Product + craft audit Current system → elite human UGC
Spielberg safely receives messy footage, keeps context, runs an editor, supports targeted revisions, and delivers one final. But the path from technically complete to hand-montaged, native, magnetic is still mostly written advice—not a repeatable product guarantee.
This is a static architecture, UX, runtime-contract, and market-craft audit—not a blinded score of customer output. The repository contains a synthetic review fixture and a real repeated-take research set, but no representative completed UGC masters or post-publish retention data. That absence is itself the largest proof gap.
01 · The verdict
A perfect montage is not more effects. It is the right 24 frames at the right moment, proof before doubt, sound carried across the cut, and every decorative move motivated.
The current system knows much of this grammar in its prompts and skills. What it does not yet do is turn those choices into required, inspectable evidence attached to the exact delivered file.
Tenant isolation, immutable raw media, resumable upload, one concise conversational reply, persistent per-montage context, idempotent delivery, and redo → fragment → approve → stitched-final semantics are unusually solid.
Today, “done” reduces to a successful assembler run plus pieces.json. Delivery can select the newest MP4 and check ownership, duration, and dimensions without a mandatory story, audio, caption, crop, freeze, or visual-proof receipt.
“Perfect” needs a gold set of messy UGC packs edited by strong human social editors, blind pairwise ratings, and—in production—retention, saves, shares, clicks, and revision decisions joined back to edit choices.
02 · Current state
Telegram or resumable web uploader stores immutable originals and basic duration/dimension metadata.
Strong primitiveOne persistent chat extracts platform, 15/30/60s, mood, captions, and music; richer intent remains free-form.
Too shallowOne autonomous editor session is expected to apply a deep library of directing, raw-edit, motion, sound, and QA doctrine.
Prose-dependentA round becomes “done” when assembly exits successfully and a pieces manifest exists.
Weak gateOne whole MP4 is sent. A targeted redo yields a fragment; approval stitches the next whole master.
Good interaction modelassemble_round exited 0
+ pieces.json exists
master.sha matches
+ quality_contract.status = pass
03 · Gap to the best human edits
This is a capability-enforcement judgment, not an empirical output score. The service is strong at correctness, has meaningful clean-up tools, and can reach higher when the agent executes the playbook. It does not yet prove those upper levels consistently.
The right tenant, montage, round, playable format, duration, and dimensions.
Bad takes, repeats, silence, captions, mix, and technical defects handled.
One promise, compressed story, early proof, earned CTA, no narrative waste.
Feels born in the feed: creator voice, safe UI, intentional imperfection, platform rhythm.
The eyebrow lift, tactile click, delayed reveal, emotional hold, and surprise that earn attention.
| Editorial decision | Service contract today | Elite human cut | Mechanism to add |
|---|---|---|---|
| Moment selection | Transcript, waveform, duplicate, and timing evidence; limited durable visual-performance data. | Picks gaze, credibility, gesture peak, microexpression, product visibility, and the revealing half-second after the line. | Performance atlas per phrase/take: emotion, eye contact, gesture, blink, energy, subject/product, technical confidence. |
| Hook | Strong written doctrine, but no required uploaded-footage hook lab or first-frame acceptance gate. | Picture, voice, text, and sound converge on one promise in the first event—not a title-card throat clear. | Three real hook assemblies: result-first, contrarian/confession, sensory demo; contact sheet + 0–2.5s preview. |
| Story compression | An agent can build beats; no required promise → proof → payoff ledger or alternate assembly. | Moves proof before explanation, deletes correct-but-unnecessary lines, preserves only setup, escalation, objection, payoff. | Edit-decision map with each beat’s intent, claim, proof asset, payoff, transition reason, and confidence. |
| Pacing | Docs reject metronomic cutting; execution remains prompt-led and has no retention feedback. | Pattern interrupts at narrative turns, holds when comprehension or emotion needs time, accelerates only for tension/payoff. | Intentional cadence: beat role + shot-length rationale + viewer-response data linked to the timeline. |
| Sound | Deterministic music and loudness tooling; limited productized dialogue repair, ambience continuity, or first-class split edits. | J-cuts the next idea, L-cuts voice over proof, smooths room tone, preserves breath and tactile click/spray/keyboard as evidence. | Five-part sound scene: dialogue, ambience, native product sound, music, sparse SFX—with independent A/V in/out. |
| Captions | Real word timing and art direction; limited languages, fixed patterns, and a generic 9:16 safety model. | Text adds the number, objection, or context voice cannot. It is not karaoke. Placement responds to face, product, UI, and device. | Semantic captions + term review + face/product/UI collision tests in TikTok/Reels device previews. |
| Picture finish | Scale/pad, LUTs, and size checks; no guaranteed semantic reframe, shot match, skin/exposure gate, or stabilization path. | Corrects exposure/WB first, protects skin and product color, keyframes crop as attention moves, keeps useful phone texture. | Tracked reframe + shot match with manual override and an explicit “keep for trust / remove as defect” decision. |
| Proof + CTA | Can be requested in free text; not mandatory or verified in canonical cut artifacts. | Shows a result, use, comparison, receipt, or real reaction while the claim is made; CTA is earned by the payoff. | Proof ledger connecting every claim to a visible beat and checking CTA wording, visibility, and safe placement. |
04 · What winning social creative teaches
The numbers below are platform-reported advertising studies and creative-center observations. They are directional craft signals, not causal guarantees for organic performance.
