Measured first.
Then changed.

41 of 53 graphics were faulty. Not one was the model's fault.

Problem

Episode 3's first automated run put 53 graphics on the review board, and most of them were wrong.

Measured

Every fault traced to where it was made: 14 in selection code handing components data in the wrong shape, 11 in the decision matrix, 11 in rules that contradicted the series' own decisions, 5 in inputs. 0 in the model: re-run on two other models, the same faults came back.

Changed

The fixes went into the joins between stages, not into prompts: a brand lexicon with an on-screen word check (the transcriber had heard "Firstmate" as "First Meet", and every writing model copied it), rows blurred until named instead of dropped, a centred chapter card with its rail, a text-over-face probe, and calls that draw nothing are refused before the board.

Still open thenEvery check that would have caught this ran after the board. Moving the checks in front of the board was the next piece of work (02).

The pipeline knew. The review board never showed it.

Problem

Episode 4 passed every check and still felt repetitive: ten ledger cards in one episode.

Measured

One script over every run's plan. Episode 4 was the first run with zero check refusals and the worst on every variety number. Its own rubric had written 5 "swap" and 2 "cut" verdicts. They were demoted to pending and never reached the approve button.

Changed

A golden set: three recorded episodes with the approved result at twelve stages. Each stage is scored twice, once on perfect input and once on the stage before it, so a fault points at the stage that made it. Checks are the gates; a model only ranks. The critic was calibrated on his blind A-or-B picks and moved in front of the board, and its cost fell from $13 to $1.85 an episode. The goal is 10 or fewer notes from him per episode, down from 53.

Still openThe whole-episode variety gate this review proposed (refuse episode 4 as it stood, pass episodes 1 and 2) has not shipped yet.

We ran the competition on our own footage, and read its code without copying it.

Problem

Is our pipeline actually better than the leading Claude Code video plugin, or does it only feel that way to us?

Measured

We installed it the way a buyer would, in a clean account, gave it the same raw 14:39 episode and let it run alone. Then we ran the same measuring commands on both cuts. Separately, we read its 46 effects (47,251 lines of Python) and measured them without running them. We copied nothing.

What the code read taught us

Its effects share a good habit: the model fills a few constants and the code does the measuring, placing and refusing. But the code is copied rather than shared: at least six ways of finding when a word starts, and 26 copies of one of them. Ours is the stronger system (tokens, a catalogue the code reads, a check that proves before render), but slower to extend: a new component touches 5 to 9 shared files. Its demo Reels run 8 to 23 graphics a minute, while its maker's own hand edits ran 2.7. That spread is why density became a setting the creator chooses (10).

Still openWe leave more pauses and cost more per episode. Making a new graphic a one-folder change is planned, not built.

An approved Short script becomes its cut, and every jump gets a cover.

Problem

A Short stitched from lines across the take showed a hard jump at each join, so its proof sheet refused it on all 5 platforms.

Measured

The cut is now one exact pass from the raw take, so its timing map runs straight from raw to final. All 95 of the script's words survive, in order. The proof sheet's seams row counts covered jumps: 0 of 2 before.

2 of 2 jumps covered (was 0) · 4 of 5 platforms delivered (was 0) · −16.08 LUFS, −2.01 dBTP

Changed

The planner puts a cover on every jump in a script cut, a takeover or a punch-in, exactly as the script names it. A script whose jumps can't be covered is refused before it is planned.

Still openLinkedIn needed its own storyboard (06). The caption after a takeover stays hidden for about a second; that is filed as its own fix.

No text across his eyes, checked before anything renders.

Problem

Hook headlines ran across his forehead and glasses, and a label sat over one eye.

Measured

His eye band is the middle of the head box from the video's matte: 38–62% of its height, 15–85% of its width. It was set from 18 frames of three cuts and mapped through the camera at every punch-in.

Changed

The frame check refuses text or a graphic across his eyes, and a graphic over another component's text, before anything renders. A hook too tall to fit shrinks, but never below 60% of its size. A graphic with no clear place is dropped, and the plan says why.

Still open thenTwo of three Shorts each lost a graphic. The eye band also left the rest of his face open (07).

