Your library does not know which hook paid for itself. Your brief writer does not know what the library holds. So every round starts from somebody's memory.
heybar closes the circle.
Knowledge writes the brief. A machine editor cuts it. You stamp the frame. It launches — and the spend lands back on the exact hook that earned it.
Not one of them remembers the last ad. You already know every item on this list, and you know which one you were looking in at 11pm on Thursday.
a Drive foldera Frame.io linkthe spreadsheeta Slack threadtimestamps in DMsa Notion docsomeone's Desktopan Airtable basea naming convention nobody followsthe editor's hard drivea WeTransfer that expiredscreenshots of Ads Managera Loom nobody watched“final_v4_REAL.mp4”an editor waiting on you
The mechanism
One loop. Six stops. Nothing gets retyped.
Every tool in your stack is a stop on this circle, and not one of them tells the next one what it learned. heybar is the circle — so what happened at stop six is already sitting at stop one when the next round starts.
01Knowledge
01
Knowledge
What your customers actually said, with the receipt attached.
Reads
Reviews, Reddit, support tickets, post-purchase surveys, competitor ads
Makes
Named avatars — a voice-of-customer cluster with a person's name on it
Guard
A quote that is not literally in the source it cites is refused.
02
Brief
A creative director that has read the doctrine, not just the dashboard.
Reads
Live winners, everything already tried, the brand's own named rules
Makes
The brief — avatar, angle, hook, destination, format, headline checks
Guard
Awareness and sophistication are decided before a word is written.
03
Make
A machine editor that holds a seat on the board, not a tab in your browser.
Reads
The brief and its reference clips
Makes
Shot list, generation, stitch, captions, mix — delivered as a card
Guard
It delivers to Needs Review. Never one column further.
04
Review
Stamp the frame. Pin the line that launches.
Reads
Nothing you have to explain twice
Makes
A note at the second, and the exact copy going live
Guard
Nothing reaches Ready to Launch without a person, except a static that passed every fidelity check.
05
Launch
One writer to your ad account. One.
Reads
Caps, ladders, kill thresholds — every one of them a setting you can see
Makes
Paused ads, budget steps, pauses, each logged with its arithmetic
Guard
The model proposes five kinds of action. A deterministic engine owns the rest.
06
Learn
The money lands on the hook. Not just on the ad.
Reads
Spend, CPA, ROAS, retention, and the orders Shopify actually verified
Makes
A number on every hook, body, CTA and archetype you have ever run
Guard
A winner locks per avatar after five purchases, then the other slots keep testing.
The roster
Four of your teammates are machines. They have job descriptions.
Not “AI-powered.” Each one is a row on the People tab with a key of its own, pinned to one brand by the server. Each ships a specific artefact. Each has a list of things it is not allowed to do — and that list is the part worth reading.
On the roster
Creative Director writes the brief
Reads
Winners, rules, avatars, competitor ads, what you already tried
Ships
A brief with the avatar, the angle, the destination and the headline checks already run
Cannot
Override a rule you switched on. The deterministic picks are made before it is asked anything.
On the roster
AI Editor cuts the ad
Reads
The brief and its reference clips
Ships
Hook, b-roll, voice, stitched, captioned and mixed — as a delivery on the card
Cannot
Move a card past Needs Review. A person still approves every cut.
On the roster
Image Studio makes the static
Reads
The brand's reference set, base prompt and fidelity checklist
Ships
Renders graded check by check; a clean pass walks itself to Ready to Launch
Cannot
Pass a render that fails a check. The checklist is the brand's, not the model's.
On the roster
AI CMO reviews the night
Reads
What the buyer did, what the kill switch did, profit, competitors
Ships
Answers to the buyer's open questions, and the ones it will not answer as yes/no tiles for you
Cannot
Write to Meta. Ever. It reads and advises and that is the whole job.
The library
Search your footage by what somebody said in it.
Speech is transcribed on our own machines with Whisper. Frames are embedded with CLIP, so you can search what a shot looks like. Faces are grouped into the people who keep turning up. Nothing leaves for a model that does not need it.
6 of 4,812 clips · 0.04s
Application“I do this before anything else touches my face.”$14,208 · $21.40 4.8xTestimonial“I was skeptical, and I'll tell you exactly why.”$31,905 · $18.90 5.4xUnboxing“Okay — it came in two days.”$6,410 · $29.10 3.1xProduct$9,882 · $24.60 3.9xB-roll$4,120 · $33.80 2.6xSwatch$21,744 · $19.70 5.1x
Attribution
Most tools score the ad. This one scores the sentence.
