workflow-ugc-automation — URL in, a reviewed batch of one-variable variants out

The transferable part of davidaistar's pipeline is the orchestration, not StarPop: brand DNA from a URL, a vision model breaking a reference into per-second beats, an LLM chunking those into ≤15 s prompts, one base fanned into variants, parallel generation, a review step. We rebuild it on our own scripts, API-first (no browser driving a SaaS UI), with our gates. Free steps can run unattended. Nothing paid runs without Fish's go for that batch, and nothing in this workflow writes to Meta.

1. Read first

• knowledge/ad-formats/_FORMAT-CORE.md: doctrine §1, lanes §2, shot tiers §2b, lane limits §3 (wrist rule), gates §4. Binding. Photoreal people = ugc-omni and its G0-G4 gates (§7 here).
• knowledge/ad-formats/_MODEL-REGISTRY.md: every generation step below names a ROLE; the model comes from there. fal is the default gateway.
• references/procedure.md: commands, file schemas (brand-dna, scenes, variants, manifest, review), prompt template, catalog refresh, worked ILLUSTRATIVE example.
• references/sources.md: the davidaistar videos, what we kept and what we refused.
• Brand: knowledge/brands/dawnbands/product-truth.md (hardware facts), knowledge/brands/dawnbands/claims-register.md, knowledge/brands/dawnbands/brand-kit.md.
• The format skill for the ad you're batching (format-ad-clone for the source-type decision; format-podcast, format-street-interview, ugc-omni, cartoon-h3 for production rules). skills/ugc-omni/SKILL.md + scripts/ugc_omni/omni.py are THE photoreal engine (Fish 10-06; another session owns them: read, never edit). A batch's photoreal base is one ugc-omni job (references/omni-job.example.json, illustrative). This workflow orchestrates; it never overrides a format's beat sheet.

2. When to use / when not

• Use when the job is a BATCH: one approved base ad → 3-12 variants, or one product URL → several formats tested in parallel, or a winner to fan out one variable at a time.
• Use step 1 alone (brand DNA) whenever a brief needs current product images, price, palette or PDP copy: it's free and read-only.
• Don't use for a single bespoke ad (go straight to its format skill), for natives (dawn-native-brief → dawn-native-write), or to clone a lane-leader competitor's script (format-ad-clone: their dissect is a gap map).
• Evidence status: davidaistar claims no CPA/ROAS for the automated outputs (f83UH: qualitative only). Our proof is the one-variable test design, not his numbers. The pipeline is untested end-to-end on Dawn; each lane under it has its own receipts.

3. Inputs

4. Outputs

All under output/ugc-auto/<YYYY-MM-DD>-<slug>/ (new slug per batch, session freshness): - brand-dna.json copy (source of truth stays output/brand-dna/<handle>/, with images/ + raw/ receipts) - dissect/ pointer to output/ad-dissect/<slug>/ (dissect.json, pacing.json, dissect.md) - scenes.json: the base ad as ≤15 s generation units, each with role, tier, Action/Audio/Quality prompt blocks - variants.json: the variant list, ONE axis + value per variant, the parent and what stays constant - manifest.json: per variant per unit: lane, registry role, model id, job dir, cost quote, actual cost, takes to approved, QA result, file (review.md rolls these up into cost per approved variant, failed takes included: starpop seedance-2-5-ads, top-5-ai-video-generators-realistic-ugc) - review.md + ads.json: Fish's review sheet and the hub payload; final/<variant-id>.mp4 (≤24 MB copies for the hub)

5. Process

• Step 0 — combo plan (feedback loop, workflow-combo-loop): python3 scripts/combo_loop/plan.py --format workflow-ugc-automation --brand dawn → pick one combo (a paying parent + ONE changed dimension) and paste its block into the pack front matter / job or spec JSON ("combo": {...}; no file → tags.py register --prefix): combo_avatar combo_angle combo_pov combo_authority combo_stake combo_stake_on combo_emotion combo_root_cause combo_mechanism combo_payoff combo_device combo_format combo_parent combo_variable combo_ad_prefix (= the launched Meta ad-name prefix). python3 scripts/combo_loop/tags.py check <pack> must PASS before concept approval; the weekly refresh reads results back by that prefix.

