Pipeline — Demo Factory

Last updated: 2026-05-04 Source: Spec v1.1 Section 4, D2, D3, D13

---

State machine

init → captured → synopsized → scripted → rendered → done

Each state transition writes to run.json in the vault workspace. Every gate is: - Resumable — user can rerun a stage without restarting from scratch - Bypassable — --yolo flag skips all gate approvals

---

Step -1 — Scope & Angle Selection (mandatory)

Before the pipeline starts, the orchestrator skill runs scope selection: - What angle to take (feature showcase, problem/solution, CTA-focused, etc.) - Who the audience is (technical, business, mixed) - Which focus keywords to surface

Output: angle_slug + focus_keywords → embedded in run workspace name and script prompt.

---

Stage 1 — Capture

Script: scripts/capture.ts (deterministic) Gate: User reviews screenshots, prunes bad ones, re-shoots if needed.

Two modes: - Replay (default): if Products/<slug>/capture.yaml exists → Playwright replays it - Record (first run): user drives a real browser; skill records every navigation/click/screenshot → auto-generates capture.yaml for future replays

Output: CaptureManifest — array of scenes with id, label, screenshot (base64 data URI), timestamp_ms. State advance: init → captured

---

Stage 2 — Synopsis (LLM-1)

Skill: skills/synopsis/SKILL.md — Sonnet Gate: User reviews/edits synopsis; provenance shown (which vault docs were read).

Input: Products/<slug>/ vault docs + repo README/CLAUDE.md, filtered by angle_slug + focus_keywords. Doc selection: list → score → read top-N within token budget. Source code is never read. Output: synopsis.md (2–3 paragraphs, angle-filtered). State advance: captured → synopsized

---

Stage 3 — Script (LLM-2)

Skill: skills/script-writer/SKILL.md — Opus Gate: User reviews/edits script; 30–60s budget enforced (engine flags if total_duration_ms > 60000).

Input: synopsis.md + CaptureManifest Output: script.json — multi-track structured object:

{
  "total_duration_ms": 42000,
  "structure": "hook-problem-demo-cta",
  "beats": [
    {
      "id": "b1",
      "role": "hook",
      "start_ms": 0,
      "duration_ms": 8000,
      "screen_ref": "scene-01",
      "text_overlay": "Scaffold crawlers in minutes"
    },
    {
      "id": "b2",
      "role": "demo",
      "start_ms": 8000,
      "duration_ms": 18000,
      "screen_ref": "scene-02",
      "vo": "Crawler Factory automates it end-to-end"
    }
  ]
}

Three parallel tracks per beat (all optional): - text_overlay → rendered by Remotion (TextOverlayLayer) - vo → fed to 11Labs (VoiceProvider) - avatar_line → fed to Heygen (AvatarProvider)

State advance: synopsized → scripted

---

Stage 4 — Render

Script: scripts/render.ts (deterministic — Remotion CLI wrapper) Gate: User reviews video; re-render if needed.

Input: script.json + CaptureManifest (screenshots as base64) + brand profile Active tracks selected via --tracks flag: - text — silent kinetic text (no external API) - vo,text — with 11Labs audio - vo,text,avatar — full stack with Heygen avatar

Formats: --format=9:16 (default) | 16:9 | 1:1 | all Output: MP4 per format at binaries/<slug>/<run-id>/renders/<format>/demo.mp4 State advance: scripted → rendered

---

Stage 5 — Upload

Script: scripts/upload.ts (deterministic — StorageProvider)

Uploads final MP4(s) to Google Drive. Returns shareable URL. Vault Runs/INDEX.md updated with run metadata + URL. State advance: rendered → done

---

Track modes

Mode	Tracks active	APIs needed	Use case
--tracks text	TextOverlayLayer only	None	Quick test, LinkedIn silent autoplay
--tracks vo,text	VO + text overlay	ELEVENLABS_API_KEY	YouTube, narrated explainer
--tracks vo,text,avatar	VO + text + avatar	11Labs + Heygen	LinkedIn founder-voice, premium

---

Per-product workspace layout

Vault:
  Projects/Video-Editing/
    Products/<slug>/
      capture.yaml       # replay script (created on first record)
      brand.yaml         # brand override vs base profile

    Runs/<run-id>/
      run.json           # { status, product, angle_slug, run_id, ... }
      synopsis.md
      script.json
      render-manifest.json

Local disk:
  ~/AI-Development/Video_Editing/binaries/<slug>/<run-id>/
    captures/            # screenshots
    audio/               # 11Labs MP3
    renders/9-16/demo.mp4