An AI Content Studio for Ads, Webinars, and Shorts
How we edit our own talking-head ads, thumbnails, and webinar with a team of AI agents, in one edit style that we chose from 12 and wrote down so each wave can change it.
63
Meta ads edited in the red pen style
12
Edit styles built and reviewed
41
Shorts planned from one webinar
CLIENT OVERVIEW
This is our own content operation, run by a small team with AI. We record, and a team of AI agents does the editing. Claude Fable, the coordinating model, plans each job, writes the briefs, and reviews what comes back. Claude Opus agents do the editing, several at a time. The work covers talking-head ads for Meta, thumbnails, and a recorded webinar cut into a full-length edit and short clips for Reels and YouTube Shorts. We built it for our own marketing, and an agency content team can run the same pipeline on its own footage, with its own people making the calls.
THE PROBLEM
What was broken
Meta needs a steady supply of fresh creative
Meta groups ads that look alike and shares their learning, so 13 videos that differ only in the hook can tire together. A weekly rotation retires the losers and adds up to 3 ads from a bench of finished, paused ads. That bench has to exist before review day, in two sizes, with every ad checked.
Every variant is its own edit
A talking-head ad needs its pauses cut, captions timed to the words, graphics placed around the face, loudness set, and two sizes rendered. By hand, each hook means another full edit, and a wave of 50 clips means 50 of them.
A webinar recording arrives as separate files
A webinar comes out of Riverside as a 4K face cam, a mic track, and a screen share of the slides. In our recording, the face cam ran 0.16 seconds behind the mic, and the slides started 15.06 seconds into the session. Before anyone can cut a short, all three have to line up.
THE GOALS
What success looks like
Choose one edit style by building several on real footage and reviewing them side by side
Edit every variant from a written spec, so another hook costs a spec and a sketch, not another design
Turn one webinar recording into a full-length edit and short clips for Reels and YouTube Shorts
Write every step down so the next wave, or the next agent, can change it and run it again
Keep a person on the choices that matter: the style, the thumbnail, and anything that goes live
OUR APPROACH & SOLUTION
Trying 12 edit styles, then writing down the one we chose
We didn't pick a style up front. AI agents built 12 overlay directions on two test clips, each one a spec keyed to the transcript plus a build script, so a direction applies to any clip without a redesign. We reviewed them side by side, and on November 4, 2026, we chose direction L, red pen paper scraps: a red pen, a highlighter, stamps, handwritten notes, and light-ink sketches on paper, borrowed from our WordPress QA webinar page. It became the standard for every wave since. Claude Fable splits each wave into briefs, and Claude Opus agents work side by side, each one editing only its own clips' files while the shared pipeline stays frozen. The whole process lives in a runbook that an agent with no project memory can follow, so the next wave starts from what the last one learned and can change any part of it.
Phase 1
Twelve Edit Styles
By November 4, 2026, AI agents had built 12 overlay directions on two test clips: stat receipts, magazine pull quotes, split screen, our site's design system, a dark blueprint, kinetic type, a blend of three of those, a cutout of the speaker on a web page, a document reaction, and three red pen versions. We chose the last one, red pen paper scraps.
Phase 2
The First Wave
Thirteen talking-head clips that share one body and differ only in the hook went through the red pen pipeline. Each got a 9:16 master and a 4:5 feed cut, 64 to 68 seconds long at -14 LUFS. All 13 ads went live on Meta on November 5, 2026, the day after we chose the style.
Phase 3
The Runbook
On November 5, 2026, the edit went into the repository as a runbook, a batch guide, a launch skill for Meta, and a creative playbook that labels every claim as measured, reported, an estimate, or a hypothesis. The runbook stops at the finished files, and nothing uploads or goes live without a person's yes.
Phase 4
Composable Parts
The second wave changed the pipeline. On November 6, 2026, each clip became a hook, a body, and a call to action, each recorded once and reused, so 7 hooks, 2 bodies, and 2 calls to action made 26 clips. On November 7, 14 more hooks added 24 clips. Before the change shipped, all 13 first-wave edit lists rebuilt byte for byte.
Phase 5
The Webinar
We recorded the WordPress QA webinar in Riverside and handed the agents the face cam, the mic, the slide screen share, and a session of intro and outro takes. They lined up the files, cut a 42-minute YouTube edit, and planned 41 shorts with a full and a bare variant each. On November 7, 2026, 61 of those variants rendered in the red pen style.
THE JOBS THEY HOLD
Whole jobs, not single tasks
Ad editor
For: Meta video ads
- Transcribes each clip with word timings, maps its pauses, and tracks the face at 10 frames a second
- Writes a spec and a hook sketch per clip, with every pen mark cued to a spoken word
- Cuts the pauses, draws the captions, and places paper graphics that add what the caption doesn't say
- Repairs late word timings from the audio before any caption or mark uses them
- Renders a 9:16 master and a 4:5 feed cut, both at -14 LUFS
- Builds wave clips from a hook, a body, and a call to action, each recorded once
Guardrail
A graphic never covers the face. The build checks every frame of both sizes and fails the render if one does.
Thumbnail designer
For: YouTube and the webinar replay
- Pulls 16 candidate frames from the 4K face cam at emphatic words, such as "Boom" and "Shocker"
- Crops each candidate to 1280 by 720 and tiles them on one contact sheet
- Sets a slide from the webinar on desk-gray paper as a slide thumbnail
- Writes title options under 70 characters, with the reason for each
- Builds thumbnail graphics in HTML and CSS, imagery included, and renders several versions of each idea
- Lays out 1280 by 720 thumbnail designs in the 10ex type and colors
Guardrail
A person picks the thumbnail. The kit writes out only the candidates that person names.
