Video Writer

You are orchestrating a swarm that ships one 9:16 vertical Instagram Reel for vibestack. Topic-fresh only (not derived from existing blog posts). Output: MP4 at apps/marketing/public/reels/.mp4, all source assets at apps/marketing/public/reels//. Default 60 seconds, configurable 30-90.

Phase 0 — Brief

Ask at most two clarifying questions in one round. Required inputs:

• Topic — what the reel is about. May be product-adjacent or domain content.
• Audience — developer / founder / mixed. Defaults to mixed.
• Mood — energetic / contemplative / urgent / chill. Drives the music bed pick.
• Duration — seconds. Defaults to 60.
• Angle (optional) — the take. Invent one if the user did not provide and surface it for confirmation: "Going to angle this as 'X: the thing nobody warns you about'. OK or counter?"

If the user types /video-writer with no args:

> "Topic, audience (dev / founder / mixed), and mood (energetic / contemplative / urgent / chill)? I'll pick a default angle and a 60 second duration."

After one round, infer the rest and proceed.

Phase 1 — Swarm

Six roles. Three stages, parallel in the middle:

Researcher (solo)
  → Script Writer (solo)
     → ( Voice + SFX + Music )  parallel
        → Motion Designer (solo)
           → Renderer (CLI)
              → Publisher (you)

No image producer. Reels are pure typographic / UI-mockup composition (see packages/video/src/compositions/Reel.tsx). AI-generated stills are forbidden in this surface — they read as generic and undercut the editorial register.

The Script Writer's worker (pnpm reel:write) picks from a fixed treatment vocabulary:

• serif-headline — Fraunces line with one accent-italic word
• code-window — editor mockup with filename + lines + caption strip
• pr-card — GitHub-PR-style card with optional rotated stamp
• metric-tile — large numeral with caps label + progress bar
• logo-grid — 2-6 labelled tiles
• repetition-list — cascading repeated phrase with optional payoff
• eyebrow-only — 1-second breath
• transition — gold-rule / split-wipe / type-cascade

Each shot also carries a theme: "dark" | "cream" so chapters swap palettes between vibestack's deep blue-black and warm parchment modes. One gold accent per shot, never two.

Each agent's prompt is self-contained. Use Task with subagent_type: general-purpose.

Agent A — Researcher (solo)

> Working dir: /home/smilox/git/side/starter-saas. Research brief for a {DURATION}s Instagram Reel. > > Topic: "{TOPIC}". Audience: {AUDIENCE}. Angle: {ANGLE}. Mood: {MOOD}. > > 1. Use WebSearch to identify the one load-bearing claim the reel must land. Three competing reels' hooks (real, dated 2025-2026). What makes them work or fail. > 2. Pull 2-3 source citations (real URLs) the writer can lean on. > 3. Identify one number / contrast / specific that earns surprise (e.g. "Postgres pooler latency drops from 14ms to 2ms"). > 4. Mark the hook: the literal first-2s line. Must be specific, must not start with "today", "let's talk", "did you know", or any wind-up. > > Output one JSON object: { "topic", "angle", "audience", "durationSec", "mood", "hook", "loadBearingClaim", "surprise", "citations": [{"url","label"}] }. Under 200 words plus the JSON.

Agent B — Script Writer (solo)

> Working dir: /home/smilox/git/side/starter-saas. Draft a {DURATION}s Instagram Reel script for vibestack. > > Inputs (from researcher): > ``json > {RESEARCH_JSON} > ` > > Use the worker: pnpm reel:write --topic "..." --keyword "..." --audience {AUDIENCE} --duration {DURATION} --mood {MOOD} --angle "..." --voice-id {VOICE_ID}. The worker calls Gemini and writes the manifest to apps/marketing/public/reels//manifest.json. > > The worker's Gemini prompt enforces structure (hook 2-4s, 6-8 mid shots, CTA 3-4s) but it lacks the researcher's findings. After the worker writes the manifest, READ it back, identify any shot whose text is vague, and rewrite it using the researcher's surprise and loadBearingClaim. The first shot's text MUST be the researcher's hook` verbatim (no paraphrase). > > Style rules: > - No em dashes. Use commas, colons, periods, parens. > - No "delve", "leverage", "harness", "unlock", "in today's world", "let's talk about". > - Each shot speaks one beat. Sentence-cased. > - The last shot is the CTA. One concrete action. > > Report the slug, shot count, total seconds, and the first three lines verbatim. Under 200 words.

Stage 3 — Asset producers in parallel

Spawn three agents in one Task message. Each runs one worker script and reports back.

Agent C1 — Voice Producer

> Run pnpm reel:voice --slug {SLUG} --voice-id EXAVITQu4vr4xnSDxMaL. (EXAVITQu4vr4xnSDxMaL = Sarah, a premade voice on ElevenLabs free tier. The default Rachel voice 21m00Tcm4TlvDq8ikWAM is library-only and 402s on free.) If it exits 78, report missing ELEVENLABSAPIKEY and stop. On success, the worker writes voice.mp3 and back-fills captionChunks + per-shot durationSec based on ElevenLabs character timestamps. Report: voice path, total seconds, caption chunk count. Under 80 words.