TikTok reports the first two seconds account for half of recall and awareness impact. Its guidance favors vertical, sound-on, authentic work where every visual and audio element has a reason.
Official source ↗TikTok’s Creative Accelerator associates product-in-use footage with higher conversion and recommends live action, natural light, facial close-ups, a shot list, and audio.
Official source ↗Respondents said brand posts should not feel too polished; TikTok’s coded ad study reports casual, UGC-friendly work nearly doubled several early-watch and engagement outcomes.
Official source ↗Meta reports lower cost per result for Reels creative built in 9:16 with audio and key elements inside safe zones versus image ads in its cited tests.
Official source ↗TikTok’s True Classic example moves oddly satisfying hook → before/after → user validation in lived-in scenes → pack/promo CTA. Vaseline Verified built an entire repeatable format by testing community hacks, then visibly verifying or debunking them.
TikTok’s surfaced high-performing creative is repeatedly described through direct before/after, familiar scenes, curiosity leads, immersive close-ups, and scenario-specific customization—not generic cinematic finish.
Professional editors string out takes, mark reactions and improvs, select around emotional beats, and sculpt one consistent performance. Tiny looks and even a blink at a cut can change perceived credibility.
05 · The conversation experience
Today, the service counts files and duration, then asks generic brief questions. The north star is a content-grounded collaborator that names moments, flags risk, proposes a story, and makes every revision exact and reversible.
Telegram invokes the persistent montage session. The review composer only appends to an inbox relayed to the human owner, while the UI says the agent will reply. In delivered one-page mode, Chat and Comments are hidden entirely.
Fix:One event log and one conversation gateway. A message from either surface invokes the same session and projects the same reply to both.
Onboarding says the brief can be spoken, but a normal voice note becomes an audio asset unless a narrow revision state is active. It is not transcribed into the conversation.
Fix:Voice in intake/brief becomes a chat turn by default; only “use this as source audio” makes it media.
The state waits for approval, but the fragment can arrive with only a label—no “approve or say what is off”—and the model context omits the pending approval target.
Fix:Every fragment names the exact delta, locked spans, and next action. Add Undo and Compare.
Goal, audience, promise, offer, proof, CTA, must-use/avoid, compliance, authenticity tolerance, crop, brand, reference behavior, and variants are not durable structured state.
Fix:A confidence- and provenance-aware creative contract; ask only the single highest-impact missing question.
User “Hook is slow. Start with the bottle drop and lower music under speech.”
Agent “I’ll replace only 0:00–0:02 with upload 4’s bottle drop and lower music 4 dB under speech. Everything else stays locked. I’ll send the opening first.”
Fragment “Opening · 0:00–0:03. The drop now lands on the beat; music is lower under the first line. Approve or tell me what is still off. Compare V1 · Undo”
User “Approve.”
Agent “Locked. V2 changes only opening 0:00–0:02 and speech/music balance; all other frames match V1.”
“That’s all” and “start” depend on model interpretation. There is no deterministic done/start path, and a failed natural-language turn is not replayed despite recovery copy promising a later answer.
The user is told the message will be saved, but archiving happens only after successful registration. Save the raw event before reading mutable state.
Real editor narration exists, but there is no client-visible milestone plan, ETA range, next verified unit, cancel, or meaningful-delay receipt. Copy can imply proactive follow-up that is intentionally disabled.
The architecture separates client, preference, and project knowledge, but normal chat writes mainly preferences; identity, audience, provenance, conflicts, and project indexing can remain incomplete.
Resumption is solid, but there are no thumbnails, remove/cancel, duplicate warning, per-file retry, reorder, or exact error recovery; one fatal path can leave completion stuck.
Chat follows the user’s language while review UI is hardcoded English, including a hardcoded operator name. Ukrainian/Russian users cross a language and ownership boundary at delivery.
A 40-turn default can interrupt revision with no remaining-turn indicator or reserved approval/blocker budget.
Tests exercise contracts and source shape, not rendered mobile WebView, localization, accessibility, offline/reload, uploader failure, or the cross-surface conversation promise.
06 · Current result surface
Telegram promises a simple conversation and one result. Desktop opens a technical timeline with Comments, Library, and Chat; mobile switches to Cuts & trims, Visual effects, and Sound. The experience changes product category mid-journey.
“Vadym’s agent will reply here” hardcodes an operator identity and suggests shared conversational memory that the web path does not have.