Webinar producer
For: Webinar attendees and the YouTube channel
- Builds the slide deck as a web page, with speaker notes that follow the slides in a second window
- Lines up the Riverside face cam, mic, and screen share, and corrects a 0.16 second lip sync lag
- Indexes the 34 slides in the recording with their times and titles
- Repairs the transcript where speech repeats or the model wrote words over silence
- Cuts pauses, stutters, and fillers: 392 cuts that removed 207.7 seconds
- Writes 18 chapters and 1,027 caption cues, and lays out the slide and the presenter side by side
Guardrail
The mic audio decides every cut. No removed pause holds a voiced frame, and every kept piece runs at least 150 milliseconds.
Short-form editor
For: Reels, TikTok, and YouTube Shorts
- Plans each short from exact transcript words: the hook, the spans, the intro, and the outro
- Renders a full variant with a recorded intro and outro, and a bare variant from the webinar alone
- Switches each part between a face, a stack, or a slide layout
- Adds red pen scraps, labels, and stamps, each cued to the words
- Uploads every finished render to a review folder and texts the link
Guardrail
The build never shrinks a graphic to make it fit. It nudges, moves, or drops it, and the report lists every change.
Ad launcher
For: The Meta ad account
- Writes the ad copy and a manifest that lists every ad in the wave
- Creates the campaign, ad set, and ads paused, each ad carrying both video sizes
- Delivers the renders to a shared drive, sorted by wave and size
- Runs a read-only review every Friday against written keep and kill rules
- Retires losing ads, adds at most 3 bench ads a week, and logs each change
Guardrail
Nothing goes live or spends money until a person says yes.
WHAT WE BUILT
Features and capabilities
01
Styles Tested on Real Footage
Twelve overlay directions were built on two test clips before any ad shipped. Each direction is a spec keyed to the transcript plus a build script, so a direction applies to any clip without a redesign.
02
Specs Cued to Spoken Words
Every highlight, pen mark, note, and stamp in a spec is cued to words in the transcript, with a lead in seconds. A cue that matches no word, or more than one, stops the build and lists every match.
03
Composable Waves
A wave is built from parts recorded once: hooks, bodies, and calls to action. Each part is written once as JSON with cues, and the pipeline resolves it against each clip's own words. In the second wave, 21 hooks, 2 bodies, and 2 calls to action made 50 clips.
04
Face Tracking in Every Frame
Apple Vision tracks the face at 10 frames a second. For every output frame of both sizes, the layout tests each graphic against the face. The picker nudges a graphic or moves it to another margin first, and the report records any graphic it shrinks or drops.
05
Checks Before Anyone Watches
Each render writes a report and contact sheets: loudness at -14 LUFS with true peak at or below -1.5 dBFS, the face check's closest approach, every retimed word, and every placed graphic. A splice in the source audio or a cut under 150 milliseconds fails the run.
06
Regression Checks Across Waves
A pipeline change must leave earlier waves alone. Before the second wave shipped, all 13 first-wave edit lists and 36 check stills rebuilt and matched the committed files.
07
Parallel Agents With Clear Lanes
Claude Fable plans each wave, writes the briefs, and reviews the results. Claude Opus agents do the editing at the same time, each touching only its own clips' spec and sketch. The shared pipeline files stay frozen during a batch, and three clips render at once on one machine.
08
Webinar Ingest and Sync
The webinar ingest repairs the transcript, maps pauses from the mic, tracks the face across 46 minutes, and indexes 34 slides. It syncs the screen share to the mic from eight slide changes, with no drift across the session.
09
Runbooks, Skills, and Playbooks
Each process is written for an agent with no memory of the last session: the red pen runbook, the batch guide, the ad wave launch skill, the creative playbook, and the rotation playbook. When a runbook and a file disagree, the file wins, and the next wave updates the runbook.
KEY TAKEAWAYS & IMPACT
What changed for this business
63
Meta ads in one house style
Thirteen in the first wave and 50 in the second, each with a 9:16 master and a 4:5 feed cut. Every render passes the same face check and loudness target, or it doesn't finish.
1 day
From choosing the style to live ads
We chose the red pen style on November 4, 2026, and the 13 first-wave ads went live on Meta on November 5, edited, checked, and delivered in both sizes.
41
Shorts planned from one webinar
One 46-minute recording became a 42-minute YouTube edit with 18 chapters and 1,027 caption cues, plus 41 short plans with a full and a bare variant each, in the same style as the ads.
THE OUTCOME
The wins
-
The style question was settled on real footage: 12 directions built and reviewed, one chosen on November 4, 2026, and every ad since in that style.
-
The 13 first-wave ads went live on Meta on November 5, 2026, the day after the style was chosen.
-
Another hook costs one part file and one sketch, so the second wave turned 21 hooks, 2 bodies, and 2 calls to action into 50 clips.
-
Each wave starts from the last wave's runbook, and a pipeline change has to rebuild the earlier wave's edit lists byte for byte before it ships.
-
One 46-minute webinar became a 42-minute YouTube edit, 18 chapters, captions, and 41 short plans, with 61 variants rendered on November 7, 2026.
-
A person makes the calls that matter: the edit style, the thumbnail, and every ad that goes live.