Making an AI Short Drama Solo, Part 5: Editing, Quality Checks and Publishing

Behind the scenesOct 18, 2026 · 6 min read

By this point every shot exists: voiced, generated, moving. It is tempting to think the episode is basically done. It isn't. In my time logs, the stretch from "all shots exist" to "episode is live" is where the most surprises show up. This post covers that last mile.

Editing: cut on the dialogue

Because the voice is recorded first (Part 3), the edit is built on the audio timeline, not on the clips. Each line sits at its real position, and the picture is cut to fit around it. A few habits made the biggest difference:

• J and L cuts. In real conversation you often hear the next person before you see them, or keep hearing someone after the camera has moved to the listener. Letting a line's audio start over the previous picture, or run over the next one, makes a two-person scene feel like people talking instead of a slideshow of speakers.
• A reaction after the punchline. The joke lands on the line, but the laugh lands on the other person's face. A short, silent reaction shot after the punchline almost always beats one more line of dialogue.
• Nothing frozen. My rule is that no shot in the body of an episode may be a still frame. Only the end card is allowed to hold. A freeze detector scans every final cut, because a frozen half second is easy to miss while editing and very easy to notice on a phone.

Text is always drawn in post

Generators are terrible at writing. Ask for a chat window or an email on screen and you get confident nonsense that looks almost like Chinese. So the generator never draws text. The monitor or phone is generated blank, and the chat bubbles, emails and code are drawn afterwards and tracked onto the screen. Screen insert shots still have to move: bubbles pop in, the AI's reply streams, the screen glow flickers. A static screenshot counts as a frozen shot.

Subtitles follow their own small rules: bold with an outline, no speaker name labels, sitting close to the bottom edge of the widescreen frame, and wrapped so a punctuation mark never starts a new line. That last one sounds fussy until you see a lone comma at the start of a subtitle.

Sound does half the comedy

Each location gets a quiet ambience bed: office hum, kitchen sounds, street noise under a car scene. On top of that go small stings that point at the joke. One surprising lesson: real recorded keyboard sounds beat synthesized ones every time. Programmers know what a laptop keyboard sounds like, and a fake one breaks the spell. Music ducks under dialogue, and the final mix is brought to the loudness the platforms recommend, so my episode doesn't jump out quieter or louder than the one before it in the feed. Every episode ends with the same short animated end card and a "follow me" line, so the series feels like a series.

Quality checks, in layers

No single check catches everything, so I stack them, cheapest first.

1. The contact sheet. Every time any image changes, even one, all the shots go onto a single sheet. That sheet is for consistency across shots: same face, same outfit, same room. It is not good at small things, so I also zoom into individual frames for hands, props, and the classic AI ghost: an extra blurry person in the background of a one-person shot.
2. A vision model pre-screens first frames. It counts people, looks for broken faces and hands, watermarks, and blank patches where the image fell apart. I only look closely at what it passes. It cannot judge whether a character is the same person from shot to shot, or whether an expression is right. Those stay human.
3. Automated shot QC. After the videos are generated, a script measures camera drift on shots that are supposed to be locked, and flags objects that appear from nowhere. It once caught a "ghost coworker" fading into the background that I had missed. Another time it proved me wrong: I was sure the camera had pushed in, but the measurement showed the camera hadn't moved at all. The actor had walked closer. Still, it is triage, not a verdict. A green report means two specific problems are absent. It can't see acting, lip sync, or fake text.
4. I watch the whole thing. Start to finish, as a viewer, before it goes anywhere. A log line that says "done" is not a review.

Rules that stop endless loops

Quality checks have a failure mode of their own: fix one thing, the next pass finds another, forever. Two rules keep that in check. If rewording the same shot two or three times doesn't remove an artifact, I stop rerolling and change the composition or hide it with a cut. And if an episode fails review three rounds in a row, the automation stops and a human, me, decides what is worth fixing.

Publishing: a checklist, not a vibe

The widescreen master is the source of truth. From it I export a version for each platform. The upload file is sized to fit the platform's limits, with the bitrate worked out from the episode length rather than a fixed number. I learned that the hard way: a bitrate that was too low once got the video flagged as "low resolution" by a platform's quality check.

Covers get more thought than you'd expect, because the cover decides whether anyone taps. I make them in several aspect ratios and follow a few rules:

• Rotate the lead character and the background color between episodes, so a new episode never looks like a repost of the last one.
• Turn the expression up. A mild frown reads as nothing at thumbnail size.
• Use hook words, not a plot summary.
• Shrink it to a quarter size, put it next to the last five covers, and check that it reads as new within a second.

The AI-generated disclosure is always ticked, and every description ends with a line saying the story is fictional. After posting, I check the platform's own content manager to confirm the post really exists and is under review or live. A button that clicked without an error is not proof. Then the whole episode folder is archived in one piece, so the finished video and everything that made it stay together.

Reading the numbers

About two days after posting, and again after a week, I record each platform's numbers: views, likes, comments, saves, shares, completion. They go into a simple table next to features of each episode: length, number of shots, shot sizes, topic. Once a week that turns into a short brief that helps me pick the next topics. With a handful of episodes, the patterns only show direction, not proof, and I try hard not to read more into them than that.

Next up

The last part is about the assistant behind all of this: Claude Code as my production assistant. How one session plays producer while specialist agents write, direct, check and edit, and why the written list of mistakes ended up being the most valuable thing in the pipeline.