An ad is not one thing. It is a hook, a body, a call to action and an archetype, and each of those has already appeared in a dozen other ads. heybar rolls the spend up to the part, so the next brief can ask for the part instead of the whole.
Every hook, headline, body and CTA carries its own spend, CPA and ROAS across every ad it ever appeared in.
Winners lock per avatar at five purchases and a profitable ROAS. The other slots keep testing.
Archetypes — ugly-ad screenshot, doctor authority, before/after — are scored too, so the director can favour a shape that keeps paying.
The ad
DT-0912 · side sleepers · v3
Spend$48,190ROAS4.4x
Hook · 0:00–0:03
“I was skeptical, and I'll tell you exactly why.”
Seen in14 adsROAS5.4x
Body · 0:03–0:19
Swatch macro, then the surgeon line, then the price reveal.
Seen in9 adsROAS3.8x
CTA · 0:19–0:24
“30-day returns. Free shipping. Link's right there.”
Seen in31 adsROAS2.1x
The other half of the test
The ad is half the test. The page is the other half.
Eight locked templates. An interview, never a guess. A second, cheaper model fact-checks every claim against your own knowledge base before anything publishes — because the expensive mistake is a page confidently quoting the wrong store's return window.
Media is reuse-first: your own library before any generator, and video is never generated — only reused.
Copy variants per avatar, matched on the click and rendered by your own theme, so nothing sits in the render path.
A base-versus-personalised test that reports P(B beats A) and refuses to call a winner before ten orders and six days.
Publish to a Shopify page, a section in your live theme, or a heybar-hosted URL.
A
2.41%
B
3.84%
P(B beats A)97%
14 days · 61 orders · floor cleared
Why the numbers look boring
We draw five dots.
Meta reports retention at five checkpoints. 25, 50, 75, 95, 100.
Never per second.
A smooth curve would be a lie, so heybar plots five dots and says so on the chart itself.
That habit runs the whole way down. Every figure carries its own arithmetic on hover — what went in, what window it covers, what it excludes. Cost per acquisition is blended with what Shopify actually verified, because Meta under-reports cost and the raw number walks the budget ladder the wrong way. Returns are bucketed by order date, not refund date, since a refund lands about forty-five days late and would otherwise be charged to a month that did not earn it.
None of that makes a better screenshot. It makes a number you can act on.
Video retention · DT-0912as reported
The dashed line is interpolation, not data.
In their words
The tell is what happens when you take it away.
“Repurposing old and new content together without approval cycles or additional stakeholders makes everything much faster.”
Emily Arons · Senior Performance Marketer
“If you want us to keep the same level of quality and speed, you can’t get rid of it. Without it, our whole team slows down — and that would affect revenue.”
There is no self-serve signup. Somebody adds your email by hand, which means we would both rather know now.
It fits if
You buy paid social at volume and ship creative every week.
You pay editors, and you want them working on one board instead of five inboxes.
You already know your cost per acquisition — and you do not fully trust Meta's version of it.
You run more than one brand and are tired of maintaining a second copy of everything.
It does not if
You want to sign up with a credit card this afternoon.
You post organic and do not buy media.
You want a tool that writes ad copy and nothing else. There are cheaper ones, and they are fine.
You need a completed SOC 2 report on the desk this quarter.
Before you ask
Questions people actually ask before they sign in.
Who is this for?
Operators running paid social at volume, and the editors they pay. If you have more than one brand and a second copy of every process, that is the case it was built for.
Does this replace my media buyer?
No. It replaces the eight handoffs between them and everyone else. The buyer keeps buying; heybar is the only thing writing to the ad account, so two systems never fight over the same ad set.
What actually runs without me?
A static that passes every fidelity check walks itself to Ready to Launch. Budget ladders, caps and kills run hourly against thresholds you set. Everything else stops at a person — video never launches without an approval.
Where do editors work?
The board, and nowhere else. Cards, an inbox, and email. There is no second tracker behind it, and an editor does not need to learn one to deliver.
Can I trust the retention chart?
Meta reports five checkpoints — 25, 50, 75, 95, 100 — and never per second. We plot five dots and say so on the chart. A smooth curve would be a lie.
What about my other brand's numbers?
Every robot holds its own key, and the key is pinned to one brand by the server, not by a paragraph in a prompt. One brand's automation reading another's numbers is not a policy here. It is a 403.
How do I get in?
An admin adds your email. You sign in with a one-time code to that inbox. There is no self-serve signup, on purpose — this desk holds live spend and unreleased cuts.
Stop briefing from memory.
Tell us how many brands you run, how many editors deliver each week, and what you spend. We will tell you whether this fits — and say so if it does not.