MervinPraison daily-single-video
Pipeline order + QA gates: repo root AGENTS.md
Social comparison variant: examples/videos/anthropic-claude-fable-5-social-comparison/
Single-topic walkthrough (one launch story, ~5–9 min). Reference run: examples/videos/anthropic-claude-fable-5-mythos-5/
June roundup pattern: examples/videos/june-2026-ai-roundup/
Handoff: create-news research/<slug>/HANDOFF.md
Do NOT use
segment-video-roundup/scripts/bootstrap-project.sh(megapost layout)- Megapost-only validators (
required_assetscatalogue gates) — hook_montage is required for daily_single
Pipeline phases
| Phase | Status | What |
|---|---|---|
| 1 — Script + B-roll + hook montage | Current | daily_single CLI: scripts → VO → ffmpeg beat assembly with phrase-synced hook heroes |
| 2 — HeyGen decks | When credits return | Per-beat + hook avatar_media_3 like June roundup (align-cues → yaml → build) |
| 3 — Publish | Shared | .cursor/skills/mer-vin-article-video-upload/SKILL.md |
Phase 1 and Phase 2 share the same script contract below so Phase 2 is a drop-in upgrade, not a rewrite.
YouTube script contract (June-aligned)
Every video follows this spoken structure. Optimise for retention: hook in the first sentence, overview before depth, explicit handoff into content.
00-hook — three parts (~55–70 words)
- Attention hook — one punchy sentence (why watch now). Label line
Hook:is stripped at TTS only. - Overview — comma-separated roll-call of topics (one phrase → one hero in montage). Plain language, non-developer friendly.
- Bridge — end with exactly: Let's get started.
June reference (segments/00-hook/script.md):
Fifteen stories in this roundup: Nemotron 3 Ultra, Gemma 4 … Meta Muse Spark on watch.
Now we are going to walk through every one in detail. Let's get started.
Single-topic example (anthropic-claude-fable-5-mythos-5/segments/00-hook/script.md):
Hook: Anthropic just released Claude Fable five — if AI is part of your work, this launch matters.
In the next five minutes: what most teams actually get, Stripe's fifty-million-line proof, benchmark scores that matter, safety without dead ends, and the website-versus-developer mistake teams keep making.
Let's get started.
Plain language: .cursor/rules/daily-single-plain-language.mdc — no Mythos before beat 02; automated pytest tests/test_audience_language.py.
Hook attention (first 5s): record-canonical-scroll → blog scroll video; validate-hook-attention exports one JPEG per second.
Hook montage (required): each comma clause in the overview maps to a hero asset during cue 2 (~2s per slide). Assembly: hook_montage.py → assemble.py _hook_montage(). Validators: hook_montage in validate-sync --runs 3 (≥5 distinct heroes, no launch-only overview).
01-cold-open … 10-* — beats
- Start after Let's get started — first beat is real content, not another intro.
- Plain language for a general YouTube audience; explain jargon when unavoidable (see
.cursor/rules/daily-single-plain-language.mdc). - ~2–4 sentences per beat; one sentence = one SRT cue (see
video-script-captionsskill).
99-outro — June CTA (no spoken mer.vin)
I hope you liked this video. If it helped, like, share, and subscribe for the latest AI updates. Thanks for watching.
Do not speak mer.vin URLs, integration tables, or technical recaps in the outro.
Phase 1 commands
conda activate test # or windsurf
cd /path/to/praisonaippt
PROJECT=examples/videos/<slug>
| Step | Command |
|---|---|
| Scripts from beat map | python -m praisonaippt.daily_single --project $PROJECT write-scripts |
| Lock hook/outro manually | Edit segments/00-hook/script.md, segments/99-outro/script.md |
| Sync canonical assets (HD) | python -m praisonaippt.daily_single --project $PROJECT sync-assets |
| Voice-over | python -m praisonaippt.daily_single --project $PROJECT synthesise-vo |
| Hook/outro HeyGen | python -m praisonaippt.daily_single --project $PROJECT bookend-media 00-hook 99-outro |
| Captions | python -m praisonaippt.daily_single --project $PROJECT build-captions |
| Assemble B-roll | python -m praisonaippt.daily_single --project $PROJECT assemble-beats |
| Transcript ↔ visual map | python -m praisonaippt.daily_single --project $PROJECT validate-display |
| Spoken ↔ visual gate | python -m praisonaippt.daily_single --project $PROJECT validate-spoken-visual |
| Pixel visual audit (every 5s) | python -m praisonaippt.daily_single --project $PROJECT audit-visual --interval 5 |
| Robust sync test (×3) | python -m praisonaippt.daily_single --project $PROJECT validate-sync --runs 3 |
| Full gate | python -m praisonaippt.daily_single --project $PROJECT validate-all |
Standard pipeline order: sync-assets → synthesise-vo → bookend-media → build-captions → assemble-beats → validate-display → validate-spoken-visual → audit-visual → validate-sync --runs 3 → validate-all
Cue-aligned rebuild (beat-06, beat-01): see .cursor/skills/daily-single-video-pipeline/spoken-visual-sync.md
Step-by-step with modular QA gates: use .cursor/skills/daily-single-video-pipeline/SKILL.md (validate-qa --when pre_build|pre_assemble|post_vo|post_build).
Hook montage artefact: segments/00-hook/hook_montage.json (phrase → hero plan; written on assemble).