Step rules: 1. Brand DNA reads only the public storefront (<pdp> HTML, <pdp>.json, first-party theme CSS) with curl. It writes images, colours, fonts, PDP copy, price, review snippets and warnings. Act on every warning: a handle redirect means ad links use the canonical URL; lit-display images are never product-in-scene or i2v refs (use knowledge/brands/dawnbands/product-cutout-unlit.png); a price mismatch with product-truth means the live PDP wins for offer lines and Fish hears about it. Hardware facts still come from product-truth only; PDP review snippets are store-curated, not VOC. 2. Dissect gives structure and measured pacing. It does NOT give our script. Source-type decision per format-ad-clone: our winner / non-competitor = mirror the structure with one innovation; competitor = gap map, must fail the competitor-swap test. 3. Scene chunking: group beats into generation units that respect the lane cap (H3 ≤15 s/≤4 panels; omni buckets 4/6/8/10 s, one script line per take; Seedance b-roll per registry). Each unit carries the dissect's role, a tier (A/B/C, core §2b), and three blocks: Action (who, doing what, framing), Audio (voice, room tone, delivery, exact line), Quality (capture device, light, aspect). Give the model the scene + the line and let it improvise camera work; lock only what the gate checks (product, wrist, words). Every ref gets one stated job in a reference map, every shot one action and one camera instruction, every seam between shots is chosen (cut or stated transition), and footage comes back clean of captions/text (procedure §3; starpop Seedance/Kling guides). 4. Variants: one axis per variant vs the base: hook (first 3 s line + action), actor (one locked character swapped for another), setting (environment only), product_angle (which benefit/mechanism beat leads), delivery (same words, a different read: calm / tired / skeptical / deadpan; starpop how-to-use-seedance-2-0-to-make-ads, how-to-use-kling-3-0-to-make-ads). Everything else is byte-identical (same body units reused, not regenerated). Two axes changed = two variants. 3 hooks on a shared body is the default first fan-out (core §4 gate 2). 5. Draft first (hub rule): the plan (scenes, hooks, start frames if any) goes up as a Draft session; Fish reviews there. A missing verdict is NOT approval. Fish's go is a chat message naming this batch and its quote; write it into manifest.json go. 6. Draft → final + fan-out (core gate 6b): where the lane has a draft route (Seedance 2.5 Draft → "Generate in 1080p" on Higgsfield; fal Flux 3 */draft → draft-enhance; H3 480P → upscale only after a side-by-side vs native 1080), iterate tier-A/B units on drafts and render only approved drafts at final quality. ugc-omni has no draft step: its G2 lock (same line good 3×) is the smoke and the proof, and every A-roll take is Omni (tiers only pick b-roll). H3 cartoon batches draft with run.py --draft (480P, own <slug>-draft dir); no other lane has a batch draft stage (NOT BUILT). Then smoke ONE unit of the base on its lane, Fish's eye on it, then the rest. Shared body first, then hooks (shared-batch race, cartoon-h3). Regens are their own calls and count against a batch cap of 8. 7. QA: review in a fixed order and pick correct over pretty: product → face/wardrobe → hands/action → dialogue → continuity → camera → audio/ending. Before any regen, name the cause (reference, prompt, scope, settings) and change only that. Freezes 0, dark frames 0, transcript = script (omni verify), captions from the script, band dark and wrist down in every product shot, no smartwatch. A generated band appears only through ugc-omni's product gate (variant/pick --product ≥7/10 + display state, prodcheck), display-off wording, worn or resting across a palm (never pinched upright), product_action locked; real footage is the fallback. Gemini samples ~1 fps: eyes are the final gate. 8. Review sheet: one row per variant (procedure §8). Stage finals into the SAME hub session (--finalize), never a second session.

