Back to the brief

Since this was written (MISC, 7:22pm Friday): Saturday's budget moves were held at 7:42pm, so the next automatic move is Monday 19 October at 12:01am Pacific. The plan: make all 21 (9:48am Saturday) and launch them (3:00pm), each only once its video is OK'd here (3:40pm). The music comes from Meta's and TikTok's free libraries; on Meta it has to be mixed into the file, and that call is in Questions.

Batch 4 TikTok slideshows as Meta ads: what the Meta ads committee says

Last updated: Friday 16 October 2026, 7:22pm Pacific, by META MISC (one of Mike's Claude Code sessions), for FM133. Nothing went live and nothing changed in Meta. This is advice for Mike; CREATION puts it in tonight's brief.

Mike, 5:40pm Pacific: "I believe pretty much all of those slideshows are EXCELLENT and could sell on Meta ... i think these are better than any ads we have so far except founder videos." And 5:46pm Pacific: "i do think these do a better job of selling the products than any still image ads we're running or are considering running".

The bottom line

• The committee agrees with Mike, and none of the seven disagreed. Two said yes outright and five said "partly". The five agree the best of the batch should beat most of our still image ads (Denney: as videos, not as carousels). Everyone who spoke to the founder videos keeps them on top (Denney, Robinson, Sanchez; Melwani sees no basis yet to rank the slideshows above them). Their hold-back is that Meta spending is the proof, and nothing has run yet. The bottom six (N5, S5, B5, S4, B7, B3, ranked 16 to 21) got no video votes; Cooper puts five of them below our better stills, and Robinson and Sanchez would not run them as built.
• The seven to run first, best first: N4 off switch, S1 cells eat themselves, B6 see it in your blood, B1 running on empty, N2 why so cheap, S3 fast or dinner, A3 scoreboard. N4 was first for five of the seven members, and N4 and S1 were in every member's top five.
• Can Claude do the animation? Yes, shown tonight. A Claude session animated S1 in code in 28 minutes. The words build in, the medal drops, the junk piles up and the pouch lands, and every slide ends exactly on its approved picture. Watch it: https://vimeo.com/1236724505/b350849450 (24 seconds, 4:5, with a music track made in code). All seven members said "run it with changes". Test 2 has their main changes, with a Higgsfield clip of the cell taking a real bite: https://vimeo.com/1236730857/fc54d33ab1 (21 seconds, 4:5, ending on the offer card). The code-only cut, without the bite, for comparison: https://vimeo.com/1236731366/f619b71d34.
• Can Claude do the music? Only partly. It made the background track in that video with code, but nobody has heard it yet (Claude can't hear audio). The best route is free tracks from Meta's own Sound Collection, chosen by mood and tempo and cut to the animation by Claude. Mike hears them on the picks page before he picks, and nothing goes to Meta until he does. A paid library is the upgrade if he wants more choice (from $299 a year).
• The best way to make all 21: code for every word, number, chart and pack, rebuilt from the build files Nick's team sent. It takes about 3 hours for all 21, in Feed and Reels shapes, with and without music, including one review round. It costs nothing beyond Claude time, needs nothing from Mike or Nick, and this session would make them.
• The one choice in the route: code moves a photo object only as a whole piece (slide, pop, zoom, squash). To make one really change shape, like S1's cell taking a bite, takes a Higgsfield clip, and the committee preferred adding those. But Higgsfield's Kling files carry an "AIGC" tag naming Kling as their maker, and Higgsfield's terms (section 5.5) say not to remove markings Higgsfield applies, which may cover it. Meta isn't known to read that tag, but it is an AI marker in the file. Code alone has neither question. Mike and Nick's call, question 1.
• Still image ads: add the slideshows alongside them first, then switch off the stills they beat. Six still ads come off once their slideshow version beats them on new-customer sales per $1, in this order: SPMD-99 "fasting benefits" plate, NMN-99 "back to 25" beakers, SPMD-99 "six more years" stat, BROQ "test or refund", BROQ lab comparison, SPMD-99 "pure source". The other seven still ads stay.
• Before any of them go on Meta, six posts need a few words changed to match Mike's settled lines (S4's retired "15 mg" line, "We bought" in B2 and A3, B1's "sold in the US", and others), and B6 needs two picture fixes (table below). The Meta copies also need to be saved clean: every slide file carries a record naming Claude as its maker, and Meta reads that record to add its "AI info" label.

