ve-produce

name	ve-produce
description	Stages 4–5 of the video-essay pipeline. For each of 3 angles — per-angle hook research, write a short script, pre-audio judge, then audio/images/covers/captioned video. Use this skill when packages.md exists but the 3 shorts aren't all assembled yet. Runs after ve-intake, before ve-publish-shorts.

ve-produce — Stages 4–5

What this skill does

Takes packages.md (3 angles) and produces 3 fully-rendered, captioned shorts ready to publish. Per-angle pipeline: research → script → judge → audio → images → covers → video. Auto-advances from script judge pass to audio; no user checkpoint inside Stage 5 (checkpoint lives at Stage 6 in ve-publish-shorts).

Stage 4: Per-angle hook research

For each of A, B, C, run a focused hook-research pass scoped to: "what's the strongest 1.7-second hook for this specific angle?" Different question from the angle-thesis. Surfaces:

• Specific quotable numbers (Bible's "$1," "5-hour half-life" power-word style)
• Visual anomalies a viewer would pause on
• Concrete named entities (companies, people, dates)
• Counter-examples that contradict the cultural prior

Run the deep-research thorough-mode flow per references/deep_research/SKILL.md, scoped to each angle's hook. Quote the angle's Hook concept: field verbatim in the per-agent prompts. Save each output to $EPISODES_DIR/<slug>/research_hook_{A,B,C}.md — three separate files. Never overwrite research_angles.md.

Parallelizable — launch 3 deep-research invocations concurrently if budget allows.

No checkpoint here. Auto-advance to Stage 5.

Stage 5: Produce 3 shorts (parallel)

For each of A, B, C. Sub-steps run per angle; angles can be parallelized across each other.

5a. Write the short script

Read references/shorts_bible.md Part 2 Commandment 4 (Jenny Hoyos framework). Write $EPISODES_DIR/<slug>/shorts/{A,B,C}/script.md:

---
episode: <NNN>
slug: <slug>
angle: A
format: short
length_strategy: loop-maximized | watch-maximized
loop_technique: visual-match-cut | narrative-callback | audio-continuity | cliffhanger-reversal | open-question-close
target_seconds: 20    # loop-max: 11-25; watch-max: 40-60
target_words: 50      # ~150 wpm
---

# <core idea, working title>

[VISUAL: first frame — peak anomaly, sound-off legible. THIS IS THE THUMBNAIL.]

Hook line. Power word. ≤10 words. Sound-off legible.

[VISUAL: foreshadowing visual]

Foreshadow line. Plants the ending. ≤2 lines, ≤3 seconds combined.

[VISUAL: body beat 1]

Body paragraph 1. "But / therefore" causation. 5th-grade readability.

[VISUAL: body beat 2]

Body paragraph 2.

[VISUAL: payoff frame — same as first frame for visual-match-cut loop, OR ending that recontextualizes the opening]

Payoff line. NO trailing second. NO sign-off. Cut on the punchline.

Rules:

• Hook + last line written first (Hoyos)
• "But / therefore" storytelling, never "and then"
• 5th-grade readability or lower (MrBeast operates at 1st)
• No power words at the end (those belong in the hook)
• Cut the trailing second — last narration beat IS the payoff
• Script economy > factual completeness. Research gives you the full table of true facts; the script picks only the ones that push the hook forward. If a true-but-tangential fact would make a cold viewer say "wait, so…", drop it.

5a.5 Script judge (two-phase, pre-audio)

Run the full two-phase protocol from references/shorts_script_judge_rubric.md against each angle's script.md BEFORE kicking off audio or images. A Gate 10 coherence fail is free to fix here (rewrite); caught at Stage 6 it costs a full audio + image + ZapCap re-roll.

Per-angle — two sequential Agent calls:

Phase 0 (naive cold read). Agent receives ONLY the script path. No Bible, no rubric, no packages, no research. Prompt: "You are watching this video for the first time. Answer the five cold-reader questions from shorts_script_judge_rubric.md Phase 0. Do not read any other file." Save the 300–400 word report to /tmp/<slug>_<angle>_naive_read.md.

