emfMax ad sprint
4 video ads (2 B2B, 2 B2C) plus a refreshed static set, built on 12 Oct 2026 by 5 parallel AI helpers from Max's own feedback. English only, lead-form destination, no health claims.
The videos
B1 · Wellness retreats & bio hotels
"Some of your guests will ask, is my room low-EMF? It's a fair question. It helps to have a real answer…"
Meta copy & notes
Primary: Adding a low-EMF suite to your retreat or hotel? We measure the site, plan the shielding with your architect, and give you a spec and a quote your contractor can build from. Start with a free 15-minute chat with Max. Based in Singapore.
Headline: Low-EMF rooms, measured properly · CTA: Get quote
QA: sync ≤0.9 frame, captions 73/73 match, −14.1 LUFS. Open: Max's studio portrait over the villa backdrop (an on-location photo would be better); confirm the post-build re-measure service.
B2 · Equipment interference (EMI)
"Replaced it twice, and the fault's still there? … It could be interference. Could be nothing. Check it before you order more spares."
Meta copy & notes
Primary: Replaced the same part twice and the fault is still there? It could be interference. Could be nothing. We come on site, measure it properly, and give you an executive summary, the readings and an action plan with cost ranges. Based in Singapore.
Headline: Measure it before you replace it · CTA: Get quote
QA: 35/35 SFX on frame, captions 71/71, −14.2 LUFS, 855 frames scanned clean. Open: brand names on Max's meters (readouts blurred); on-site in MY/TH unconfirmed.
C1 · New home, "before the cot goes in"
"Before the cot goes in, measure one more thing. Once the reno's done, you won't really shift the cot, bed or desk…"
Meta copy & notes
Primary: Just got your keys? Before the cot goes in, get your home's EMF readings done. We come down, check every room and give you a written report with photos and what we'd suggest. Start with a free 15-minute chat with Max.
Headline: Measure the room before the cot · CTA: Get quote
QA: 27/27 SFX within 1 frame, captions 0 mismatches, −14.1 LUFS. Your key-shot note is fixed (line-art key and door). 4:5 is the blurred-background version.
C2 · Substation next door
"Got a substation next door? … If moving the bed is enough, we'll tell you."
Status
Stopped at wrap-up. This is the first full draft (loudness-normalised for preview). Known fixes not applied: the offer label above the headline, small captions, a bigger real-meter moment, the transparent logo, Max on the end card. No 4:5 yet. About one revision round plus a final render from done.
Static set (Max's 12 Oct feedback)
Dropped angle 4 and the Mandarin c5, merged the duplicate new-home angles, added 2 wellness-centre B2B angles (c11, c12), and rewrote all copy in plain spoken English. Real meter photos and transparent logos throughout.
Cost per video (SGD)
USD converted at S$1.29 per US$. Google = the broker ledger (Veo clips plus Gemini checks). OpenAI = GPT Image 2.5 stills; the API returns no price, so these are estimates. Claude = Opus 5.5 tokens priced at API list rates (US$4/M input, US$20/M output, US$0.20/M cache read, US$8/M 1-hour cache write). On your Claude subscription the real marginal cost is usage allowance, not cash. ElevenLabs runs on subscription credits, so it isn't cashed out.
| Item | OpenAI | Claude (API-equiv.) | Total | |
|---|---|---|---|---|
| B1 wellness | S$15.74 | ≈S$1.94 | S$52.07 | ≈S$69.75 |
| B2 EMI | S$16.06 | ≈S$1.48 | S$45.39 | ≈S$62.93 |
| C1 new home | S$15.70 | ≈S$1.94 | S$45.90 | ≈S$63.54 |
| C2 substation (draft) | ≈S$8.11 | ≈S$1.29 | S$23.18 | ≈S$32.58 |
| Static set (15 images) | S$1.11 | S$2.53 | S$15.66 | S$19.30 |
| Orchestration (this chat, incl. 5–12 Oct onboarding) | n/a | n/a | S$73.08 | S$73.08 |
| Total | S$56.72 | ≈S$9.18 | S$255.29 | ≈S$321 |
Cash actually spent outside subscriptions: about S$66 (Google + OpenAI). Per finished video, excluding Claude: about S$17–18.