Questions for Mike

1. Make the animated videos? - A (recommended): yes, all 21, code only, onto the picks page tomorrow, both shapes, with and without music, after the wording fixes. No credits, no AI marker in any file. Photo objects move as whole pieces but don't change shape. The committee gave N5, S5, B5, S4, B7 and B3 no video votes, so they'd sit at the bottom of the picks page. - B: yes, all 21, code plus a Higgsfield clip wherever an object should really move (S1's cell biting), as the committee preferred. About $9 to $13 of credits. The 218.75 credits left cover 25 clips; past that it needs more credits, which only you can buy. The final files would either leave out Kling's AIGC tag, which Higgsfield's terms may bar removing, or keep it, which Meta isn't known to read but leaves an AI marker in the ad file. - C: the seven above first, either way. - D: wait until you've watched test 2 (both cuts). 2. Music: - A (recommended): Meta's Sound Collection, which is free and is the library Meta itself uses for ad music. Claude picks and cuts the tracks, and you hear them on the picks page. (Meta's pages don't spell out baking a track into an uploaded video; the Sound Collection page says "Free to use in any videos you create and share on Facebook and Instagram".) - B: buy Epidemic Sound Business ($75 a month, or $360 a year). It covers a brand's own ads, and ads published while subscribed stay covered after you cancel. - C: Artlist Pro ($299 a year, yearly only). 3. First test (the seven posts in section 5, in the four live ad sets, budgets unchanged): - A (recommended): yes, once the wording is fixed and you've picked them. - B: carousels only at first. - C: hold.

1. Do they agree with Mike, and how they rank all 21

The vote: 2 yes (Jared Robinson, Savannah Sanchez), 5 partly (Alex Cooper, Ash Melwani, Dara Denney, Sarah Levinger, Stefan Georgi), 0 no. Nobody said "No opinion on this" here.

What "partly" means, in their words:

• Cooper: "On message, Mike is mostly right." The best of these carry a whole argument where most stills carry one line, but "'better at selling' is decided by spend carried and new-customer return, not by how good an ad looks to us." BROQ "winter ready" and the skin and hair stills speak to people no batch post speaks to.
• Melwani: "these posts are built the way I build supplement ads for cold traffic", but there's no basis yet to say they beat the founder videos.
• Denney: as videos, yes; as carousels, no ("more of a waste of time than anything" is her line on carousels). Founder videos stay top.
• Levinger: yes on how they sell (a "Wait... what?" opening, a named enemy, then proof), but expect five to nine of the 21 to carry and the rest to get little spend.
• Georgi: the concepts are strong and the bar is low, but "I wouldn't call it until they win in the account."

Their point of reference: the 13 live still ads have brought in $595.70 of new-customer sales on $955.82 of spend ($0.62 per $1, 6 orders), as read Thursday 7:28am Pacific. Six members used that number.

The ranking (points across all seven lists: 21 for first, 1 for last):

B3 came last on all seven lists ("How do you say BROQ?" only stops people who already know BROQ).

Why the top seven (one line each from the members):

• N4: one picture, an off switch in your cells, carries problem, cause, fix, a randomized trial and the pack in five cards (Cooper, Levinger, Melwani).
• S1: a surprising true fact, the Nobel Prize as proof, and "Without skipping a meal." as the payoff for anyone who has tried fasting (Levinger, Denney, Sanchez).
• B6: turns the question every supplement buyer has, "how do I know it's working?", into the ad and answers it with a blood test and the guarantee; the best fix for BROQ, the weakest ad set this week (Melwani, Denney, Levinger).
• B1: an empty fuel gauge and "Over 50%" stop the scroll, then the 20-brand lab proof (Melwani, Georgi).
• N2: answers a question every NMN shopper has, with a sourced number and our own tested amount, and lands on the page of our best NMN still (Georgi first, Melwani).
• S3: two plates, two choices, understood in one second; every line is already live on the fasting still (Denney, Georgi, Sanchez).
• A3: three head-to-head lab wins on a scoreboard, made for rolling numbers, once its three wording problems are fixed (Levinger, Denney).

