davidaistar teardown → our formats → Dawn test plan (2026-10-06)
Sources: 62 videos (23 tutorials + 39 shorts). 59 transcripts plus Gemini video passes on the 3 that rate-limited. 13 unique Foreplay ads (#10 and #11 are the same link), each downloaded and broken down by Gemini.
Raw data: ~/research/davidaistar/{txt,vids,analysis,foreplay}/
0. The headline
- We already run his core pipeline. cartoon-h3 was adopted from his MiniMax H3 workflow (
knowledge/references/minimax-h3/). His "Singing ads (Suno + Minimax)" tutorial is the same as our song lane: song first, then a character sheet, 2x2 contact-sheet storyboards, H3 with sliced audio, native audio muted, master track laid back over. Our song lane does that and is Dawn's best video format (sl0919-h2: $40 CPA; songs: 3.8-4.7% CTR at half the frequency build). - What's new is in his podcast, clone-masterclass and Seedance videos. These are dialogue/photoreal tricks, styles we don't run yet, and the clone method. See §2.
- His model calls are already out of date (Flux 3, Nano Banana 2, Seedream 4.5, Kling 3, Seedance 2 Mini). The steps hold up; swap in what we have: H3-max (fal/Higgsfield/kie), GPT Image 2.5, Suno V5 via kie, Hailuo-02, Kling Avatar/OmniHuman, H3 lip-sync.
1. His pipelines, one line each
| Video | Format | Pipeline core | We have it? |
|---|---|---|---|
| AEkzwbp6RRU | H3 animation ads | Reddit/TikTok research → full VO → dead-air cut → word timings → batch beats to the 15 s cap → H3 + audio slice → CapCut | YES = cartoon-h3 |
| 8NzJJ3e7uZ4 | Singing ads | Lyrics from research → Suno FIRST → char sheet → 2x2 contact-sheet storyboards → H3 (panel + char ref + sliced song) → mute native audio, lay master back over → 2-3 s hyper-motion product end card | YES = song lane (end card is new) |
| TozlaC3-rCA | Pixar character ads | Seedream-type CGI image model for Pixar → char + env sheets ("3D floor plan view" for sets) → i2v → VO | YES (cartoon-h3 Pixar) |
| N9Cl7PKrlxQ | Claymation | Clay-style stills → i2v → VO | YES (claymation batch, 2 live) |
| 867Fbi-lyCw | Paper animation (~$1/ad) | Nano-Banana-type paper-cutout stills → cheap i2v ("simple motion" model) → VO | NO: new style on the existing lane |
| IolBFaQIrl4 | Zach-D style animated | Exaggerated 2D cartoon style, VO-led | NO: style swap |
| bw1LPGZWf-k / HBIENgQIQVU | Clone any ad | Hand-tag the reference's beats (hook/agitation/objection/mechanism/CTA) → LLM rewrites at the same pacing → hard cuts are easy, transitions need start/end frames → do split-screens and overlays in the editor → 80% proven structure / 20% new | PARTLY (we clone structures; no frame-by-frame dissect tool) |
| GIczkree_W0 / pYNC49BHsHQ | AI podcast | Pinterest real-photo refs → characters → put ALL of one speaker's lines in ONE generation (beats the min-duration limit, holds voice) → cut line by line in the edit → generate 16:9, crop to 9:16 (reads more native) → patch clipped last words with an EL clone | NO |
| PmXqgiFn9_Y | UGC + song lip sync ~$1 | Still image → audio-driven lip sync | Tested in yapper (closed) |
| CZwfTy7jZcQ / qXiHN3 / OvAyKctgNE4 / f83UH | Seedance 2 UGC | Omni-reference with MANY real product photos (angles, in hand, packaging) → in-model edit/extend for long takes → change ONE variable per pass → Claude skill: Gemini per-second breakdown → ≤15 s scene prompts | Partial (Seedance unlimited expired 9-09; fal blocks realistic faces on ref-to-video, i2v works) |
| gpzy2ofG0Dw / Rqim7SlMtZg | TikTok Shop clones | Mine GMV winners (FastMoss/Kalodata) → port the psychological mechanism across categories → AI actor | Concept only |
| _Y7KFZHxMe0 | Copy research | Reddit/review mining → LLM script with hooks sourced from named threads | Ours is deeper (atlas, VOC, combo map) |
2. Techniques worth stealing (we don't do these yet)
- Hyper-motion product end card (2-3 s) on every song/cartoon. Cheap, and could lift CTR. Build once, reuse on every ad.
- One generation per speaker. All of a character's lines go in one H3/Seedance call, then get cut in the edit. This unlocks two-character dialogue formats without per-line voice drift.
