1080×1920

Instagram Reel

Audience retention

100%75%50%25%0

{{ R.retHoverPct }} watching at {{ R.retHoverTime }}

0s{{ R.q1Fmt }}{{ R.midFmt }}{{ R.q3Fmt }}{{ R.lenFmt }}

Drop-off heatmap

0s{{ R.midFmt }}{{ R.lenFmt }}

Holding Dropping Lost

DEATH MOMENTS · biggest post‑hook drops · what's on screen & said

Engagement breakdown

Engagement rate {{ R.engRateFmt }}

{{ e.value }} · {{ e.rate }}

Engagement rates

% of views

Skip & repost rates + Higher/Lower benchmarks are read from your Instagram insights recordings (the Graph API doesn't expose them).

Reel audit

every element checked against your Lab's measured best practices

Auditing against the Lab…

No audit yet — this reel isn't in the measurement DB (no stored transcript/frames). Extract it first, then the audit fills in automatically.

AI breakdown

VISUAL · 45%

CREATOR · 55%

Stufe {{ R.dv.curFrame.intensity }}

FRAME {{ R.dv.counter }}

🎞 {{ R.dv.curFrame.motion }}

GESPROCHEN

ON-SCREEN TEXT

UP NEXT →

0s {{ R.dv.curTime }} {{ R.dv.totalTime }}

VISUAL-RHYTHMUS · geschätzt

FILMSTRIP

{{ R.dv.selCount }} selected · click to toggle, then Download

Hook & drop-off

Suggestions

Transcript

Turn any reel into script & visuals

Paste an Instagram reel to deconstruct it frame‑by‑frame — then scrub the footage, rewrite the script, and generate a ready‑to‑paste Claude design prompt.

▶

{{ CC.frameCount }} frames

9:16

VISUAL · 45%

CREATOR · 55%

Stufe {{ CC.curFrame.intensity }}

FRAME {{ CC.counter }}

🎞 {{ CC.curFrame.motion }}

GESPROCHEN

ON-SCREEN TEXT

UP NEXT →

0s {{ CC.curTime }} {{ CC.totalTime }}

VISUAL-RHYTHMUS · geschätzt

FILMSTRIP

{{ CC.selCount }} selected · click to toggle, then Download

ORIGINAL SCRIPT · SPOKEN

{{ CC.transcript }}

NEW SCRIPT · AI

{{ CC.gscript }}

PER-FRAME VISUALS

Scene‑by‑scene descriptions from the footage. These feed the prompt on the right.

✳

Claude Design Prompt

SCRIPT-QUELLE

{{ CC.assembledPrompt }}

Frames · {{ CC.totalTime }}

Talking head

UI / B-Roll

Ø Visual-Intensität

Visual-Intensität über die Zeit

Geschätzte Stufe 0–4 pro Szene (aus Bewegung + On-Screen-Text). Klick öffnet den Frame.

0–1 Ruhe

2 Support

3 Hero

4 Scroll-Stopper

Hook

Pacing

Höhepunkt

SUGGESTIONS

Animations‑Breakdown · Reverse‑Engineering

Bricht das extrahierte Reel Szene für Szene auf: die echten Frames, was animiert wird, und eine Spec, um jeden Frame sauber als Mockup nachzubauen.

Erst ein Reel im Frames‑Tab extrahieren — dann brichst du hier dessen Animationen Frame für Frame auf.

„{{ sec.narration }}"

ANIMATION BRIEF

ANIMATIONS-FLOW · {{ sec.frameCount }} FRAMES

{{ fr.n }}. {{ fr.caption }}

→

Hit Break down animations — the reel is split into scenes with its real frames + an animation brief each, plus a one‑click clean mockup.

Studio

Give a topic. The advisor learns the patterns that actually drove engagement on your reels, writes a script grounded in your winners, and you refine it in the chat. It cites the real reels it learned from.

REEL BRIEF

OPTIMIZE FOR

◆

Your winning formula

›{{ f }}

LEARNED FROM

{{ ref.shortcode }} {{ ref.likes }}♥ · {{ ref.comments }}💬

◆

Enter a topic on the left and hit Generate.
The advisor pulls the patterns from your reels and writes a grounded script — then refine it here in chat.