Several posts tell the same idea (Cooper, Denney): N4, N1, N3 and N6 are NAD+ falling with age; S1 and S3 are autophagy without fasting; B1 and B2 are the 20-brand lab test; A3 and A2 are the lab head-to-head. Cooper and Denney rank the best version of each idea first and keep the others as later variations, because Meta groups similar ads together. Every member's reasons for their own top 10 are in committee/.

How the posts were judged: six Opus reviewers looked at every slide first, one post at a time (reviews/, summary in DIGEST.md). Then the seven members judged from those reviews, the slide strips and the S1 test. Nine posts (B1, N2, A2, B3, B4, B5, N1, A3, B6), four of N3's five slides and all of B7 had not synced to this Mac. Those were judged from the one-sheet overview plus the build files, which hold every word exactly as printed. That includes four of the top seven (B6, B1, N2, A3). All of them synced at 7:03pm Pacific, and a separate full-size check (reviews/FULL-SIZE-CHECK.md) found every word matches its build file, no score changes and no reason to move any of the top seven. It found one thing the overview hid: on B6's slide 1, the red in each blood tube is a flat band floating mid-tube, which looks fake at Feed size. It also corrects a few square-crop measurements (B1, A3, B3, B4) and notes small spacing fixes (N2's "NAD+ ." and "99+ %", N3's "100 %", B5's slide 1). Every member was told up front that Claude sessions would do the animation automatically, taking none of Mike's or Nick's time.

2. As Instagram and Facebook carousels

Which: N4, B6, N2 and S3 (all seven; Denney only as a later format behind her videos), then B1, A3 and S1 (all but Denney) and A1 (all but Georgi).

What Meta needs (from the reviewers; most members said "No opinion on this" on sizes and placement, so these are the reviewers' calls):

• First card: most posts keep their own slide 1. Three change order: B7 leads with its lab-stamp card, B3 with its "3.5 pounds of raw broccoli" card, and S5 with its proof card or a line naming spermidine.
• "Highlight carousel card" off for every post that tells a story in order (all but A1 and B7), so Meta can't move a later card to the front.
• Sizes:

- 4:5 for Instagram and Facebook Feed. The TikTok size works as is, but the empty bottom fifth left for TikTok's caption should be used or cropped. - 1:1 for Facebook Feed carousels. Several posts need the square cut from the top or a line moved; the reviews give the pixel rows. - 9:16 for Stories and Reels, with nothing important in the top 14% or the bottom 35%. - Text size: every post clears the 24 px minimum in our ads knowledge base (Robinson, the reviewers). Darken the grey lines on slide 3 of B2, B3 and B4 and B6's footnote, keep S2's citation at 28 px or more, and raise B5's subline to about 40 to 44 px.

• End card: every post gets one, with the live offer and guarantee from its landing page. For example SPMD-99: "Save 15% + free shipping" and "180-day money-back guarantee. No need to send anything back.", with the annual-plan small print.
• Landing page: one live page per post, matched to what it says. For example N4 goes to nmn-99-back-to-25, B6 to broq-test-or-refund, S1 and S3 to spmd-99-fasting-benefits, and A3 to lab-tested. Every post's page and headline are in DIGEST.md.
• Text: the primary text is the TikTok caption without hashtags. TikTok-only settings come off. Advantage+ enhancements that crop, rewrite or animate the cards stay off.
• Meta ad policy: one real risk in the whole batch. N3's card 1, "Your NAD+ at 50." over a half-empty battery, speaks to the viewer's own body at a stated age. The three-ages fix below removes it. B6's "you can see the results in your blood test results" is low risk as long as its footnote stays on the same card. Nothing else.