One storyboard, checked on every platform, LinkedIn's title bar included.

Problem

The plan was checked only on the main build, which has no title bar. On LinkedIn, two logo pops landed on the bar, and the run stopped at step 8 of 15.

Measured

The planner's log reproduced the three faults word for word: graphic over text at 23.5 s: pop 'logo alexandria' … over 'I got seven AI agents'. After the fix: 6 of 6 platform checks clean, and the step the run couldn't reach before exits 0.

Changed

Once the main build passes, the planner builds LinkedIn's own version from the same storyboard and checks it too. A graphic that collides there is moved or dropped on the one storyboard, so every platform stays a pure swap. The unit tests use the three real faults verbatim, and all 65 planner tests pass.

Still open thenThe check only knew his eye line, so the nearest clear place was his nose, on all five platforms. We shipped it with that known gap and fixed it the same day (07).

Graphics keep off his whole face, not only his eyes.

Problem

A pill reading "Way more helpful" sat on his cheek for three seconds. The proof sheets said "his face is never covered: pass" on all 15 files.

Measured

That check covered only graphics without text, and a pill is text. Re-checked with the fix: text across his face at 22.0 s: 'Way more helpful'. The face box runs from the eye band to the chin, measured on the matte.

eye band no text or graphic crosses it, as before · face box new: no graphic crosses it. The hook card stays under his eyes, as the motion rules intend.

Changed

The frame check refuses any graphic, pill or label across the face box. The proof sheet's face row reports the check's own result ("pass: 0 of 111 boxes across his eyes or face") instead of "not checked".

Still openOne Short lost its only graphic over the footage: there was no clear place for the pill on LinkedIn (56 px of room, the pill needs 88). Putting graphics back is the density work (10).

A hook card ends on a clause, never halfway through a thought.

Problem

Hooks read "So with / Alexandria" (no claim), "Building is fun, but / marketing" (a subject still waiting for its verb), and a takeover ended on "about".

Measured

Every card and takeover on episode 3's cuts, before and after, plus the frame check on all four builds: 0 fails.

Changed

One rule in one place. A card ends on the sentence's end, a clause mark, or a word followed by a conjunction, pronoun or preposition. It never ends on a function word or a subject before its verb, and never opens on "So" or "Because". The script check refuses a hook that opens on "Because". It uses closed word lists, with no new language model in the loop.

Still openWhen no clause fits in three lines, the card falls back to the landing phrase on its own, as before.

The length estimate is the cut's own length.

Problem

Scripts were ranked on an estimated length up to 9.7 s longer than the cut they produced.

Measured

The transcriber pads words with silence, so word times can't predict a cut. The estimate now makes the same pause cuts the cutter makes, which takes 2 s per take.

What honesty costMeasured honestly, every earlier Short was under the approved 25 s minimum, and 0 of episode 3's scripts passed. The bar was lowered to 20 s, measured on the cut, and recorded in the PR so it can be reversed.

Graphics a minute, counted, then set by the creator.

Problem

The planner dropped any graphic that overlapped a seam cover, and nothing counted how busy a Short was.

Measured

Every proof sheet now has a graphics-a-minute row, counted from the plan. On episode 3's three Shorts: 6.8, 5.0 and 3.0 → 6.0 a minute, all short of the 8–10 target.

Changed

A graphic in a cover's way moves to just after it, keeping its planned length. Then the brand picks the pace: Calm (3–5 a minute), Standard (8–10, the default), Busy (16–20), or an exact number. The plan never pads to reach the band; when it falls short it says why, word by word.

Still openOff target is advisory: if it refused delivery, all three of episode 3's Shorts would be refused today. Explain-and-compare sentences still have no graphic that fits them.

A camera per span of the cut, set from where his head is.

Problem

One fixed vertical crop for the whole Short cut his head off whenever he leaned. It was set by hand.

Measured

The share of his head's width on the 9:16 frame, for every span of the cut and for every second. The check refuses any span under 80%.

Before (left of each pair): one camera at the centre. Now (right): the camera that span gets, cut on the jump. The source picture is never moved.

