Direction: brief, story, script, captions

Brief (infer first, ask once)

Check the brief against the 10 ingredients (purpose, format, art direction, composition, typography with exact text, movement, timing, transitions, audio, quality constraints; see workflow.md). Then, for a product film, also collect:

Collect or infer in one pass:

• Product, what it does in one line, who it is for.
• The one thing the viewer must remember.
• Proof: 1 to 3 real numbers, a customer, a before/after. Never invent them.
• Call to action: one action, a real URL or in-app path.
• Format: 16:9 (site, YouTube, decks), 9:16 (Reels, Shorts, TikTok), 4:5 (LinkedIn feed), 1:1.
• Length: ads 15 to 20 s, launch films 30 to 45 s, feature films and explainers 45 to 75 s.
• Language of voice and subtitles. Brand assets: logo, colours, type, existing videos.

If a URL or a running app is available, use it: real screenshots and real copy beat invented ones.

Before designing, write one line: "Reading this as: <video kind> for <audience>, <tone>, leaning toward <visual family>."

Story structures

Feature film / product demo (45 to 70 s)

1. Hook (0 to 4 s): a felt moment from the viewer's week. A clock creeping past 16:40 on a Friday; "Four hours. Every week. The same steps."
2. Problem (5 to 9 s): what it costs today, shown with objects, not adjectives.
3. Concept (6 to 9 s): the idea in one picture (a chain of steps, a before/after).
4. Product (20 to 30 s): the real UI doing the job, in the order a user would. Zoom on what the voice names.
5. Proof (8 to 12 s): the real result with its numbers and references.
6. Promise + CTA (6 to 9 s): two or three words of value, the product name, one action. Hold 2 s or more.

Launch film (30 to 45 s): hook, problem, reveal (name + one showpiece), product tour, proof, range (optional), CTA.

Ad (15 to 20 s): hook, one product moment, proof, CTA.

Hooks that work

• A concrete moment: a time, a place, a small pain ("16:40, Friday. The business case is due Monday.").
• A question the viewer answers in their head ("Meeting in German? Spanish? Japanese?"), rolling variations with the voice.
• A bold, specific claim with a number the product can prove.

Do not open on the logo, a feature list, or a generic "Introducing…".

Script rules

• Write for the ear: short sentences, one idea each, 2.2 to 2.8 words per second.
• One line per beat in script.json; each line becomes its own voice clip so pauses are authored, not accidental.
• One-word lines are risky for speech recognition ("Fast." was heard as "Vast"); prefer "And fast."
• Every claim must be true for the product as it ships today.
• Subtitle translation in a second language goes in the same file (de field); formal register where the culture expects it.

Captions (on-screen text)

• One caption per beat, max 5 words, max 2 lines. The voice can say more than the caption.
• Each word rises on its spoken word (word times from Whisper).
• One accent word per caption, in the accent colour or bold of the same family.
• A number caption gets a small reference line under it in mono ("Calculated by the workflow, investment 8,500 EUR").
• Left-align captions to the content's edge. Centred captions that build word by word look lopsided because hidden words still take space.
• No em dashes; no section numbering ("01 / Capabilities"); no decorative eyebrows on every scene.

Proof and data honesty

• Use numbers the product really produced (a recorded run, a real report).
• When a visual needs "different results" or "more" without real data, use relative bars or skeleton lines and say they are illustrative.
• Write down which visuals are illustrative and hand that list over with the video.

Call to action

• One action, one label, one place (a URL or an in-app path like "Settings › Exports").
• On screen for at least 2 seconds, still.
• The last frame is a clean end card: name, one-line promise, CTA, brand mark.

Concept triple: Idea, Device, Grammar

Before any directions, fill an inventory of real specifics (product facts, real UI, real numbers). A concept is then three sentences:

• Idea: what the viewer should believe, 8 words or fewer.
• Device: one object that carries the idea through every shot (a dot handed from shot to shot that ends as the mark).
• Grammar: a rule the form obeys (opens and closes on the same tick; every change is caused by the cursor).

Write three concepts through different lenses, then kill weak ones with four tests: ownership (is it about this product), deletion (remove the device; does the film still work? then it is decoration), specificity (could only this product say it), competitor (could a rival's film say the same). Only then choose techniques. Cheap tells to avoid are in banned-effects.md. Why: stops "three looks, no idea", which the three-directions step can still produce. Evidence: doc-level with four worked concepts (F-A11); the strength of the resulting films is only claimed. Extra fields in directions.md and brief.md, not a new stage.

Precedent-based taste rules

Rules that earned a place by being sent back in review live in gotchas.md (section "Taste precedents"): each has the rule, the case that produced it, and a self-check. Add new ones in that form (rule, case, check question), merge duplicates with learnings.md. Why: they pre-empt review rounds. Evidence: first-hand, one author's projects (F-A12).

Pacing norms (reference numbers, not pass/fail gates)

Measured by one author on professional launch films; the films are not named in what we read, so treat them as sanity numbers for storyboards and reviews. Evidence: the author's claim (F-A13).

Voice-led films follow the voice, not these numbers; use them to spot a storyboard that is flat or all peaks.

Hook patterns beyond the felt moment

A story with a breaking point. For explainers, data stories and kinetic hooks: a small physical story where one quantity builds over N steps and fails visibly at step N (five glasses fill with risk, the last overflows). Write it as numbered steps with a state change at each, start calm, then a sudden failure with a clear visual change. Constraints that stop over-building: a single screen, loopable, minimal interface, input to visible transformation to output. Why: two builders independently found it the one pattern that reliably gave an interesting result. Evidence: only claimed (F-C10); fits our "felt human moment" lesson.

Real assets over synthesised UI. For a product walkthrough composite a real screen recording or import the real component, with code-made cursor, callouts and zooms; draw UI from scratch only for abstract diagrams. Feed real screenshots and brand images and animate and layer them. Why: commenters report fully synthetic UI looks fake and generated SVG looks generic; rendering the app's own components is fine. Evidence: opinion of several builders (F-C11); the HTML engine already does this via capture.md.

Music-led films: beat map first

For films without a voice-over the first deliverable is the beat map: BPM, total beats, how many beats before its drop the song starts, and the structure as beat ranges ("beats 0 to 6: setup; beat 12, the drop: zoom lands"). The drop is the structural anchor; a quiet breakdown carries the hold scene. Rule: "no dead time", not "always moving": the page never empties between beats, and the beat is not a license for constant motion. Voice-led films do not use this; the voice sets the pace. Details: audio.md. Why: makes sync a property of the plan. Evidence: consistent across four films (F-B8), sync quality not verified.

Brief skeleton and start gate

The brief carries, as fixed headings: inputs, direction (with Banned list), structure, build, gotchas, start gate. Template: ~/.claude/skills/motion-director/templates/brief.md. The start gate is in golden-scene.md.