6. Tools & models

Roles come from _MODEL-REGISTRY.md; re-read it each batch. Lane per format: | Format (his five) | Our lane + entry | Registry role(s) | Status | |---|---|---|---| | Handheld selfie talking head | ugc-omni scripts/ugc_omni/omni.py (cost, lock, gen, verify, broll, assemble; dry run until --spend) | Photoreal / multi-ref video with locked character; Image gen (avatar/frames) | Built (another session owns ugc-omni: reference, don't edit) | | One-shot UGC with b-roll | ugc-omni A-roll + broll; or Seedance b-roll scripts/seedance_broll_runner.py | + Cheap simple-motion i2v (tier C) | Built. Runner models are Higgsfield seedance1_5/seedance_2_0 only | | Single-take native speech (yapper) | ugc-omni mode: oneshot (one location, no b-roll, chained segments on exact frames, anchor_every: 3, reveal via end_frame) | Photoreal talking take (Omni) | Built. H3 native voice via yapper.py = side arm only (one variable vs a locked omni baseline); its keyframe chain stays closed (Fish 10-05) | | Pixar / cartoon object | python3 -m scripts.cartoon_h3.run (+ matrix for hooks × leads) | Stylized video, ref + audio clock; TTS | Built, full gates in cartoon-h3; no lip-sync (core §3): his talking-object shot is staged off-mouth | | Podcast two-shot | format-podcast | per that skill | per that skill | | Statics from the same DNA | scripts/statics/gen_statics.py (dawn-statics) | Image gen; Exact text on image | Built | | Cinematic short (no VO) | not run | — | Skipped: he rates it weak for DR (f83UH) | Provider conflicts, flag once per batch (not resolved here): (1) the registry says fal is the default gateway, but seedance_broll_runner.py is Higgsfield-only (seedance1_5/seedance_2_0, no 2.5) and skills/vo-broll/SKILL.md (9-15) routes generated clips to Higgsfield; (2) ugc-omni runs on kie, not fal; (3) image arms: pass --arms gpt2,sd5 (Fish 10-06: Nano Banana retired, Seedream 5 Pro is arm 2; omni.py's own default still lists nbp, its owner changes that). Ask Fish at the batch go which applies.

Shared: dissect + QA = Video understanding (scripts/ad_dissector.py, scripts/gemini_qa.py, scripts/cartoon_h3/clipqa.py); scripts = LLM writing (scripts/claude_cli.py); staging = ~/ad-batch-hub/stage-session.py, stats = ~/ad-batch-hub/scripts/refresh.sh (read-only Meta GET); cartoon audit site = python3 -m scripts.cartoon_h3.hub.

NOT BUILT (needs Fish's go): - Built 10-06 (free, scripts/ugc_tools/): scene_chunker.py, batch_variation.py (plan / approve / check / log; check exits 2 without Fish's go), review_sheet.py. Not built: per-lane job dirs from a variant (each lane's own entry point still runs the units: matrix.py, omni hook lines, gen_statics --only), and no lane runner calls batch_variation.py check yet. - Seedance 2.5 (fal) in seedance_broll_runner.py; Video extend, Video edit by instruction, Motion transfer/recast roles (all NOT BUILT in the registry). - Static swipe-file library (his 136 ripped templates) under knowledge/. - A fal catalog refresh/diff script (procedure §0 is a curl loop + a python one-liner). - Browser-driving a generation SaaS (his Claude-in-Chrome step): deliberately not built; API entry points replace it.

7. Gates & QA checklist

• [ ] Brand DNA run today; every warning acted on (canonical URL, lit images excluded, price mismatch told to Fish).
• [ ] Parent named with CPA; batch purpose + constant + variable written (hub test_design).
• [ ] Base script went through the format beat sheet + copy routing; problem-aware line 1 names HER by a symptom; stake on her, caused by the mornings; recipe A (story: mom narrator, accuser, explainer inside the story, vindication) or recipe B (callout: brand voice, her daily burden, kid-independence payoff) per the format.
• [ ] Competitor source → gap map only; concept fails the competitor-swap test.
• [ ] variants.json: every variant changes exactly ONE axis; constant units reused, not regenerated.
• [ ] Every unit ≤ its lane cap; tier set; tier-A only on product/hands/CTA/hook.
• [ ] Photoreal units = ugc-omni G0-G4 (same in format-podcast / -street-interview / -ai-actor-demo / -motion-transfer / -whistleblower / -reaction-duet): G0 script + hooks + board (no paid step before it) → G1 real ref frame + avatar (+ gated product frames) → G2 lock: same line good 3× (the 9-30 motion proof) → G3 every A-roll line + b-roll + product check → G4 the assembled cut. Every gate is Fish's; the batch go is on top. An actor variant = a new G1 + G2. No face-blur or filter workarounds.
• [ ] Generated band frames only through ugc-omni's product gate (pick --product ≥7/10, right display state); everything else real footage. Seedance 2.5 product shots stay rejected (Fish 10-06).
• [ ] Draft → final run where the lane has a draft route (step 6); tier recorded per unit in manifest.json.
• [ ] Draft staged; Fish's verdicts pulled; Fish's batch go recorded with the quote; balances checked (fal, Higgsfield, kie).
• [ ] Smoke unit approved by eye before the fan-out; regens ≤8; --candidates 1; never Higgsfield UI Rerun.
• [ ] QA: script-true transcript, captions from script, band dark + wrist down, no lit face, freezes 0; Fish watched.
• [ ] Hub finalized in the same session; nothing sent to Meta from this workflow.

