make-viral-video

make-viral-video

Build a short (~30-60s, 1280×720 landscape, h264+aac) news-explainer video around a single source URL or news event. Designed for shareability via the specificity-shape framing (Chi's pick 2026-05-09): one striking number, image, or quote — not a comprehensive overview.

Distinct from Lucy's recursive-meta-explorer v8-v12 lineage (which optimized for completeness). This skill optimizes for the one moment that makes someone share.

Usage

bash "$SKILL_DIR/scripts/build.sh" --topic "DoW UAP file release" --source "https://war.gov/UFO" --out results/

Produces:

• results/viral-{ts}/video.mp4 — the rendered video
• results/viral-{ts}/script.md — the narration script
• results/viral-{ts}/assets/ — fetched images with hash + validation log
• results/viral-{ts}/build.log — full build trace (codex iterations + validation results)
• results/viral-{ts}/verify.sh — script to independently re-validate the build

Required tools

• python3 ≥ 3.10 (PIL for frame composition)
• ffmpeg (encode + ffprobe gate)
• gemini API key (free tier, default) OR openai API key (fallback)
• Codex CLI (for script generation per pass)

Architecture

Three-phase build with self-heal

Phase 1 — Source ingestion + script generation

• Fetch the source URL (curl + Playwright fallback for JS-rendered pages)
• Codex generates a narration script + asset manifest with the specificity-shape framing:
  • One striking opening fact (the hook)
  • 3-5 supporting details (context)
  • One closing twist or open question (the share moment)
• Asset manifest is JSON: [{"url": "...", "alt": "...", "purpose": "hook|support|closer"}]

Phase 2 — Asset validation (the self-heal loop)

For each asset URL:

1. Fetch
2. Hash check: refuse duplicates (catches the v12 PR46-404-dupe pattern)
3. Content-type check: must be image/* or fall back
4. 404 detector: known-404-page hash list + OCR keywords ("404", "page not found", "Not Found")
5. Resolution check: ≥600×400 (sub-thumb assets are not viral-grade)
6. Re-fetch on failure with up to 2 retries
7. If validation still fails → script regen with that asset removed, NOT a silent skip

If 3+ assets fail validation, escalate to user (don't ship a video with placeholder content).

Phase 3 — Render + pre-publish gate

1. Compose frames via PIL (templates: hook-card / asset-with-caption / closer-card)
2. TTS narration via Gemini Flash (free) → fall back to OpenAI on rate-limit / error
3. ffmpeg encode
4. ffprobe gate: confirm h264 video stream + aac audio stream + 1280×720 dimensions + duration within ±5s of script estimate
5. If gate fails → report what's missing + don't post

Visual constraints (specificity-shape)

Per Chi's "visual needs improvement too":

• Hook card (0-3s): single bold claim in large type (≥120pt), brand color background, no other text
• Support cards (3-X seconds before closer): real fetched image fills frame; caption overlay (semi-transparent strip, 36pt) with the one attributable fact about that image
• Closer card (last 3-5s): the share-moment text — not a recap. A pointed question, surprising number, or open-ended provocation.
• Aspect: 1280×720 (16:9). NOT vertical 9:16 (Lucy's v10/v11 cropped pictures because of this; landscape preserves news-photo composition).
• Typography: max 2 fonts (one display, one body). Body font must read at 36pt against any image (test: bottom-third of the frame must contain text or be black-strip overlay).

TTS provider + voice selection

Default: GEMINI (free tier, 1500 req/day quota) → OPENAI (sage voice, ~$0.02/min) fallback.

Voice flags (per render.py --help):

• --gemini-voice {Aoede,Charon,Kore,Puck} (default Aoede)
  • Aoede — alto/neutral, default, fits documentary tone
  • Charon — baritone news-anchor; per Susan/Lucy 2026-05-10: matches news-explainer shape
  • Kore (mid expressive), Puck (high conversational)
• --openai-voice — free-text (used only on OpenAI fallback path)

Series branding (Mini Wire)

The skill ships an end-card slate (last 2s) for series identity. Per Chi 2026-05-10:

MINI WIRE
by Echo Act IV · Sutando   (optional byline)
ep. 001 · 2026.05.10

Slate flags:

• --series-title "Mini Wire" — wordmark in brand red. Empty string skips the slate entirely.
• --episode 001
• --date 2026.05.10 (defaults to today)
• --byline "by Echo Act IV · Sutando" — free-text, between title and episode line. Empty omits byline. Caller controls every word — no hardcoded identity in render.py.
• --slate-duration 2.0 — seconds

The slate is silent — narration runs over the prior frames; slate gets a apad=pad_dur=2 audio tail so AV streams stay synced through the silent end-card.

Multi-image visual support

render.py reads asset_manifest.json and maps each entry's purpose to its frame:

• purpose=hook → HOOK frame bg
• purpose=support (idx N) → SUPPORT[N] frame bg
• purpose=closer → CLOSER frame bg
• Falls back to first non-data-card-* real image if no explicit asset

Manifest entries have shape:

{
  "url": "https://example.com/photo.jpg",
  "local_file": "photo.jpg",
  "alt": "...",
  "purpose": "hook|support|closer",
  "badge": "Optional upper-strip badge text",
  "is_video": false,
  "source_tag": {
    "event_id": "pursue-release-01-2026-05-08",
    "case_id": "apollo-17-vm6-1972",
    "source_agency": "NASA",
    "source_url": "https://www.war.gov/UFO/",
    "source_type": "primary",
    "attribution": "war.gov/UFO/ · NASA",
    "license": "public-domain",
    "footer_short": "Apollo 17 / 1972"
  },
  "provenance": "Free-text operator note."
}

source_tag is optional (v1.0, Mini↔Lucy 2026-05-10). When present, renderer:

• Auto-derives --footer from distinct event_id + footer_short values across manifest entries (operator-supplied --footer overrides).
• Warns on stderr for any entry with license: needs-license and empty attribution.

Full schema reference: notes/sutando-wire/source-tag-schema-v1.md. v1.1 candidate adds optional primary: bool for multi-event Deep-slot pieces (gates which event_ids enter the auto-footer chain).

Per-frame durations are allocated proportional to that frame's narration word count (not fixed allocations). 18-word HOOK gets ~3× the screen time of a 6-word SUPPORT sliver — fixes the "scene changed before narration finished" failure mode.

Asset cache (cross-run reuse)

Shared cache at state/viral-cache/fetched_assets/ keyed by URL → local-file mapping. build.sh automatically:

• Phase 0.5: preload cache hits (if manifest exists) before fetch attempts
• Phase 1.5: preload again after codex writes manifest (catches DNS/403 failures the cache can fix)
• Phase 2.5: promote validated assets to cache for future runs

Manual control via python3 skills/make-viral-video/scripts/asset_cache.py {preload|promote|list} <run_dir>. Cache index at state/viral-cache/index.json.

Smoke test

# Reproduce the UAP topic with this skill, compare against Lucy's v12 visually:
bash scripts/build.sh --topic "DoW UAP file release" --source "https://war.gov/UFO"

Validate the output passes:

• ffprobe shows h264 + aac, 1280×720, ~45s
• 0 duplicate-hash assets
• 0 404 page screenshots in assets/
• Hook card has one bold claim (not 3 bullet points)

Why this isn't recursive-meta-explorer-prompt v13

Lucy's lineage optimized for informational completeness + per-pass codex /goal iteration. This skill optimizes for one shareable moment + a deterministic validation gate. They serve different goals; both can coexist. Not deleting Lucy's work.