phyo thu

Lab

Work

About

Work with me

phyo thu

Lab

Work

About

Work with me

phyo thu

Lab

Work

About

Work with me

← Lab

To The Moon · Part 2 · AI music video · 2026

Don't Let Go

A month after To The Moon, the same director and AI producer made the sequel: longer, darker and in three days. This is what a month of practice changed.

One director and an AI producer made a 75 second music video in two days. Here’s what that means for creative teams.

days, brief to graded master

$156

in image and video generation (USD)

shots in the cut, from 400+ stills and 80+ clips

versions of the edit before picture lock

Overview

Don't Let Go is the 90 second sequel to To The Moon, with an original song, made in three days for about $156 in generation costs. The moon bubble bursts, Luna is out cold in a recovery pod, her robot cat sends an SOS, and a crew on Earth launches a rescue.

Same setup as last time. I directed every creative call. Claude ran production: the brief, the world bible, prompts, generation through the Leonardo API, measuring the song, code for every graphic, review tools and the edit prep. Specialist models made the song, stills and video.

I didn’t do this to prove AI can make a video. I did it to find out how a creative lead should work when production is no longer the bottleneck. This is what I learned.

The difference is that this time I didn't start from a blank chat. After Part 1, I turned the whole workflow into a Claude skill, so Part 2 started from a playbook with last time's lessons baked in. I wanted to see what a month of practice changes when the process itself is something you can save and reuse.

I didn’t do this to prove AI can make a video. I did it to find out how a creative lead should work when production is no longer the bottleneck. This is what I learned.

01 · Baseline

Same team, one month apart

Part 2 is 20% longer with 55% more shots, and it took one extra day for about $19 more. Here's how the two compare.

A year earlier I made “Jenna Ortega” for the artist Stellar. It won the Artist’s Choice Award at the 2025 OpenArt Music Video Awards. It was the same kind of brief: an original song, a lead character, a world, and a story with a turn. What changed was the process.

Oct 2026 · Don't Look Down

Nov 2026 · Don't Let Go

Length

1:15

1:30

Time

About 2 days

About 3 days

Generation cost

About $137

About $156

Shots in the cut

Stills

About 280

400+

Video clips

About 107, 38 used

80+, 59 shots used

Starting point

A blank chat

A Claude skill built from Part 1

How we worked

One long chat over four days

A crew of sessions sharing one notes file

Feedback

A review page from the first rough cut

A storyboard review board from day one, then a review page on every cut

Camera

Decided shot by shot

A DOP plan: one camera style per world

The extra spend went into design, not waste. The pod Luna lies in went through nine designs and cost around $25 of stills on its own (more on that in section 05). Video was the opposite: a $10.75 model test up front meant far fewer clips thrown away per shot used.

It isn’t a like for like test. The 2025 film is longer, and it was made under competition rules with sponsored credits. Per second of finished film, it’s still about three times faster.

The bigger change is in where my hours went. In Part 1 I was still learning what to ask for. In Part 2 most of the asking was already written down, so I spent the time on the story, the look and the performances.

Speed isn’t the biggest change, though. In 2025 I carried every box myself. In 2026 the AI is the crew, and my hours go into decisions. The taste that won the first award is what directed the second film.

02 · The playbook

Part 1 became a skill

Six phases

The most useful thing I made after Part 1 wasn't a video. It was a Claude skill: the whole workflow written down as phases, rules, scripts and lessons, so any new session starts from it. It covers six phases: brief and song, world bible and look, storyboard and stills, motion, edit and finish, delivery and launch. Each phase ends with a handoff file, and the next session reads those files first. It also carries the tools from Part 1: the Leonardo API client, the dot matrix HUD font, the After Effects build script, the vertical cut and the covers.

The cost of an idea has collapsed, and the cost of a bad idea has too. Ten full versions in two days means you can test creative the way you test copy. The question is no longer “can we afford to produce this?” but “do we know what we want to say?”

Lessons baked in

The lessons part mattered most. A few that Part 1 taught me the hard way, now rules from day one: decide every delivery format on day one (vertical was an afterthought in Part 1 and cost two re-renders). Listen-check that every section is actually sung before locking the song. Attach the face anchor to every generation from shot one. Build the review page from the first rough cut, not the fifth. Run one session per phase and hand off, instead of one giant chat. Part 1's Claude cost was mostly the chat re-reading its own history.

Taste is now the bottleneck, and that’s good news for us. Every hour I’d normally spend keyframing went into direction: the idea, the casting, the performance notes, which take, which frame, when to stop.

