entries clipped and re-based onto a timeline starting at tick 0, plus a frame lane. It spawns no child process.
• watch --source s1 --subject centre --out_shape 9:16 derives the framing path. The result is cached by source, subject and shape, so a second watch with the same three spawns nothing.
• Everything after that is the ordinary chain.

Which span becomes a clip is a human decision. carve takes the span from the caller and splicedeck does not score or rank candidates — that exclusion is deliberate (docs/roadmap.md, Not in v1). Ask, or use a span the person named.

Subject tiers, and which one actually works

--subject takes largest, centre or none.

Tier	Needs	State
centre	nothing	works in a fresh clone
largest	the reduce locator, or model weights	reduce tier works; model tier has no acquisition path
none	nothing	no tracking, static window

SUBJECT_MODEL_ABSENT does not mean vertical is impossible. It means the model tier is unavailable. SPD_SIGHT_LOCATOR selects the locator, and the geometric reduce tier needs no weights at all. Start with --subject centre when you just need a vertical export; reach for tracking when the subject actually moves.

The refusals particular to this path

Code	Means	Do this
NEEDS_FRAMING	the sheet has no frame lane for this canvas	run watch for that shape; the refusal carries the call
SOURCE_TOO_NARROW	filling 9:16 would upscale past the ceiling	report the scale factor in plain; a wider source is the only fix
WINDOW_ASPECT	a framing entry fails the exact integer cross-product against the canvas	a bug in whatever wrote the path, not a caller error
LEGIBILITY_FAIL	caption text falls under the contrast or size floor at this width	vertical is the narrowest canvas, so this bites here first
SAFE_AREA_FAIL	an element lands where a platform's own interface sits	move it; the profile declares the insets
WORD_TOO_WIDE	a single word cannot fit the line box	shorter caption text, or a smaller style