The result moment should lead with “Does this story feel right?” and three grounded actions—not expose an NLE-shaped surface before the creator asks for control.
07 · Video quality and pipeline
Bind the approved master path and SHA to the round. Require full decode, expected fps/duration/audio, black/frozen frames, word-boundary joins, A/V sync, caption text/safe zones, face/product crop, loudness/true peak, and required story beats. Any failed check repairs or blocks delivery.
Per asset: content hash; codec/fps/VFR/HDR/rotation; audio health; shots; word-level ASR; faces, hands, product and action tracks; blur/shake/exposure; duplicate clusters; emotion/performance; native-sound events. Present contact sheets and ranked ranges, not just file count.
Manifest-driven strategies for testimonial, product demo, founder/coach, lifestyle, mini-vlog, event/travel, interview/multicam, screen recording, and mixed photo/video. Each cut records selection reason, story role, native sound, crop intent, and confidence.
Normalize CFR, pin one tested HyperFrames version, restore missing composition-library gates, use the proven capture path, fail on zero visual samples, and attach snapshot/motion checks to the delivered master—not only to documentation.
Dialogue isolation/denoise/de-reverb/EQ/de-ess, room-tone continuity, beat/phrase-aware music, tactile nat sound, sparse SFX, semantic reframe, stabilization, exposure/WB/skin/product match, multilingual term-aware captions.
Show the creative contract, story map, three hook candidates, and a low-cost proxy first. The user can lock a concept and important moments before a full-quality render. This reduces expensive random rerolls.
{
"master": { "path": "round-02/master.mp4", "sha256": "…" },
"story": { "promise": "pass", "visibleProofBy": "00:06.2", "cta": "pass" },
"picture": { "blackFreeze": "pass", "faceProductCrop": "pass", "safeZones": "pass" },
"sound": { "wordCuts": "pass", "dialogue": "pass", "loudnessTruePeak": "pass" },
"captions": { "scriptDiff": "pass", "collision": "pass" },
"status": "pass"
}08 · Concrete output examples
27s product testimonial
Best for: beauty, wellness, home, apps, small consumer products.
Human choices: first frame already has product motion; strongest skeptical line from any take; keep the eye-roll or laugh; L-cut voice across the close demo; preserve the snap/spray/click; hold 8–12 frames after genuine relief.
Avoid: logo card, karaoke captions, every-cut whoosh, unmatched before/after, compulsory LUT, or CTA before proof.
30s founder / coach insight
Best for: SaaS, creators, education, professional services.
Human choices: moving screenshot in the first frame; hard cut on eye contact; J-cut the next phrase under proof; slow down for the receipt; carry clicks/keyboard as native sound; select the confident microexpression, not only clean diction.
Avoid: “stop scrolling,” counterfeit reply-to-comment UI, unreadable screen recording, or “follow for more” unless follower growth is the actual goal.
30s lifestyle / mini-vlog
Best for: travel, food, events, fitness, transformations.
Human choices: keep the spill, miss, laugh, or doubt that creates stakes; non-metronomic 0.4/0.8/1.3s setup holds; cut on movement and native sound; alternate close/medium/wide; give the reveal 1.5–2.5s and include the reaction after it.
Avoid: equal clip lengths, file order, stabilizing every honest wobble, crash zooms on a timer, or looping a motion that destroys comprehension.
09 · North-star product
Telegram, uploader, result page, timestamp comment, and voice all become contextual events in one montage session.
“I chose take 3 for eye contact; kept clip 6 only for its close-up; proof appears by 0:06.” Users can correct facts and taste.
Conversion / Organic / Energy—or three hooks on one locked body—with clear differences and a fast compare path.
Frame, word, scene, upload, and subject references; unchanged spans lock; every action has a diff, Undo, and Compare.
Receiving → validating → analyzing → selecting → rough cut → captions/audio → QC → rendering, with ETA range and blocker.
Pace, shot density, caption energy, crop tolerance, nat-sound preference, zoom restraint, imperfection budget, and accepted revisions.
10 · Prioritized roadmap
Every final is artifact-bound and QA-passed; users see a footage-grounded plan before render; the benchmark can say whether output beats a baseline.
11 · Evidence trail
Method: traced the live upload, session, editor, assembly, delivery, review, revision, memory, and QA paths at commit cc9fc44; ran the repository’s full 888-test suite; opened the current review surface at desktop and mobile sizes; inspected engine skills, fixtures, and evaluation artifacts; benchmarked official TikTok/Meta guidance, public high-performing creative patterns, professional editing workflows, current AI editor controls, and primary technical research.
Interpretation rule: documented ability is not counted as a product guarantee unless the runtime produces durable evidence and the delivery gate requires it. Platform metrics remain attributed to the platform.
Final thought
It should be able to prove: “I understood what happened, chose why each moment belongs, protected what makes it human, and verified this exact file before I sent it.”
Back to top ↑