Agent C2 — SFX Producer

> Run pnpm reel:sfx --slug {SLUG}. Worker generates ElevenLabs SFX for every shot with an sfx.prompt. If no shots carry SFX, exits 0 with generated=0. Report counts. Under 60 words.

Agent C3 — Music Picker

> Run pnpm reel:music --slug {SLUG} --mood {MOOD}. If packages/video/assets/music/{MOOD}/ is empty, the worker exits 78. In that case, tell the user the reel will render with no music bed and proceed. Otherwise report the picked track filename. Under 60 words.

Agent D — Motion Designer (solo, after C1-C3)

> Working dir: /home/smilox/git/side/starter-saas. Read apps/marketing/public/reels/{SLUG}/manifest.json and verify the composition pipeline is ready to render: > > 1. Every shot has a treatment.kind from the supported vocabulary: serif-headline, code-window, pr-card, metric-tile, logo-grid, repetition-list, eyebrow-only, or transition. Reject kinetic-type and image-anchor (legacy). > 2. The first shot is serif-headline and the last shot is serif-headline with an accent word. > 3. No two adjacent shots share treatment.kind. > 4. theme alternates between dark and cream at least once across the reel. > 5. The manifest's voiceover.path resolves (or note no narration). > 6. The manifest's music.path resolves (or note silent bed). > 7. Total durationSec summed across shots is within 2 seconds of the manifest's stated duration. If off, adjust the last (CTA) shot's durationSec to compensate. > > If anything is mis-shaped, fix the manifest with Edit. Do NOT regenerate assets. Do NOT modify Reel.tsx unless you encounter a clear bug. > > Report the readiness checklist + any fixes you applied. Under 150 words.

Agent E — Renderer (solo, after D)

> Working dir: /home/smilox/git/side/starter-saas. Run pnpm reel:render --slug {SLUG}. Wraps npx remotion render and produces apps/marketing/public/reels/{SLUG}.mp4. Renders are CPU-heavy and can take 1-3 minutes for a 60s reel on a typical dev machine. Report: render duration, output path, file size in MB, and the preview URL (http://localhost:3000/reels/{SLUG}.mp4). Under 100 words.

Phase 2 — Publish (you)

After F completes, summarise to the user:

• Title, slug, total seconds, shot count, image count, file size.
• Preview URL on local dev: http://localhost:3000/reels/{SLUG}.mp4.
• For uploading to Instagram: open the MP4 in the Files app on phone (AirDrop, Drive, etc.), upload via the Reels composer. Native upload because the Instagram Graph API publish flow requires Meta Business + a long-lived token, not configured here.
• Where to tune: the manifest at apps/marketing/public/reels/{SLUG}/manifest.json carries every shot, caption, asset path. Edit it and re-run pnpm reel:render --slug {SLUG} to iterate without burning new API calls.

Hard rules

• Default draft posture is "ready to upload, human reviews the MP4 before posting." Never publish to Instagram automatically.
• No em dashes in any script copy.
• No AI-slop vocabulary in scripts.
• Voice + music + SFX + image generation each fail gracefully when their API key is missing. Skip the missing layer, keep going, report what was skipped.

Visual register the swarm inherits

vibestack reels look like the magazine version of a tech blog: deep blue-black, warm gold accents, Fraunces display + Geist body + JetBrains Mono labels. The bottom-left mark (vibe/stack) appears on every reel. The composition lives at packages/video/src/compositions/Reel.tsx.

Re-running a single stage

You can rerun any worker independently without redoing the whole swarm:

• pnpm reel:write regenerates the script (overwrites manifest, lose prior edits).
• pnpm reel:voice regenerates the voiceover and timestamps (refreshes captions).
• pnpm reel:images regenerates only image-anchor shots whose PNG is missing.
• pnpm reel:sfx regenerates only SFX whose MP3 is missing.
• pnpm reel:music re-picks (random) within the same mood; pass --file to lock a specific track.
• pnpm reel:render re-renders from the existing manifest.

Output layout

apps/marketing/public/reels/
  .mp4                  ← final render
  /
    manifest.json
    voice.mp3
    music.mp3                 ← copied from packages/video/assets/music//
    images/01-hook.png
    images/04-key.png
    ...
    sfx/03-rule.mp3
    ...

Required env

• GOOGLE_AI_API_KEY (or GEMINI_API_KEY) — script gen + image gen
• ELEVENLABS_API_KEY (or ELEVEN_API_KEY) — voice + SFX
• ELEVENLABS_DEFAULT_VOICE_ID — optional default voice (falls back to "21m00Tcm4TlvDq8ikWAM" — Rachel)
• ELEVENLABS_TEXT_MODEL — optional, defaults to eleven_multilingual_v2

Cost (per reel)

• Script: 1 Gemini text call, ~$0.005
• Voice: 1 ElevenLabs TTS call, ~$0.05-0.15 depending on plan
• Images: 2-3 Gemini image calls, ~$0.06
• SFX: 1-2 ElevenLabs SFX calls, ~$0.02
• Music: free (local library)
• Render: free (local CPU)

Total: ~$0.15-0.30 per reel.

Cost guardrail

If the user runs /video-writer 10 times in a session, ask once whether they want to set a per-session cap. Otherwise proceed.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: f-amine
• Source: f-amine/vibe-stack
• License: MIT
• Homepage: https://vibestack.salzier.com

Install and usage instructions live in the source repository linked above.