8. Failure modes

• Wrong product page: an old handle silently redirects (Dawn's default PDP does) → links point at the redirect; use canonical_url.
• Lit-band refs: the PDP hero and colour shots show a lit display → H3/Seedance draws a smartwatch. Product refs = the permanent real-photo set knowledge/brands/dawnbands/product-refs/ (Fish 10-06): pack_off.jpg by default, pack_on.jpg only when the lit time is the point; never PDP renders.
• Hand/product fidelity: ungated band-in-hand scored 3-4/10 (yapper 10-04) → product frames go through ugc-omni's gate (exact first frame scored vs the real pack; close framing; product_action locked, else the hand drops at 0.25 s) or real footage. A new model candidate gets a one-unit proof first.
• Variant creep: hook + actor changed together → the read is meaningless. Split.
• Accidental spend: seedance_broll_runner.py submits unless --dry-run; omni/run.py are dry until --spend. Always quote first.
• Voice drift across units: separately generated takes change voice → one locked prompt/voice per actor (omni lock/voice), one voice pass over all units.
• Price drift: product-truth said $39.99 while the live PDP said $44.99 (10-06) → offer lines from brand-dna, Fish told.
• Silence read as approval on the hub → only explicit verdicts + a chat go count.
• Parallel race: hooks generated before the shared body finishes → duplicate paid body batches.

9. Sources

~/research/davidaistar/REPORT.md; analysis/gemini_f83UH-qmNUs.md, analysis/gemini_aqo2L5wYGEg.md, analysis/ugc_seedance.md (f83UH, aqo2 sections, cross-video #6-7), analysis/guide_copy_images_shorts.md (shorts wGVBPiaFYiw, cPQ9e0RfiQE, J7k_eNpLO3o, 2zOlhDAOwb4; variation mode), analysis/diff_ai_ugc.md, analysis/diff_cloning.md. Ours: project_ad_batch_hub.md, feedback_higgsfield_credit_discipline.md, feedback_meta_upload_preflight.md, feedback_meta_link_match.md, reference_shopify_stores.md (public storefront only). Detail: references/sources.md.

Not built yet (needs Fish's go)

Every line in this skill that mentions something not built, roughly de-duplicated. Some items repeat in different words.

• 6. Draft → final + fan-out (core gate 6b): where the lane has a draft route (Seedance 2.5 Draft → "Generate in 1080p" on Higgsfield; fal Flux 3 /draft → draft-enhance; H3 480P → upscale only after a side-by-side vs native 1080), iterate tie
• NOT BUILT (needs Fish's go)
• Built 10-06 (free, scripts/ugc_tools/): scene_chunker.py, batch_variation.py (plan / approve / check / log; check exits 2 without Fish's go), review_sheet.py. Not built: per-lane job dirs from a variant (each lane's own entry point still ru
• Seedance 2.5 (fal) in seedance_broll_runner.py; Video extend, Video edit by instruction, Motion transfer/recast roles (all NOT BUILT in the registry)
• Browser-driving a generation SaaS (his Claude-in-Chrome step): deliberately not built; API entry points replace it
• Changed / refused: no motion swap (registry Motion transfer role is NOT BUILT); no uncensored-model routing; no Sora-style "remix" of someone else's video
• analysis/diff_cloning.md: P0 dissector (built: scripts/ad_dissector.py); P1 batch_variation.py (built 10-06 as scripts/ugc_tools/batch_variation.py, scene-set variants, not image batches) + static swipe archive (NOT BUILT); competitor disse
• everything-you-need-to-know-about-seedance-2-0: extension length = the new segment, not the total; up to 12 input files with 15 s combined video/audio refs. Registry-level (Video extend role NOT BUILT); proposed note there