Thinking…

GENERATED SCRIPT

{{ STU.script }}

HOOK VARIANTS

VISUAL-INTENSITY PLAN

Lab

One pipeline, measured end to end: 1 what's on screen each second and whether it holds viewers → 2 the exact moments they leave → 3 what your opening lines do → 4 whether held attention converts to views. Trained on your frames + your exact retention, cross‑validated by reel. No LLM opinions.

{{ LAB.attSeconds }} per‑second observations · {{ LAB.attReels }} reels · group cross‑validation (no reel in both train & test)

Training the attention model on your per‑second data…

1 · ATTENTION MODEL gradient‑boosted trees · per‑second drop rate · cross‑validated by reel

MODEL ACCURACY

out‑of‑sample R² on per‑second loss

TIME‑DECAY ALONE

knowing only the timestamp · content adds {{ LAB.attLift }}

AVG LOSS RATE

{{ LAB.attHazard }}%/s

of remaining viewers, each second

WHAT HOLDS vs LOSES VIEWERS · green holds · red bleeds · per‑second effect, all else equal · levers the model barely uses (<0.5% weight) are hidden as noise

BLEEDING MORE THAN CONTENT PREDICTS · something unmeasured is hurting

HOLDING BETTER THAN PREDICTED · rewatch to find the hidden edge

Per‑second effects compound: at your average loss rate, 30 seconds keeps ~18% of viewers — hold one green edge of −0.7pp/s over that stretch and ~22% survive, a quarter more audience. The two lists flag reels where something the model can't see (delivery, topic, audio) is helping or hurting — those are worth rewatching.

2 · DEATH MOMENTS where viewers actually leave · exact per‑second retention × what's on screen

The 0–3s scroll‑off is universal. What's actionable is the second wave — the biggest drop after the hook — and which visual CUTS cost more retention than the reel's own natural decline. Every reel's own death moments are on its detail page.

Finding your death moments…

AVG HOOK SURVIVAL @3s

{{ LAB.dropHook }}%

still watching after the scroll‑off

BIGGEST POST‑HOOK CLIFF

where most reels take their next hit

WHICH CUTS COST RETENTION · n = how often the cut appears · −pts = extra retention lost vs the reel's own trend

⏳ Scene‑cut costs appear once the structured visual backfill finishes.

STEEPEST POST‑HOOK CLIFFS · click to open the reel

WHAT MAKES THEM SWIPE · skip‑rate model · LOO R² {{ LAB.skipR2 }} · n {{ LAB.skipN }} · effect in skip‑rate points · green = fewer skips

3 · HOOKS your opening lines, measured · what is SAID in the first 3.5s · {{ LAB.hookN }} reels

Deterministic patterns (regex on the spoken opener — no AI opinions) contrasted on the two things a hook controls: the swipe and the 3‑second hold. Descriptive averages at small n — where they agree with the models above, trust them.

TOPICS · keyword‑detected, descriptive only — topic hurt the predictor when ablated (thin cells) · ⚠ = under 5 reels, read as anecdote

TOPICNMED VIEWSSKIPHOLD @3s

OPENERS THAT HELD · lowest skip · click to open

OPENERS THAT GOT SWIPED · highest skip · click to open

4 · VIEW PREDICTOR content → retention → engagement, stacked · leave‑one‑out CV · maturity‑controlled

CONTENT ALONE

what you control up front

+ RETENTION

+ hook hold, completion, skip

+ ENGAGEMENT

+ saves/shares/comments per view

VIEW GROWTH

WHAT MOVES VIEWS · standardized effect, each controlled for the rest

R² is out‑of‑sample (leave‑one‑out): 0 = no better than guessing your average, 1 = perfect. Content covers the deterministic script + format + posting features (regex over your transcripts, counted frames, the timestamp). The engagement tier only exists after posting and its rates use views as denominator — read those coefficients as quality signals, not levers you can pull in advance. Too‑fresh reels are excluded so unsettled view counts don't skew it.

Everything above is measured and honestly cross‑validated — strong enough to rank what to try next, not to prove cause. The proof is changing ONE lever in a new reel (e.g. cut the stale scene at its predicted death moment) and watching its curve beat the model's prediction.

Performance overview

Daily reach & plays

Engagement mix

Avg engagement rates

What's working

Needs attention

All reels · {{ reelCount }}

{{ R.cap }}