Paid off straight away

The skill paid off straight away. Measuring the song caught that Suno never sang the break line or the "hold" hums. Instead of fighting Suno, the line became on-screen type and "HOLD" moved into the HUD.

The edge isn’t the prompt. Prompts and models change every month. What lasts is the system around them: a written bible, a measured edit, approval gates, and code for anything that has to be exact.

03 · Operating model

A director and a crew of sessions

Director · Me

The idea, the look, casting, every approval, the edit decisions, the grade, the finish

Producer · Claude

A crew of sessions (storyboard, pod design, QC, motion, edit, launch): brief, world bible, storyboard as data, every prompt, the beat map, Python graphics, review tools, Premiere assembly, the vertical cut, thumbnails, this write-up's notes

Specialists

Suno, Seedream 5.0 Pro, Nano Banana Pro, Seedance 2.5, Hailuo 03, Wan 3.0, via the Leonardo API. The song, stills, small surgical fixes, motion and lip sync.

The two seats stayed the same, but the producer seat became a crew. Instead of one long chat, each phase ran in its own Claude session, with sub-agents for the heavy reading.

The model doesn’t depend on a particular tool. The producer seat orchestrates whichever specialist models are best this month. Swap any of them and the structure still holds.

The sessions don't share memory, only files. So the project has one notes file every session reads first (CLAUDE.md), a running handoff log, and a rule: one writer per file. That's how a new session picked up the pod design or the motion batch without me re-explaining anything.

The model doesn’t depend on a particular tool. The producer seat orchestrates whichever specialist models are best this month. Swap any of them and the structure still holds.

I learned this one mid-project. Early on, the storyboard chat and a pod chat ran side by side and drifted, including two running cost totals that had to be reconciled. After that, it was one main session with helpers that report back.

The model doesn’t depend on a particular tool. The producer seat orchestrates whichever specialist models are best this month. Swap any of them and the structure still holds.

Claude also does QC now. Every take gets checked against a list before I see it: direction of movement, faces in frame, screen colours, hands, and the cat staying kitten-sized. My time is for taste, not spotting extra fingers.

The model doesn’t depend on a particular tool. The producer seat orchestrates whichever specialist models are best this month. Swap any of them and the structure still holds.

04 · The production system

Seven things that held up

One phrase doing three jobs

Part 1 was "to the moon": getting high, number go up, a literal launch. Part 2 is "don't let go": hold your bags, stay alive, hold your breath. The market went red, so the whole film went red. Every shot and graphic had a test: does it say hold? That gave us the hands motif (paw meets hand, paw on the glass, the rescuer's glove) and the O2 clock in the HUD.

The brief centred on one phrase doing three jobs. “To the moon” is euphoria, a valuation going up and a literal launch. Every shot hangs off it. AI will happily generate a hundred directions. A sharp brief is what makes the one you pick mean something.

For brands: one sharp line gives every asset a yes or no test.

Review the story before anything moves

This was the best new thing. Claude built a storyboard review board: the song plays, each shot's still comes up on a monitor, and the shots sit on a timeline to scale. I tapped OK, Change or Cut on each card and left notes as it played. It felt very much like working with a creative team. The board went through 44+ versions at the same link, and every note was saved, so Claude could read them back and make the next version.

For brands: an animatic with the real audio, on stills that cost cents, is where the story should be approved.

Give every world a camera

Before any video, Claude wrote a DOP plan. Every shot is tagged by the world it's in, and each world gets one camera style: Interstellar as the house look (locked off on the moon, no shake in a vacuum), Dune for the creepy watched shots and the Boss, Blade Runner 2049 for the red void inside her head, First Man for the crew on Earth. Camera and acting calls got so quick, because most of them were already made.

For brands: a camera rule per setting keeps 60 shots feeling like one film.

Sheets before shots

The biggest lesson. Every recurring thing gets a sheet before any scene shot: the face anchor, Luna in the pod, her expressions, the cat, the zombie astronaut, the crew suit, the four characters in costume, the pod itself and the tank. The pod design kept changing mid-shots and cost about $25 in re-rolls, and it turned out old pod references were leaking into new prompts. Once one locked pod sheet was the only reference, the shots matched.

For brands: lock each product or character with one approved reference, and retire the old ones in writing.

Test before you spend

Before the motion batch, we tested the models on 2 shots for $10.75. Seedance 2.5 became the default, Hailuo 03 did the simple shots for about 8 cents a clip, and Wan 3.0 did the lip sync because the lip motion and expressions just looked better. The prompts stayed short: the camera and the beat in plain words, nothing over-directed.