Phase A+B (rubric grading). Separate Agent. Receives: Bible, rubric, script, Phase 0 report, angle's packages.md entry. Images/video not rendered yet — the Phase 0 report is load-bearing evidence for Gate 10 and hook-vs-body coherence in Phase B. Visual gates (1, 6, 7, 8, 9) get checked "pending render"; Stage 6 re-verifies them.

Decision:

• ≥7/10 + no Gate 10 fail → proceed to 5b audio
• Gate 10 fail OR <7/10 → rewrite script, re-run both phases, do not advance
• Rewrite is cheap; audio costs real money and images cost transient credits. Never advance past 5a.5 on a failing script.

No user checkpoint — the judge is authoritative on coherence. Only escalate on <5/10 (multiple fails) or if the rewrite loop doesn't converge in two iterations.

5b. Generate audio

ve-audio <slug> --format short --angle A

Reads shorts/A/script.md, writes shorts/A/audio.mp3 + shorts/A/audio_timings.json. Uses ElevenLabs by default (voice from script frontmatter or pipeline/config.toml).

5c. Generate visuals (vertical 9:16)

ve-images <slug> --format short --angle A

Outputs into shorts/A/images/. Style preset stays the same as the rest of the channel (typically editorial_illustration).

5d. Render covers (3 per angle — Bible-grade)

The in-video first frame does scroll-stop in-feed. Covers do profile-grid + Search + browse. Different jobs, different images.

# Reels profile grid (3:4 tall, since Jan 2025)
ve-thumbnail <slug> \
    --aspect 3:4 \
    --brief "<cover_concept from packages.md>" \
    --text "<2-5 word overlay derived from hook>" \
    --text-position center-bottom --text-color yellow --text-stroke black \
    --out shorts/A/covers/cover_reels.png

# YouTube Shorts mobile profile (9:16)
ve-thumbnail <slug> \
    --aspect 9:16 \
    --brief "<cover_concept>" \
    --text "<overlay>" \
    --text-position center-bottom --text-color yellow --text-stroke black \
    --out shorts/A/covers/cover_shorts.png

# TikTok profile grid (9:16)
ve-thumbnail <slug> \
    --aspect 9:16 \
    --brief "<cover_concept>" \
    --text "<overlay>" \
    --text-position center-bottom --text-color yellow --text-stroke black \
    --out shorts/A/covers/cover_tiktok.png

cover_shorts.png and cover_tiktok.png can often be the same render (both 9:16); produce both filenames so per-platform overrides work later.

Cover overlay text rules (per references/judge_rubric.md Scroll-Stop Visual Power):

• 2–5 words, ≤20 chars
• High-contrast fill + stroke (yellow on black is the channel default)
• Legible at 120px scaled down
• Different text from in-video first-frame — cover sells the click on a profile page; first frame is already inside the video.

5e. Assemble video (with burned-in captions via ZapCap)

ve-assemble <slug> --format short --angle A

Outputs shorts/A/video.mp4 (captioned) and shorts/A/video_nocaptions.mp4 (raw). Requires ZAPCAP_API_KEY — verified by ve-doctor. On missing key, ve-assemble hard-fails before ffmpeg burn (exit 2). On a caption error mid-run, video_nocaptions.mp4 is preserved; a re-run of the same command skips ffmpeg and retries captioning only.

End state per angle: shorts/A/script.md, audio.mp3, audio_timings.json, images/, video.mp4, video_nocaptions.mp4, covers/cover_reels.png, covers/cover_shorts.png, covers/cover_tiktok.png.

Handoff to ve-publish-shorts

When all 3 angles have shorts/<angle>/video.mp4 + 3 covers, hand off to ve-publish-shorts — it runs Stage 6 (visual gate) then Stage 7 (publish to YT/IG/TikTok).