Words to fix before Meta (checked against Mike's settled rules; these are what would need changing for Meta, and nothing here was raised about TikTok):

Clean files for Meta: I checked the slide files myself. Each one carries a Content Credentials (C2PA) record naming Anthropic and Claude, and Meta says it reads exactly that record to put "AI info" on ads. Meta copies get saved without it, per Mike's rule that no ad is labelled as AI. The Nano Banana photos inside some slides (the cart in N2, the fuel gauge in B1, and others) go through the usual watermark removal first (_SYSTEM/GEMINI-WATERMARK-REMOVAL.md).

3. As videos with the elements animated and music

Which: S1, N4, B1, A3, B6 and S2 (all seven members), then B4, N3 and N2 (six of seven). N3 and B4 make better videos than carousels: the draining battery and the capsule riding a transit map need motion.

How long: 15 to 22 seconds, plus a 15-second cut for Stories. Make 9:16 the master for Reels and Stories, with a 4:5 cut for Feed.

How it should move (agreed by the members who looked at the S1 test):

• The hook is fully readable on the very first frame, and motion comes after the words (Denney, Georgi, Levinger).
• One object really moves in the first two seconds: S1's cell takes a bite, N4's switch flips, B1's needle sags to empty (Levinger, Cooper).
• Shorter holds: test 1 holds about 3 seconds after each build, which Cooper and Georgi found slow. Cooper and Levinger want about 18 to 20 seconds in all, Robinson about 18, Sanchez 15 to 18.
• It ends on an offer card in the landing page's own words.
• Every post's review has a second-by-second plan: what counts up, what pops, what fills. For example A3's scoreboard digits roll like an odometer and stop at 25, 1,139.6 and 21.9 mg, and B6's 90 dots fill one by one.

What music: six of seven said "No opinion on this." Savannah Sanchez: no AI music; a track chosen by ear, built into each video so the beats land on the cuts; for carousels, a song chosen by hand from Meta's Sound Collection at setup, never Meta's automatic pick. The reviewers' moods per post are in DIGEST.md, for example warm minimal electronic at 85 to 95 BPM for N4, and a stadium anthem at 120 to 128 BPM for A3.

4. Mike's question: can Claude do the animation and the music, and if not, what can?

Animation: yes, Claude can, and did tonight.

• The test: S1 animated entirely in code, rebuilt from its own build file. Words pop in, the definition card types on, the medal drops and catches the light, the junk piles up, and the pouch lands with its shadow.
• It matches: the last frame of every slide matches the approved slide to about 1 colour level in 255, which the eye can't see (test/S1-match-check.jpg).
• How it looks: members called it crisp, clean and on brand (Cooper: "Crisp, clean, premium"; Georgi: "clean, on brand and expensive-looking"). The gaps were no words on the first frame, a cell that never moves, long holds and no offer card.
• Time: 28 minutes for both shapes, with and without music. Report: test/ANIMATION-TEST.md.
• Watch test 1: https://vimeo.com/1236724505/b350849450 (4:5, 24 seconds, with the code-made track).
• What code can't do: make the clay cell itself chew. Higgsfield's Kling 3.0 can. In a test it made S1's cell close on the junk and chew, true to the photo, for about 43 cents (test/HIGGSFIELD-TEST.md). Claude drives Higgsfield through its own tools, so this stays automatic too.
• Test 2 has the four changes most members asked for: the hook on the first frame, the cell taking a real bite, shorter holds, an offer card. Watch it: https://vimeo.com/1236730857/fc54d33ab1 (4:5, 21 seconds, with the code-made track). The headline is there from the first frame; the cell pulls the gears in, closes and chews between 1 and 2 seconds, smoothly at phone size; the offer card shows the whole pouch with the page's own lines. One thing to know: mid-bite, the cell's glassy glow and pink nucleus fade within about a quarter of a second and a few loose gears slip behind it; at phone size it reads as a gulp. It took three Kling tries (the first two warped) and 26.25 credits, about $1.29. A second cut moves the cell by code only (so it has the other three changes): a push-in, the junk flying in, a small squash and bounce. It's clean and calmer, because the cell never eats. Watch it: https://vimeo.com/1236731366/f619b71d34. Both cuts carry no AIGC, C2PA or Kling marker. One fix for the toolkit: the end card needs one more beat, since the guarantee lines are fully on screen for only about a second.