Visual audit artefact: merge/visual_audit_report.json + merge/visual_audit_frames/ (JPEG every 5s + cue midpoints).
Skip existing media unless scripts changed: add --skip-existing to synthesise-vo / bookend-media.
After any script edit: re-run VO → bookends (if hook/outro) → build-captions → assemble-beats → validate-display → validate-spoken-visual.
Before first assemble (and after handoff changes): run sync-assets — crawls canonical news URL images and re-downloads YouTube demos at 720p+ (not 360p progressive). Patches beat-map.json with carousel clips and biology charts.
create-news upstream
| Step | Where |
|---|---|
| Motion clip analysis | analyse_motion_clips.py --slug <slug> |
| Beat map | build_beat_map.py --slug <slug> |
| Cards / slides | generate_daily_cards.py --slug <slug> |
| Master script | research/<slug>/video-script.md (beats 1–10; hook/outro live in project segments/) |
Bootstrap: scripts/bootstrap-daily-single.sh
Outputs
| File | Purpose |
|---|---|
merge/final.mp4 |
Loudnorm final |
merge/final.srt |
Script-aligned captions |
merge/display_sync_report.json |
Each cue → on-screen visual |
merge/sync_validation_report.json |
3-run robust gate + YouTube quality |
merge/asset_sync_report.json |
Canonical crawl + HD video inventory |
Captions rules: .cursor/skills/video-script-captions/SKILL.md
Phase 2 roadmap (HeyGen + June template)
When HeyGen credits return, upgrade hook segment only first (same script contract):
PROJECT=examples/videos/<slug>
# Hook segment — June compositor path
python -m praisonaippt segment-video sync-media --project $PROJECT --segment 00-hook
python -m praisonaippt segment-video align-cues --project $PROJECT --force 00-hook
python -m praisonaippt segment-video yaml --project $PROJECT --force 00-hook
python -m praisonaippt segment-video build --project $PROJECT --force 00-hook
python -m praisonaippt segment-video validate-hook --project $PROJECT --segment 00-hook
# Re-merge: replace beats/00-hook.mp4 with segment build, then assemble-beats
Phase 1.5 hook_montage.json becomes input to sync-media hero selection. Full per-beat HeyGen:
- Keep
segments/*/script.mdunchanged (same contract). - Add per-segment
segment.yamlusingexamples/heygen-50590-video-audio-heygen-images.yamlpattern. - Run segment-video-roundup stages per beat:
sync-media→align-cues→yaml→build→normalize-audio→merge. - Continuous HeyGen under hook montage slides; outro = HeyGen only.
See .cursor/skills/segment-video-roundup/SKILL.md and reference.md.
Quality gates before publish
-
sync-assetspass — all handoff images + videos ≥720p from canonical page - Hook has three parts: attention → comma-separated overview → Let's get started
- Hook overview shows ≥5 distinct heroes (not single launch B-roll for whole overview)
- Hook attention uses first hero slide (not vintage launch B-roll at t=0)
-
audit-visualpass — no generic/off-topic frames (merge/visual_audit_report.json) - Beat 1 starts content (no duplicate intro)
- Outro = June subscribe CTA; no mer.vin spoken
-
validate-sync --runs 3pass (script lock + hook + hook_montage + image + YouTube quality, idempotent) -
build-captions+validate-display+validate-spoken-visualpass -
merge/spoken_visual_sync_report.json→ok: true - Plain language — readable without developer context
Testing (spoken ↔ video ↔ image)
Three layers (mirrors June roundup, adapted for Phase 1 B-roll):
| Layer | What it checks | Command / test |
|---|---|---|
| Caption lock | SRT text == locked script.md (not Whisper) |
validate-sync → caption_script_lock |
| Spoken↔visual gate | Windows, charts, transitions, coverage | validate-spoken-visual → spoken_visual_sync_report.json |
| Image mapping | Each cue midpoint → correct asset file + keyword score ≥0.35 | validate-display → display_sync_report.json |
| Hook structure | Cues 1–3 = attention → overview → Let's get started | validate-sync → hook_structure |
| Hook montage | Overview cue → ≥5 distinct heroes; alignment ≥0.45; not launch-only | validate-sync → hook_montage |
| Visual audit | Pixel match + topic score every 5s; blocks generic B-roll | audit-visual → visual_audit_report.json |
| Idempotency | Same report on 3 consecutive runs | validate-sync --runs 3 |
| YouTube quality | Compelling hook, plain language, pacing, outro CTA, alignment depth | validate-sync → youtube_quality |
pytest tests/test_daily_single_display_sync_unit.py \
tests/test_daily_single_sync_validation.py \
tests/test_daily_single_hook_montage.py \
tests/test_daily_single_visual_audit.py \
tests/test_cue_slide_sync.py \
tests/test_spoken_visual_sync.py -q
python -m praisonaippt.daily_single --project examples/videos/<slug> validate-spoken-visual
python -m praisonaippt.daily_single --project examples/videos/<slug> validate-sync --runs 3
Outputs: merge/sync_validation_report.json, merge/display_sync_report.json, merge/spoken_visual_sync_report.json
Additional resources
- Full documentation: docs/daily-single-video.md
- Phase 1 vs 2 detail: reference.md
- REUSE bridge:
examples/videos/<slug>/REUSE.md - Captions:
.cursor/skills/video-script-captions/SKILL.md