Sung character ads
One integrated vocal-and-music track, then its film.
Ready to use when creating the new project
A format we love.
New angles, products and people.
This is the implementation brief shaped by your answers. Keep the familiar trend studio, reuse the working code, and make the reverse-engineered format the thing that travels between films.
Paste the brief or this page URL into the new project. The kit contains instructions and manifests; production code and videos come from the verified donor sources.
01 / The direction is decided
The new brief replaces the earlier open questions. Intermediate acceptance happens inside the workflow; you review the complete final film.
One integrated vocal-and-music track, then its film.
Persona is optional. Choose the route explicitly and keep its real results.
Keep the music, visual finish, acting and pace close; adapt product and passports.
Use it for other angles, products and matching new trends.
Keep executing internal reviews and repairs within the recorded brief and budget.
Adjust the blocking creative criterion visibly; preserve every earlier result.
The exact product evidence bounds what the film can show.
Source, reverse-engineered flow, recreate controls and finished adaptations.
02 / This is what we reuse between films
Keep the source's recognizable creative character. Separate the reusable mechanisms from the facts that belong to one product or one story.
Format matching stores reasons and differences. If the new trend does not fit, mark the mismatch and create an explicit new format. Keep old runs pinned to the format version they used.
Source playback, format label, matching evidence and adaptations.
What makes it work, what stays, what varies, and which 10/10 examples guide it.
Product + geo + angle + passports + budget. Show the selected music route.
Complete playback, version, cost, compromises and approve / revise.
03 / Reuse the wheel
82 pinned file entries in 12 groups. Every link points to a specific commit. The new project gets its own code and data; it must not depend on a live worktree.
Showing all 12 groups.
Reuse renderWorld, buildProgramHtml and renderFactory interaction patterns; replace catalog assumptions with format → trend → adaptation data. Do not copy the 2,000-line app wholesale.
web/trends, web/formats, web/adaptationsUse mechanisms, evidence, preserve/change decisions and source-fidelity structure. Lift named cast, product, geography, role counts and story details into per-format slots; the Target example is not a universal template.
formats/<id>/format.json and source evidenceReuse actual upload-cover, task persistence, explicit take selection, audition and authorized edge-edit code. Keep song_contract, song_comparison and musical fingerprint dependencies together while extracting. Retain tests and adapt geo/language assumptions.
app/providers/suno.py and app/audio/Reuse the adapter and tests; it depends on gen helpers. Creation failed in the inspected experiment. Port it as an optional route, never mark its capability proven or force every film through it.
app/providers/suno_persona.pyReuse video_reference_paths, clip_input, actual negative-prompt delivery, reference hashes and paid-ledger merging. gen and harness/thumbs imports need explicit rewiring. Keep documented multimodal mode constraints and tests.
app/providers/visuals.pyReuse complete PCM identity verification, listening receipt capture and independent evaluation; identity imports gen and music_dna_reel. Preserve exact media contracts. Deep groove diagnostics stay available without becoming mandatory research for every run.
app/audio/identity.py and app/quality/audio.pySmall standard-library module: copy with its tests, preserve behavior, then parameterize language only where needed. Do not reimplement word fragmentation and phrase grouping.
app/media/song_captions.py and tests/Extract Worker.assemble_video, caption_ass, ass_time, ass_text, clip_sources, lyric_pills and their required helpers/constants. Supply explicit paths, dimensions and caption font. Do not instantiate the legacy Worker in the new app or copy GoPure one-off scripts.
app/media/assemble.pyRetain exact product/cast/state bindings and actual motion evidence. review_problem depends on gen, music_dna_research and the harness contracts. Replace fixed taste thresholds with the owner's versioned fifth-rejection policy; never rewrite historical outcomes.
app/quality/visual.py and contracts/Reuse existing Python/Postgres/R2 contracts and failure semantics. Keep ObjectStore's config/httpx dependencies. jobs and worker are coupled to the old Factory; extract leases, task reconciliation and event projection rather than importing its full workflow or adding another queue stack.
app/jobs, app/storage, app/dbCopy the skills and required references as a starting point, then reconcile explicit owner policy in PROJECT_START.md before running imported validators. Preserve source-specific observations; 10/10 is owner taste, not a computed metric or a universal musical preset.
references/music_dna_raw and skills/Reuse the dossier structure and inspect the worked problem passport. Replace all GoPure facts, prices, geography, characters, images and endpoint when adapting a different product. Recorded seller claims are not automatically verified product evidence.
products/<product>/ and passports/<cast>/These are inspected reuse candidates, not a claim that the 82 files form a standalone runnable app. The implementation must resolve imports, fonts, media tools, storage bindings and meaningful regression tests. Missing dependencies fail visibly.