The committee on the route: all five members who gave a view said a mix with code first. Code does every word, number, chart and the pack. Kling does only the one object per post that has to feel alive, as clips that move one way and never play backwards. Nobody would use Meta's own "Add animation", which labels the ad as AI automatically. Melwani and Georgi said "No opinion on this."

Savannah Sanchez would use Kling only for S1's cell, and only if a full-size check finds nothing that looks AI-made. Her reason: Higgsfield's Kling files carry an "AIGC" tag naming Kling as their maker, and Higgsfield's terms (section 5.5) say not to remove markings Higgsfield applies, which may cover it. Meta isn't known to read that tag (it reads C2PA and IPTC records, which Kling's files don't carry), but it is an AI marker in the file. The code route carries no tag at all.

Music: partly.

• What Claude can do: make a simple background track with code. Test 1 has one; nobody has heard it yet, and Sanchez would replace it. Claude can't compose a produced track or judge one by ear.
• The best route: Meta's Sound Collection.

- It's free. Meta says it is "Free to use in any videos you create and share on Facebook and Instagram", and it is the library Ads Manager uses when you add music to an ad. - For videos, Claude picks tracks by mood and tempo, bakes them in and cuts the animation to the beat. - For carousels, a track is chosen by hand at ad setup. - Meta's Reels ads page says Reels can't use licensed music and points to royalty-free music such as the Sound Collection. Epidemic and Artlist sell royalty-free licences for ads, but Meta doesn't name them.

• The paid upgrade (prices read 16 October), both covering paid ads for a company:

- Epidemic Sound Business: $75 month to month, or $360 a year. It covers a brand's own ads, and anything published while subscribed stays covered after you cancel. - Artlist Music & SFX Pro: $299 a year, yearly only, covering companies of up to 50 people.

• Not recommended: AI music. Suno is allowed commercially, but its files could trigger Meta's AI label; Udio's terms bar commercial use. Higgsfield has no music tool.

The best way to make it happen, if Mike and Nick say yes:

• Route: code from the build files in _working-files.zip. If Mike and Nick pick the Kling option (question 1 B), a Kling clip goes in only where the motion is the point (S1's cell, and a few others decided post by post). Test 2 leaves Kling's AIGC tag out of its render, which Higgsfield's terms (section 5.5) may bar; keeping it is the other way. Code alone avoids that question, but then the cell itself never changes shape. Then the clean-file check (C2PA out always; the AIGC tag out too unless Mike chooses to keep it), the match check and a look at every contact sheet.
• How good: test 1 is the floor, and test 2 shows the committee's changes, in both routes.
• How long: about 30 to 45 minutes once, to turn the S1 code into a toolkit. Then about 25 minutes of helper time per post; with seven helpers at once that's about 3 hours for all 21, both shapes, with and without music, including one review round.
• What it costs: code only, nothing beyond Claude time. With Kling, one clip per post where it helps is 21 to 30 clips, 180 to 260 credits, about $9 to $13; the Higgsfield test priced two or three per post (50 to 70 clips, $22 to $30), and Cooper and Robinson expect that many; Denney expects fewer, and Levinger would keep to one per post. After test 2, 218.75 credits are left (checked), enough for 25 clips; beyond that it needs more Higgsfield credits (a purchase only Mike can make) or code alone for the rest. Music is free with the Sound Collection.
• Who: META MISC (this session) builds them, since the toolkit is here. They go on Mike's picks page like every other ad, and CONTROL CENTER uploads only what's picked there.
• What we'd need from anyone: nothing for the animation in code; with Kling clips, more Higgsfield credits past 25 clips (only Mike can buy them). For music, a session downloads the Sound Collection tracks in the background Chrome, which is already signed in, unless Mike buys a library (only he can make a purchase).