Claude token breakdown
| Agent | Output tokens | Cache read | Cache write | API-equiv. |
|---|---|---|---|---|
| B1 | 74,988 | 49.0M | 3.63M | US$40.37 |
| B2 | 71,182 | 81.0M | 2.19M | US$35.19 |
| C1 | 41,710 | 41.7M | 3.30M | US$35.58 |
| C2 | 50,459 | 41.6M | 1.08M | US$17.97 |
| Statics | 32,596 | 41.9M | 0.39M | US$12.14 |
| Orchestrator | 184,768 | 119.6M | 3.63M | US$56.65 |
| Total | 455,703 | 374.8M | 14.2M | US$197.90 |
Uncached input is negligible (under 3k tokens in total). Most of the cost is re-reading context (cache reads), which is how long agent sessions work. The orchestrator row covers the whole chat since 5 Oct, not only today.
How the system carried it
| Piece | What it did |
|---|---|
| Parallel helpers | 5 Opus agents (4 videos + statics) worked at the same time, each with its own folder, while the orchestrator audited 6 rounds and sent 40+ steering messages. |
| /follow-up | Pulled the proven pipelines from the UltraRevive, Cathy and Renobros chats (code, voices, sync rules, your feedback) in about 1 minute. |
| Gemini / Veo | About 41 Veo 3.1 clips for environments; Gemini ran script naturalness checks, blind voice ranking and harsh critiques. |
| OpenAI GPT Image 2.5 | About 50 photoreal first frames and plates (Codex was mostly busy, so the paid API was used). |
| ElevenLabs | Approved voices, about 35 voice takes, about 21 music beds, 140+ bespoke SFX, and speech-to-text caption verification. |
| Real client assets | Max's portraits cut out with only tonal edits; 9 of his real meter product shots background-removed; transparent logos rebuilt. |
Cloud compute
| Ad | Cloud jobs | Done | Cancelled | Machine time* |
|---|---|---|---|---|
| B2 | 22 | 20 | 2 | ≈249 min |
| C1 | 17 | 13 | 4 | ≈170 min |
| B1 | 10 | 5 | 5 | ≈98 min |
| C2 | 6 | 5 | 1 | ≈20 min |
| Total | ≈55 | ≈43 | 12 | ≈540 min |
*Job wall time as reported by the job service, including queue wait, so treat it as an upper bound. Each cloud machine is 4 CPU / 16 GB. We hit the 30-starts-per-hour cap once, because B2 split every render into 5 machine starts and re-rendered sound tests. After that, review drafts went to 540p single jobs and 4:5 versions were made locally. This server stayed light: the chat used about 0.6 GB of RAM, with roughly 2.8–9 GB free across the box during the run.
Honest comparison
| Scenario | Estimate | How |
|---|---|---|
| This run (parallel + cloud) | ≈3 h wall-clock to 3 finished videos + statics | Measured: helpers launched about 13:30, last final at about 16:30 SGT. |
| Same agents, one after another | ≈11–13 h | Sum of each agent's active time (about 2.5–3 h each × 4) plus statics. |
| Rendering on this box only, no cloud | + ≈6–9 h of renders, with high risk of memory kills | About 540 job-minutes run serially on a shared 15 GB box, where only one heavy Chrome render is safe at a time. |
| Human motion designer | ≈4–8 working days | About 1–2 days per polished 25 s ad with VO, SFX, captions and 2 formats, plus the statics. |
What didn't go well: 2 session restarts interrupted the helpers; B2 over-used cloud starts early on; C2 stalled on a cancelled render and was stopped as a draft; and the review loops ran longer than needed before you capped them.
Before it goes to Max
- Max picks the angles; nothing is uploaded to Meta.
- Singapore advertiser verification (CHOIZYA PTE. LTD.) must show Verified before launch.
- Confirm with Max: the brand names on his meters, the post-build re-measure (B1), on-site work in MY/TH (B2), and the "we'll tell you straight" wording.
- C2 needs one more round if you want it in the first test.