A · 0.0–1.6 s94% → 100%

A · 4.8–8.6 s53% → 100%

A · 11.1–13.4 s47% → 100%

B · 7.6–9.0 s31% → 98%

B · 11.0–12.7 s35% → 99%

C · 25.8–36.9 s42% → 100%

Changed

The planner sets a camera for each span from his head on that span, with no hand edit, and the check measures every span before the Short is composed.

Still open13 of 81 seconds are still under 80%. In 12 of them a raised hand or the notebook widens the measured head box past what a 9:16 frame can hold.

Review stills from a settled frame, not mid-animation.

Problem

The review screen showed counters caught at 16 of 17 and blurred cards. Each still was taken a fixed second after the graphic arrived.

Measured

Each still is now taken at the middle of the graphic's hold, between its last arrival settling and its exit, read from the composition's own events. That's 45 graphics on episode 3.

Still openOne graphic still fails to snapshot at any time, as it did before. A few are mid-move by design (a reveal's blur, a canvas pan) and stay that way.

Each brand writes in its own voice, not the founder's.

Problem

An audit for hard-coding found every brand's captions and newsletter were written from one file describing the founder's own show, with "his voice" and "a founder's weekly video" in the prompts.

Measured

We recorded the exact prompt sent for a second test brand, for a brand with no voice, and for the founder's brand, then counted the founder's lines that leaked into the others.

Changed

The voice comes from the active brand's own kit (capped at 12,000 characters) and nowhere else. The prompts say "the creator" and "their own words". A run with no brand set stops rather than falling back to anyone's voice.

Still openThese checks used a stand-in model that records prompts; the wording a real model writes in each voice is judged on the review board, not by a test.

Posting times belong to the brand, in its own time zone.

Problem

Rules said "one a day from next Monday", but nothing turned that into a time. A real delivery arrived with no scheduled time, and the only 09:00 was in test data.

Measured

We onboarded a New York brand and delivered on a Sunday afternoon, New York time, with the test clock set to 25 Oct 2026. Then we tried to break it.

Changed

Each brand sets a time zone and its posting times (09:00 in the creator's own zone by default, set at onboarding). Every post carries both, and the Studio shows each brand's times on its Brand page.

Still openA kit onboarded on Windows, where no zone can be read, is refused until the zone is set. Two deliveries can still be given the same slot.

The kit on Linux and Windows, tested without a CI bill.

Problem

Content OS has to run on macOS, Windows and Linux, but had only ever run on a Mac. The three-system CI job couldn't run because the account's CI billing was blocked.

Measured

The same CI job, run on a Mac for free: natively for macOS, and in a fresh Ubuntu 24.04 x86-64 container for Linux, each on a clean clone of the PR.

Changed

Linux breakages fixed where they showed (a keep-awake lock systemd 255 refuses, a 25 fps fallback on FFmpeg before 8). Windows fixes were inferred from them: retrying file writes while Windows holds a file open, UTF-8 output, and finding a working Python rather than the Store stub.

Still openWindows has no local stand-in. Its fixes are proven by unit tests that simulate it, and the first Windows tester's run is the real check.

A tester's iPhone clip: from a crash to delivered.

Problem

A dry run with a 45 s vertical iPhone clip (HDR inside an SDR label, rotated) and a brand-new tester kit crashed the planner and never delivered.

Measured

The same clip, re-run on the fix, gap by gap.

Still openIts proof sheet (shown in 10) says 1.7 graphics a minute, under the 8–10 target, and two checks didn't run on this kind of video. It ships anyway, and says so.

The Studio, clicked through: 11 findings, 9 fixed, 2 still open.

Problem

A full click-through of the review app found eleven things a creator would trip on, from "asleep since undefined" to a step bar that pointed backwards.

Measured

Every screen screenshotted before and after, at phone (390 px) and laptop (1280 px) widths. That's 34 page loads with no console errors, no sideways scroll and no "undefined" text.

Still openTwo episode stills are still blank or blurred, and two transcription errors remain in Short 1's seeded captions. A new nit: the step bar's fill stays at a third while its label says 4 of 4.