04 / Our creative north star
The exact folder is data/music_dna_raw. All five originals are present and match the saved corpus hashes. They are the creative examples you love, across music, visuals, acting, pacing and story.
6:29 · original video
SHA-256 verified6dc251be3d0ca7a1…3:05 · original video
SHA-256 verifiedf90f964fb987783e…8:01 · original video
SHA-256 verifiedb8e3bc06f423d498…4:40 · original video
SHA-256 verifiedd22b2030c289e856…3:14 · original video
SHA-256 verified1ad8be404d159994…Calibrate reviews to the closest loved example. Do not flatten five different references into one compulsory voice, genre or plot. The 10/10 rating expresses your taste; product claims still need their own evidence.
Verified source directory on this host: /workspaces/UGS/LevchykUGS/data/music_dna_raw/. Import read-only into references/music_dna_raw/ or content-addressed storage and verify every SHA-256 in the manifest.
On another host, transfer those same original files through the authorized media store. The public startup kit contains their manifest, not the videos. If an original is missing or mismatched, stop the import; generated remakes and thumbnails do not replace it.
Download the verified reference manifest ↓05 / Your new iteration policy
“More than four” means the fifth observed rejection for the same creative criterion. It triggers a recorded policy adjustment, not another identical loop.
Count by asset lineage, criterion and creative intent. A new take ID does not reset the count.
Version the softer threshold. Save the five rejection IDs, old/new bounds and the reason.
Test the best preserved candidate under the new policy. Keep the old rejection and accepted siblings.
Examples: exact reaction timing, intensity above basic readability, a minor groove-similarity score or nonessential style precision.
Suggested implementation: a 0.5 reduction on a 1–5 subjective scale, down to 3.0; or a 20% tolerance increase, at most twice. These are configurable implementation defaults, not numbers you supplied. Wait for five further rejections before another adjustment.
Do not loosen product evidence, required words, intelligibility, identities, concern ownership, actual use and elapsed time, observable required faces/states, complete media or budget limits.
A provider error or an inaccessible review is not a creative rejection. It does not increment this counter. Show any relaxed criteria with the final film; never relabel a prior failure as an unchanged strict pass.
This is an intentional change from the donor's blanket “never lower a threshold” rule. The new project's policy and ported validator must agree. Existing runs in the donor project are not rewritten.
06 / Speed without starting over
Reuse a verified format instead of analyzing it again. Reuse matching accepted assets. Keep provider dependencies warm, run independent ready scenes concurrently and repair only the failed descendants.
Use tested provider payloads, complete-audio transport and the selected format's reference pack.
Code handles uploads, polls, captions and encoding. The planner handles creative decisions and specific repairs.
Internal acceptance continues automatically within authority. Deliver one complete film with honest status.
Measure the new implementation's time and cost per accepted minute. The earlier 45–90-minute estimate remains a hypothesis, not a delivery promise.
07 / Give this to the new project
Use this page URL as project context, copy the full brief, or attach the startup ZIP. The instructions name exactly what to reuse, where to find it and what to change.
Your decisions, the domain model, donor files, import sequence, validation policy and definition of done.
Read PROJECT_START.md ↗Copy the supplied instructions to the new repo's AGENTS.md. They reconcile your latest policy with historical skills.
Download instructions ↓Commit-pinned source files and five hash-verified, owner-rated 10/10 originals.
Reuse map ↓Reference manifest ↓82 source entries · 5 loved originals · your latest decisions
This is the startup brief and reusable-source inventory. The new application has not been built by this page, and this page makes no paid generation calls. Donor GitHub links require access; the ZIP does not include private source code, secrets or the original videos.