Floom Launch Film v13 — Holistic Audit

Full 69s cut · two independent judges (Gemini 3.1 Pro vision + Codex) · product-accuracy pass · 2026-07-01

Gemini 3.1 Pro · vision
65/100
Arc clear: YES
Strong narrative hook & clean UI, but the premium illusion is broken by webinar-style PiP and disjointed stock-like B-roll at the end.
Codex · gpt-5
56/100
Arc clear: NO (ending breaks)
Understandable founder-to-product arc, but lands closer to a polished AI explainer than an Apple/Fastlane launch film — the visual language fragments and the product section reads like a Loom demo.

Top 5 consolidated issues (both models, deduped)

1

Founder Picture-in-Picture through the whole product section

The founder shrinks to a bottom-right corner box from the name reveal through the entire feature walkthrough (~half the film). It reads as Loom/Zoom/webinar convention and destroys the cinematic premium register.

Beat L07–L12 · 21.4s–53.3s · Gemini P0 · Codex P0 · both, top issue

2

Ending B-roll whiplash — three films stitched together

Sterile light UI → dark moody office B-roll (L14) → hyper-warm golden-hour field-running (L16). The tonal jumps feel like stock footage with no causal link back to Floom; a first-timer loses the thread. Codex marks the arc as breaking here.

Beat L14–L16 · 55.6s–66.6s · Gemini P1 · Codex P0

3

Static dead-center orb reads templated / AI-slop

The blue orb sits dead-center and nearly motionless for the ~20s product stretch while cards pop around it. No camera choreography, no push-ins on the card the VO is naming; it becomes monotonous and machine-generated in feel.

Beat L08–L12 · 25.6s–52s · Gemini P2 · Codex P1

4

Word-by-word blue-highlight captions = TikTok template

Burned-in captions highlight the active word in blue, the hallmark of cheap Shorts/TikTok auto-caption templates. It undercuts the high-end feel and, during L08–L12, competes with the on-screen UI text.

Full film (esp. L01–L12) · Gemini P1 · Codex P1

5

Feature-dense beats drag (L10 & L12 ~8s each)

L10 crams autonomy + approvals + schedules + triggers into one 8s span; L12 stacks a six-logo ring + zero-lock-in card for another 8s (reads like a pitch-deck slide). Both drag and blur the single idea each beat should own.

Beat L10 · 32.2s–40.7s & L12 · 45.5s–53.3s · Codex P1 · Gemini (monotony) P2

Per-model issue lists

Gemini 3.1 Pro

P0L07–L12 · 21.4–52s

Founder PiP box screams Loom tutorial / Zoom webinar; destroys the Fastlane/Apple aesthetic.

Fix: Remove the PiP; let VO carry the UI beats or hard-cut to full-screen founder for emphasis.

P1L14–L16 · 55.6–64s

Visual whiplash / stock-footage feel: sterile UI → dark hacker room → golden field = three different videos.

Fix: Grade the B-roll to match the warm founder room; integrate the orb into the live-action to tie product to the real world.

P1L01–L12 · 0–52s

Word-by-word blue-highlight captions read as cheap TikTok/Shorts template.

Fix: Elegant phrase-based fade-in subtitles, clean sans (SF/Helvetica), subtle white/grey, no bouncy highlight.

P2L08–L12 · 25.6–46s

Product UI is monotonous — orb dead-center 20s+ while cards pop; lacks Apple-ad camera work.

Fix: Subtle 3D camera drifts; push in on the Sandboxed / Schedule card when the VO names it.

Codex

P0L07–L12 · 21.4–53.3s

Founder PiP for ~half the film turns the launch reveal into a webinar/tutorial convention.

Fix: Remove PiP after the name reveal; full-frame product choreography with occasional founder cutaways.

P0L14–L16 · 55.6–66.6s

First-timer loses the thread: dark office → golden field jumps pain→lifestyle with no causal link to Floom.

Fix: One coherent human payoff (saved time → real life), graded in the same system as founder/product.

P1L08–L12 · 25.6–53.3s

Dead-center static orb makes it feel templated / AI-generated; card pop-ups repetitive.

Fix: Choreograph the orb as an active agent moving through prompt→approval→schedule→log→providers.

P1L10 · 32.2–40.7s

Too many concepts in one 8s span (autonomy + approvals + schedules + triggers) = feature density.

Fix: Split into two sharper beats (approval-trust; scheduled/event wakeups), each with distinct action, cut faster.

P1L12 · 45.5–53.3s

Provider-logo ring + zero-lock-in card at 8s risks looking like a pitch-deck slide.

Fix: One premium portability metaphor; reveal only the essential logos briefly as proof.

P1L05–L07 · 17.8–25.6s

Strongest line ("The vision was right. The infrastructure wasn't") cuts to text-card/reveal before the pivot lands.

Fix: Let the founder own the first sentence on camera, then the text card as an impact frame before the reveal.

P2L15–LCTA · 59.4–69.4s

Three brand/end beats stack (This is Floom / Time is life / floom.dev) — ending feels stacked, not inevitable.

Fix: Make "Time is life" the emotional close; final card shorter, quieter, visually continuous.

Product accuracy — L08–L12

No misrepresentations found. The product beats track Floom's real capabilities.

L08 · platformWorker cards (Signup Digest, Outreach Drafter, Weekly Report) around the orb, "a platform for autonomous background agents that actually work." — accurate to "spawn/run background AI workers." OK
L09 · spawnPrompt-input card ("Summarize today's…") for "spawn agents with a single prompt." Plausible spawn-from-prompt UI. OK
L10 · run/approve/scheduleApproval card ("Approve if drafted replies?" Approve/Reject) + schedule card ("Every weekday 9:00 AM"). Matches run-autonomously / ask-approval / schedule & event triggers. UI plausible. OK
L11 · logged/secure/sandboxedLog rows + "Sandboxed" card for "every run is logged, secure, and sandboxed." OK
L12 · model-agnostic + OSSLogo ring reads as Anthropic, OpenAI, Gemini, Meta, Llama, Mistral — all legitimate model providers, none misrepresented. "Zero lock-in / Switch models, hosts & clients any time" card + "fully model agnostic and open source." No fake or off-brand logo. OK

Verify with Federico (not a defect, just confirm): (1) the six providers shown match the currently-supported set (Mistral + Llama both shown — confirm both are real integrations, not aspirational). (2) "available via WhatsApp" (L02) is stated as part of the OpenClaw promise that broke — confirm the film does not need to also assert WhatsApp for Floom, or it may read as an implied Floom feature.

What's working (both models)