Daily single video pipeline

Daily single video pipeline

Build ~5–9 minute YouTube walkthroughs from create-news research handoff: plain language, June-style hook/outro, script-locked captions, and multi-layer spoken↔visual validation.

Pilot project: examples/videos/anthropic-claude-fable-5-mythos-5/
June reference: examples/videos/june-2026-ai-roundup/
Agent skill: .cursor/skills/daily-single-video/SKILL.md

Architecture

Two repos:

conda activate test
cd /path/to/praisonaippt
PROJECT=examples/videos/<slug>
python -m praisonaippt.daily_single --project $PROJECT <command>

Do not use segment-video-roundup/scripts/bootstrap-project.sh for daily_single projects.

Standard pipeline order

Run in this order for every rebuild (repeat after script or handoff changes).

Build-only (legacy)

PROJECT=examples/videos/anthropic-claude-fable-5-mythos-5

python -m praisonaippt.daily_single --project $PROJECT sync-assets
python -m praisonaippt.daily_single --project $PROJECT synthesise-vo
python -m praisonaippt.daily_single --project $PROJECT bookend-media 00-hook 99-outro
python -m praisonaippt.daily_single --project $PROJECT assemble-beats
python -m praisonaippt.daily_single --project $PROJECT build-captions
python -m praisonaippt.daily_single --project $PROJECT audit-visual --interval 5
python -m praisonaippt.daily_single --project $PROJECT validate-sync --runs 3
python -m praisonaippt.daily_single --project $PROJECT validate-all

QA-gated pipeline (recommended)

Insert modular QA gates between phases so failures surface early:

PROJECT=examples/videos/anthropic-claude-fable-5-mythos-5

daily-single -p $PROJECT write-scripts                    # if segments missing
daily-single -p $PROJECT validate-qa --when pre_build
daily-single -p $PROJECT sync-assets
daily-single -p $PROJECT synthesise-vo
daily-single -p $PROJECT validate-qa --when post_vo
daily-single -p $PROJECT bookend-media 00-hook 99-outro
daily-single -p $PROJECT validate-qa --when pre_assemble
daily-single -p $PROJECT assemble-beats
daily-single -p $PROJECT build-captions
daily-single -p $PROJECT validate-qa --when post_build    # includes visual audit + sync×3
daily-single -p $PROJECT validate-all                     # optional confirm

Full testing breakdown: Daily single testing.
Step-by-step agent skill: .cursor/skills/daily-single-video-pipeline/SKILL.md.

After any script.md edit: VO → bookends (if hook/outro) → assemble → captions → audit → validate.

CLI commands

Console alias: daily-single (see pyproject.toml).

Project layout

examples/videos/<slug>/
  manifest.json              # points at create-news research + beat-map
  segments/
    00-hook/script.md, heygen.mp4, hook_montage.json
    01-cold-open … 10-alignment/script.md
    99-outro/script.md, heygen.mp4
  beats/00-hook.mp4 … 99-outro.mp4
  merge/
    final.mp4                # loudnorm final (1920×1080)
    final.srt
    timeline.json
    display_sync_report.json
    sync_validation_report.json
    visual_audit_report.json
    asset_sync_report.json
    visual_audit_frames/
    qa/                        # modular validate-qa reports
      summary.json
      s*_report.json
  scripts/config/protocol.json

create-news assets: research/<slug>/review-assets/<slug>/ (PNGs + videos/*.mp4).

Script contract

Hook (00-hook) — three SRT cues

1. Attention — one punchy sentence. Label Hook: is stripped at TTS only.
2. Overview — comma-separated roll-call (one phrase → one montage hero).
3. Bridge — exactly: Let's get started.

Example (Fable pilot):

Hook: Anthropic just dropped Claude Fable five — if AI is part of your job, this launch changes everything.

In the next five minutes: Fable versus Mythos, Stripe's fifty-million-line proof, benchmark scores that matter, safety without dead ends, and the app-versus-API mistake teams keep making.

Let's get started.

Beats (01–10)

Content starts after Let’s get started. ~2–4 plain-language sentences per beat; one sentence = one SRT cue.

Outro (99-outro)

June CTA only — do not speak mer.vin URLs:

I hope you liked this video. If it helped, like, share, and subscribe for the latest AI updates. Thanks for watching.