- Voice from the video model, then clone it. Generate the character's voice natively in a video model, clone it in EL and run the full script through the clone. More natural than stock EL voices for dialogue characters.
- Formats with one shot per person (street interview, vox-pop). Every clip is a different person, so there's no identity to hold. This dodges the consistency failure that killed Side Effects and yapper.
- Obscured-face formats (masked whistleblower, silhouette, back-of-head) hide imperfect lip sync. The format solves the uncanny problem.
- Pick the model per style and test 2 models before standardizing (paper vs clay vs Pixar).
- Fix with the edit model; don't re-run the expensive one.
- Organic format hijacking. Take a viral organic format (science explainer, "rate these", "day 1 vs day 21") and graft the product onto it. Don't clone competitor ads (matches our 10-05 rule).
- Restate refs and model name on every call when an agent drives the generation, or defaults drift.
- Generate 16:9, crop to 9:16 for photoreal. The crop hides edge artifacts and reads as phone footage.
- Static templates by awareness stage (HEa4AVv-ths). TOF: "unexpected cause" / "pain gone". MOF: pain→dream, timeline, "3 reasons why". BOF: testimonial, iMessage, tweet, offer. Map these onto the recipe-B statics: Dawn is problem-aware, so the MOF set.
- Improvise first, then fix in an edit pass. Forcing exact copy into a one-shot image prompt lowers quality, so generate loose and then make a narrow edit (same lesson as our 10-04 "short plain prompts" correction).
- Product reference = a hand-held/in-context photo, never white-background only, or the scale comes out wrong. Applies to every band shot (pull real buyer photos, per the 10-05 rule).
- Pick the model per task. Kling for consistent talking characters, Seedance for multi-step product demos, Kling Motion for motion transfer (swap character + background onto a viral clip's motion).
3. The 14 Foreplay formats: can we make them?
| # | Format (example brand, length) | Can we make it now? | Lane | Call for Dawn |
|---|---|---|---|---|
| 1 | AI skits (Beaverpro, 15 s slapstick) | YES | cartoon-h3 short | Low. Problem-aware gate means it needs a symptom callout, not just absurdity |
| 2 | Whistleblower (Resilia, 3:42 masked confessional + b-roll) | PROOF FIRST | H3 lip-sync on a still + Hailuo b-roll; the mask hides sync | Medium. Insider confesses ("school attendance officer: what we see every morning") |
| 3 | Celebrity (Yuri, 70 s) | NO | n/a | Skip: real-person likeness = impersonation risk + Meta policy |
| 4 | AI drama (Koriderm, 4:32 photoreal dialogue) | ANIMATED YES / photoreal no | cartoon-h3 dialogue | High: recipe A as spoken drama (tests song vs story) |
| 5 | AI founder-led (Azure, 3:10) | Technically yes | avatar lane | Skip unless it's the real founder's voice/face; a made-up "founder" is deceptive |
| 6 | Yapper (Yuri, 2:24 talking head + PIP) | Tested, closed 10-05 | yapper | Parked by Fish |
| 7 | Timeline (IM8, 1:51 "if X did Y for 3 weeks") | YES animated | cartoon-h3 | Medium. Must carry a character; no-character explainers got 0 purchases |
| 8 | Authority whiteboard (Skincu, 2:37, failed solutions rated 2/10→100/10) | YES animated | cartoon-h3 | Medium. Rated-failed-solutions structure is strong; authority-person reads as testimonial in statics (loses) |
| 9 | Street interview (lipstick, 54 s) | PROOF FIRST | Seedance/H3 i2v with native dialogue, 1 take per person | Medium-high as the cheapest photoreal entry |
| 10/11 | Character arguments (Penrose, 62 s 3D debate) | YES | cartoon-pipeline talking objects / H3 two-voice | High as a courtroom version (exhibit device = best song CPA 2.48 ROAS) |
| 12 | "Musical" (Resilia, 7:25 narrative VSL) | YES | song lane | Already our best format |
| 13 | Suno songs (Smooche, 5:37 Pixar musical) | YES | song lane | Confirms going ≥3 min |
| 14 | Mini-movie (Rosabella, 11:51 photoreal) | Animated yes / photoreal no | cartoon-h3 long-form VO | Low until spoken drama (#4) reads |
Gemini's tool guesses for these ads (Midjourney/Runway Gen-3/HeyGen) come from its stale training data, so treat them as style hints, not facts.
4. What Dawn should test next (ranked)
One variable per ad set, mom-1st POV, stakes that land on her and were caused by the mornings (from dawn-combo-map-2026-10-05/MAP.md).
1. Keep the song queue moving (truancy letter → CPS → own kid → boss) and add the hyper-motion end card (technique 1) to every new song. It's zero-risk.
