Generate the content

You've chosen a starter. Now decide what goes into it. Every starter declares a set of variables — the text and colors a user can change. Your job is to produce a value for each one based on the user's prompt.

Steps

1. Open the chosen composition's manifest.json. The variables array lists every variable: its id, type, label, and default. (The authoritative declaration also lives in index.html's data-composition-variables, but the manifest mirror is easier to read.)
2. For each variable, write a value that reflects the user's prompt:
   • string variables (headlines, taglines, feature lines, CTAs): write real copy. Keep it tight — respect any maxLength. These render as on-screen text, so punchy beats wordy. A headline is a few words, not a sentence.
   • color variables: choose hex colors (#rrggbb) that match the mood the user described. If they said "calm pastel", pick soft, low-saturation tones. If they said "bold" or "dark", go high-contrast. Make sure text color reads clearly against the background color.
   • enum variables: pick one of the declared options values.
   • number / boolean: stay within any declared min/max.
3. If the user didn't give you something for a particular variable, make a sensible choice that fits the overall brief rather than leaving the default — the default is generic placeholder content.

Writing good on-screen copy

• Lead with the benefit, not the feature name.
• One idea per line. The feature lines in app-trailer and the steps in explainer each get a single short phrase.
• Match the user's tone. A playful app and an enterprise tool should not sound the same.
• Don't invent claims the user didn't make (no fake stats, no "#1" superlatives unless they asked).

Output of this step

A plain JSON object mapping each variable id to your chosen value, e.g.:

{
  "app_name": "Strand",
  "tagline": "Notes that think with you.",
  "feature_1": "Capture ideas in a tap",
  "accent_color": "#7c9cff"
}

Write this to /.agents/workspace/output/proposed.json.

Also write the narration script

Every video gets a spoken voiceover by default. Alongside the on-screen variables, write a short narration script — what the voice says — sized to the starter's duration_seconds (roughly 2–3 spoken words per second, so a 6s starter is ~12–18 words). It should complement the on-screen copy, not just read it verbatim — set the scene, land the benefit, end on the CTA. Keep it tight; the voiceover step will trim it further if the synthesized audio runs long.

Save it to /.agents/workspace/output/voiceover_script.txt. (Skip only if the user explicitly asked for no narration.)

Then move to generate-voiceover, which synthesizes the narration, and then customize-composition, which validates everything against the composition's real schema before rendering.