Captions detail: video-script-captions skill.

Hook montage (required)

June roundup uses many hero swaps during the hook overview. daily_single Phase 1.5 implements phrase-synced montage via ffmpeg (not full compositor yet).

Assembly (assemble.py _hook_montage()):

1. Attention → first hero slide full-screen (not launch B-roll at t=0).
2. Overview → N hero PNGs (~2 s each, word-weighted).
3. Bridge → HeyGen PiP over hero background.

Artefact: segments/00-hook/hook_montage.json
Module: praisonaippt/daily_single/hook_montage.py

Gates: validate-sync → hook_montage — ≥5 distinct heroes, overview alignment ≥0.45, no launch-only overview.

Asset sync (sync-assets)

Module: praisonaippt/daily_single/media_sync.py

yt-dlp format (do not use best[ext=mp4] alone — yields 360p):

bestvideo[height<=1080][ext=mp4]+bestaudio[ext=m4a]/bestvideo[height<=1080]+bestaudio/best[height<=1080]/best
--merge-output-format mp4

Fable motion inventory

Upstream harvest: create-news scripts/harvest_motion.py (same HD format in media_download.py).

Beat assembly routing

VO duration drives clip lengths in assemble.py:

Validation layers

See Daily single testing for a plain-language guide to every test type, pytest commands, and the recommended checklist.

Modular QA (validate-qa)

Eleven stages (s00–s10) run at pipeline phases. Reports: merge/qa/summary.json.
Full reference: Video QA.

Legacy gates (still used at publish)

These run inside s10-final-composite but can be invoked standalone:

1. validate-sync --runs 3

Idempotent suite; fails if three consecutive reports differ.

Output: merge/sync_validation_report.json

2. validate-display

Per-cue midpoint: keyword alignment ≥ 0.35.

Output: merge/display_sync_report.json

3. validate-spoken-visual

Full-video spoken↔visual gate — stricter than display sync alone:

Output: merge/spoken_visual_sync_report.json — publish requires "ok": true.

Cue-aligned assembly (beat-06 safeguards, beat-01 views): run build-captions before assemble-beats so merge/final.srt drives slide durations. See Daily single testing — spoken↔visual and skill spoken-visual-sync.md.

daily-single -p $PROJECT build-captions
daily-single -p $PROJECT assemble-beats
daily-single -p $PROJECT validate-display
daily-single -p $PROJECT validate-spoken-visual

4. audit-visual

Samples final.mp4 every 5 s (+ cue midpoints). Pixel similarity vs planned asset:

Blocks generic B-roll (claudeai-launch.mp4) when enabled in protocol.
Output: merge/visual_audit_report.json, frames in merge/visual_audit_frames/.

5. validate-all

Final gate: tools, 1920×1080 output, duration 280–540 s, beat coverage, HeyGen bookends, media inventory ≥720p, all reports pass.
Output: validation_report.json at project root.

Package modules

Environment

Loaded from praisonaippt/.env and ~/elevenlabsAutomation/.env:

Tools: ffmpeg, ffprobe, yt-dlp, curl.

Phase 2 (HeyGen compositor)

When HeyGen credits return — hook segment first:

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

Same script.md contract. Phase 1.5 hook_montage.json feeds sync-media. See segment-video-roundup skill.

Tests

Full guide: Daily single testing

# Unit tests (after code changes)
pytest tests/test_video_qa.py \
       tests/test_daily_single_sync_validation.py \
       tests/test_daily_single_display_sync_unit.py \
       tests/test_daily_single_hook_montage.py \
       tests/test_daily_single_media_sync.py \
       tests/test_daily_single_visual_audit.py \
       tests/test_daily_single_youtube_quality.py \
       tests/test_daily_single_captions.py -q

# Project gates (after rebuild)
daily-single -p $PROJECT validate-qa --when post_build
daily-single -p $PROJECT validate-sync --runs 3

Run ×3 sync after pipeline changes for idempotency confidence.

Known issues

create-news upstream

Bootstrap: scripts/bootstrap-daily-single.sh

Related documentation

• Pipeline overview
• Video QA (modular stages)
• Daily single testing
• Commands reference — daily_single section
• Video examples
• Fable pilot REUSE
• Video export