2. Re-skin a proven song (exhibit A or oneweek): same audio, new visual style (claymation or paper). Only the style changes. If it holds CPA, we have a cheap fatigue extender for every winner. Paper is his ~$1 lane.
3. Spoken drama, recipe A without the song. Mom is accused (husband or ex), the nurse explains, mom is vindicated. Animated, two-voice, one generation per speaker. This tells us whether the song or the story is carrying the winners.
4. Courtroom character argument. Mom vs the accuser, with the alarm as "exhibit A" and the band as the witness. Format #10 combined with our best device.
5. Photoreal entry test: vox-pop street interview ("How many alarms does your kid need?" 6-8 moms, one 5-8 s take each → product). Per the binding avatar gates: a motion proof on 1 clip first, then Fish's yes before any build.
Don't: celebrity, fake founder, yapper (closed), photoreal mini-movie, college/kid-money stakes, character-less explainers.
5. Before any spend
Check fal / Higgsfield / kie balances (all three went dry on 10-03). Scripts and hooks need Fish's approval before builds (cartoon-h3 gates: storyboard.approved + --spend). Nothing in this report spent credits beyond Gemini analysis.
Part 2 (10-06, after Fish: "no way you didn't have anything to add"): step-by-step diffs vs our code
The first pass compared formats; this pass compares his steps against our actual scripts and skills. All 23 long-form tutorials were watched by Gemini (on-screen prompts + UI), and 7 lane agents diffed them. Full tables with file:line citations: analysis/diff_<lane>.md.
| Lane | Biggest real gaps found | File |
|---|---|---|
| Songs | Song storyboard rules are missing from the splitter (no-mouthing, name everyone, wake staging, wrist, no early product): 3 songs hand-rewritten and ~11 regens on 10-04. H3 told the song is "narration". Possible price-table inversion in h3.py (ledger may read low). Captions possibly inside the Reels UI zone. One hook per song. Clips can't be speed-nudged. On-camera sung chorus never tested. | diff_songs.md |
| Animation styles | Mascot-explainer (solution as hero) never tried. One model for every style (paper could run on a cheap model). No last-frame chaining across batches (hard cuts at every seam). No Fish Audio option. | diff_animation_styles.md |
| Podcast/dialogue | Native-speech H3 i2v was proven 10-04 but only single-speaker; no duo generation and no two-speaker intercut assemble; real-photo refs optional; no banter script rules; no clipped-last-word patch. | diff_podcast_dialogue.md |
| Statics | No timeline or iMessage-thread formats; exact-text forced on the first generation for new layouts; no recycle-winner text swap; no hand-held scale ref; single composition candidate. | diff_statics.md |
| AI UGC | Never tested real Seedance 2.0 (omni-ref, extend); hand/product failure never isolated on another model; segment chain is hand-built; no brand-DNA intake; long 80-90 s format vs his 15-30 s. | diff_ai_ugc.md |
| Cloning | No standing dissector (built 10-06: scripts/ad_dissector.py); no organic/TikTok Shop GMV sourcing; no cheap video-to-video swap lane; one-off batch scripts instead of a variation tool; the competitor-vs-organic source rule wasn't written down. | diff_cloning.md |
| Copy | No two-person dialogue script doctrine; no long-form "free value" sell-script sub-mode; no on-demand research pull mid-brief; song-lyric method stuck in a dated output folder. Ours is deeper on VOC verification, angle grounding and claims. | diff_copy.md |
Built 10-06
scripts/cartoon_h3/endcard.py+ run.py hook: motion end card from packend_card:(--clipslot for an AI 3D spin, not rendered yet).scripts/cartoon_h3/reskin.py: same audio + beats, new style. Exhibit A → claymation staged (dawnbands-song-reskin-exhibita-clay-2026-10-06).scripts/ad_dissector.py: any ad → measured pacing + timed beat sheet + rebuild recipe.scripts/check_format_skills.py: consistency gate for format skills.knowledge/ad-formats/_FORMAT-CORE.md+_FORMAT-SKILL-TEMPLATE.md;skills/format-*(one per format); cartoon-h3 SKILL rules 9-12; ugc-omni add-on reference.
Format skills (10-06): ~/.openclaw/workspace/skills/format-*
song · spoken-drama · courtroom · mascot-explainer · timeline · rated-solutions · skit · podcast · street-interview · whistleblower · ui-mockup · ad-clone. Core: knowledge/ad-formats/_FORMAT-CORE.md. Gate: python3 scripts/check_format_skills.py → 12/12 PASS. Open items needing Fish's go: memory project_format_skills_2026-10-06.md.