SuperAPI AI Directory

storyboard

Design & creation

View repository

Updated 10-28 21:24

About this skill

Storyboard

The cheap gate before motion. A note here costs minutes; after production it costs an hour. The poses are real artefacts (real tokens, real copy, the real logo), so nothing is rebuilt from a description later.

A revision to an approved or delivered video (a word, a colour, a re-time) does not reopen the storyboard: edit the composition, re-render, run video-gate, look at the changed frames, deliver.

The unit is a change, not a slide. A beat has an intention and an observable result; entrance, development, read and handoff are its phases. A pose is one of the two or three moments a beat needs, not the scene.

0. Intake

State sensible defaults and proceed; ask only what blocks (at most two questions). The intake heads projects/<slug>/storyboard.md:

  • Core message: the one thing the viewer leaves with.
  • Audience: who is watching and what they already know.
  • Claims ledger: every assertion the film makes, its source, and whether it is verified, measured or illustrative. Keep provenance off the frame.
  • Close: the product name and the ask. Verify any URL or package name before it goes on an end card; links belong in the post.
  • Channel, aspect, duration: 16:9 (1920×1080), 9:16 (1080×1920) or 1:1, from the brand pack's formats (/workspace/brand/brand.toml). Default 16:9 and 20-30 s. A vertical version is a reflow of the layout, never the wide frame cropped.
  • References: a film the person handed you outranks everything; measure it with film-breakdown.
  • Must-includes: required copy, marks, real product surfaces, things to avoid.

Exact requirements that must survive revisions go in projects/<slug>/artifact-contract.json (node /workspace/tools/artifact-contract.mjs --contract ... --phase plan). When supplied artwork must move, start from templates/layer-manifest.json: each moving part is an isolated transparent layer with bounds, anchors and attachments (check it with --layers). A flat image is not animation-ready; ask for the separated layers rather than faking them with guessed crops.

1. The argument, then the beats

Write the argument as prose first and read the beats aloud as one sentence: if each is a self-contained label, you have a deck. Find the turn, the one concrete thing the film pivots on. Then one row per beat and one row per cut: the fields are in references/beats-and-handoffs.md. The two fields that carry the film: the change (what develops after the entrance; a beat whose every tween is an entrance is a slide) and the carrier (what crosses each cut; where nothing does, say what the semantic handoff is). List the attention demands per frame rather than asserting a number: more than two is two beats.

Sound is part of the board when the brief has sound: the music arc and each beat's sound accent, with a plan.json started beside it (sound). With voiceover, generate the lines now: the film is cut to the voiceover, and timing locks at animatic approval. A silent deliverable says sound: none.

2. Poses, shot from real HTML

Per beat, one self-contained HTML page at the target size, styled from /workspace/brand/tokens.css (var(--brand-…), never a typed hex or family), with real copy where it carries meaning, the real mark from /workspace/brand/logos/, and the real product surface where the beat says "this is the product". Give the elements the build will animate stable ids (scene-prefixed, #s3-card; a carrier unprefixed, #carrier-issue), with explicit top/left/width/height and no self-running animation.

How many poses: a static beat needs one; a beat with a change needs two (before and after); a cut with a carrier needs three (outgoing, midpoint, incoming), and the midpoint matters most: a perfect endpoint can hide a broken journey.

node /workspace/tools/html2png.mjs projects/<slug>/board/beat-03a.html --out projects/<slug>/board/poses \
  --name beat-03a --size 1920x1080 --brand-fonts
node /workspace/tools/brand-lint.mjs projects/<slug>/board/beat-03a.html

When the scene already exists as a HyperFrames composition, hyperframes snapshot --at <t1>,<t2> shoots the poses from it instead. A look plate for a photographic or illustrated beat can come from image-generation (no text in it; type is set in HTML on top). Emit the poses and later the film from one generator (build.mjs with the beat table, copy and shared CSS), so the delivered composition is provably the approved artwork and a re-time is one number per beat.

3. One contact sheet

node /workspace/tools/review-thumb.mjs --sheet projects/<slug>/board/poses/*.png \
  --out projects/<slug>/board

Look at it yourself first against the craft review in the reference (squint: one or two things per frame; work, not decoration; a concept, not brand primitives; one accent; a named change per beat; a carrier or a named handoff per cut). Redesign only the beats that fail.

If someone cannot picture a framing, draw it instead of writing more prose: a skeleton sheet with the real copy at its intended scale or a labelled placeholder box per cell. Ship one story with the reasoning visible; three options that share a flaw read as three more misses.

When Paper is connected and people want to nudge the board, draft it there as well (paper-design); the thread still gets the sheet.

4. Sign-off, then build

Present the argument, the beat sheet, the handoff sheet and the contact sheet (one upload_files), then wait for a clear go or specific redlines. Someone who already said "storyboard it and build it" has given the go: present the sheet and keep going.

The build assembles the approved poses rather than re-authoring them (motion-direction, hyperframes): namespace each pose's styles to its scene, mount the scenes in one host composition with the carrier on its own track, then add development, camera and micro-interaction around the named ids. Then a draft render, a review through every cut, video-gate at zero errors, the final render once the look is confirmed, and the score (sound).