5. Their recommended first test

All seven: the new ads go into the four live ad sets. No new ad set, no Meta A/B test, no budget moved by hand, no AI label. Meta splits each ad set's budget among its ads, and the automatic 48-hour rule keeps moving money between ad sets on new-customer return (next at 12:01am Pacific, Saturday 17 October).

The test where their answers overlap most:

• Carousel or video: most say carousels first, because they need only small changes, with videos added as they're built. Denney would run videos only; Cooper and Melwani would run both from the start.
• How much budget: all of each ad set's budget, shared and not carved out. Four (Cooper, Levinger, Sanchez, Georgi) said "No opinion on this" on a dollar figure.
• How to read it: new-customer sales per $1 only. Read at the 48-hour rule (Cooper) or after 7 days (Melwani, Sanchez). Robinson would switch off any single ad that has spent about three times the target cost of a new customer without one (about $300 if the target is about $100), and says Mike's number wins.
• Second wave (Denney, Levinger): the batch posts on TikTok from Sunday 18 to Saturday 24 October, and whatever wins there picks the next Meta ads.
• Melwani's check first: look at landing page views per outbound click on the BROQ pages, because the BROQ weakness may be the page or the tracking as much as the ads.

6. Still image ads: replace or join?

Join first (six of seven). The batch 4 ads go in beside the live stills, and a still comes off only when its batch version beats it on new-customer sales per $1. Melwani alone would replace now. His reason: Meta rewards different ideas, and a still on the same page and angle is a third version of the same thing.

Live stills that would come off, in this order (by how many members named each; Georgi and Cooper would start with "pure source" and the lab comparison, the two biggest spenders of the six), each only once its batch twin beats it:

Keep (all seven): BROQ "winter ready" ($5.45 per $1 on a single order, and the only winter immunity ad), NMN-99 "proven NMN" ($0.96; N2 joins it on the same page), SPMD-99 "everyday shield" ($1.24), BROQ "longevity secret", SPMD-99 "fuller look" and SPMD-99 "inner glow" (too little spend to judge, and they speak to people the batch doesn't). Five of seven also keep NMN-99 "test or refund".

Planned stills on the picks page:

• Where a planned still repeats a batch idea, the batch version goes first: c2-spmd-146 (five members; S2 says the same), n13-broq-runner-up (four), and n13-broq-same-package, c2-broq-lab-bench, n13-broq-one-of-11 and c2-broq-rank-1 (two each).
• Planned stills with a picture of their own still add variety and stay in Mike's picks: c2-broq-first-cough, c2-broq-snow, c2-top5-258, c2-spmd-village.
• No planned still is rejected; it's Mike and Nick's pick as always.

Spend numbers are from the Thursday 15 October 7:28am Pacific read (STILL-ADS-NOW.md), new customers only.

Files

All in MARKETING AREAS/META ADS/MISC/batch4-for-meta/:

• committee/<name>.md: the seven members' full answers.
• reviews/<id>.md: each post judged slide by slide; reviews/FULL-SIZE-CHECK.md: the 11 posts first judged from the overview, rechecked at full size.
• review-of-brief/: the two checks of this brief.
• DIGEST.md: all 21 in one table, with every change, claim, animation plan and music note.
• STILL-ADS-NOW.md: the 13 live and 69 planned still ads.
• test/:

- ANIMATION-TEST.md, S1-animated-4x5.mp4, S1-animated-9x16.mp4, the -music copies, S1-contact-sheet.jpg, S1-match-check.jpg, and the code in code/. - HIGGSFIELD-TEST.md and the Kling clips in higgsfield/. - MUSIC-AND-TOOLS-RESEARCH.md: Meta's music and AI-label rules, library prices, other tools. - music/: the code-made track.

• PROGRESS.md: how the job ran.

Back to the brief