The HUD, radar, countdown, warning bars and titles are not AI generated. Claude wrote them in Python: an original dot matrix font defined as data, plus small generators for each graphic. A change like “make the countdown say IMPACT” was one line and a 4K re-render.

For brands: a small paid test is cheaper than a batch on the wrong model.

Anything that must be exact is code

The shot that inspired the film: Luna in an oxygen mask with a full HUD. She can't lip sync with the mask on, so the lyric types on inside the HUD, timed to the sung words from the measured vocal. Then the HUD powers down dot by dot at the end. All of it is Python, so a timing change is one line and a re-render.

For brands: type, legal lines and end frames stay out of the diffusion model.

Review every cut the same way

Same idea for the edit. Every rough cut came with a review page: the cut plays next to a card for every shot, and I tapped Keep, Fix, Re-roll or Trim, or left a note at the exact second. Claude read it back, made the next cut and listed what changed, so I only had to watch those bits.

For brands: a decision log on every cut means no note gets lost between rounds.

05 · What didn't work

Nine pods

The pod Luna lies in was the hardest thing in the whole film. It went through nine designs over two days and ate about $25–30 of a $56 stills budget. Nearly all of that was one mistake repeated, and it took me too long to spot it.

Here's how it went

1. A round escape pod with a porthole. Fine as B-roll, but it didn't feel like somewhere you'd keep someone alive.

2. A tunnel tent. I asked for her pod to sit inside a rescue tent. Claude invented a ribbed tunnel tent to match the first interior it had generated, and every shot after that inherited it. I never asked for a tunnel tent.

3. A glass tank from my own references. It looked fake: a 2.5 m tank in a 2 m tunnel, far too clean for a survivors' camp, flat pink fluid, and dead centre in every frame.

4. A top-down capsule in a bigger geodesic dome. I liked the elegance and the rubber floor mats. But the proportions squashed and the fluid level changed from shot to shot.

5. A hex window showing only her face in the mask. Closer, but it only worked as a close-up.

6. An upright hex pod and then a tilted one. They still felt odd and didn't sit in the room.

7. A hex pod sheet, locked and then retired.

"you kept using the wrong asset for the pod. that's why they all came out weird. stop wasting my credit and do it properly."

What I told Claude at that point

Claude went back through the logs and I was right. The pod prompts had been passing the old round-pod interior and a sphere-pod sheet as references the whole time. The tilted tests used the capsule room as the set. So every new design had the old shapes leaking into it, and my own locked pod shots were never used as the design source.

The fix was boring. One design (the cradle: a glass-front, metal-back cylinder leaning back at about 30°), one pod sheet and one Luna-in-the-pod sheet as the only references for every pod shot, plus a written list of retired assets nobody is allowed to pass again. From then on the shots matched.

What I'd do differently: lock the design on one sheet before any scene shot, check which references each prompt actually sends, retire old designs in writing, and never let the AI add set pieces I didn't ask for.

The rest of the list

What broke

Why

What fixed it

The kitten-sized cat kept coming out huge

Models make the hero subject bigger, and the layout reference fixes its size even at low strength

After several rounds of edits, my own Photoshop fix

Luna's face drifted in close-ups

A plate passed as a reference carries its face, and an edit of an edit gets worse

Face anchor first at high strength on every frame. For the ending I said stop and reused the original plate

$13.10 on clips that looked like After Effects motion graphics

Reference mode on the wrong shots

Re-made those shots, and it went into the lessons file

My storyboard notes from v2 onwards were lost

The review page only saved when you pressed Save

Autosave 1.5 s after every edit

A rewrite pass replaced the two shots that inspired the film

Nothing protected them

A rule, checked in code: those shots never get replaced

Two chats drifted apart, with two different cost totals

Parallel sessions don't share memory

One main session, helpers that report back, one notes file

The whispered "look up" ending in Suno

Five fixes, all worse than the original

Kept the original. Try one fix, then stop

06 · Stepped in

Where I still stepped in

Claude did more of the heavy lifting this time, but a few things were faster by hand.

Photoshop when a model wouldn't listen

The kitten-sized robot cat kept coming out huge next to the pod. After a few rounds of edits, I fixed it myself.

Lip sync in Leonardo's web app

The API doesn't take audio, so I made the two lip sync shots on Wan 3.0 in the web app, from vocal sections Claude cut for me.

A few of my own clips

Like the opening and ending in 16:9 and a flower insert.

The edit and the grade

Claude assembled the cuts in Premiere and built the After Effects finish to the beat map. I shuffled the intro, trimmed frames and finished the grade by hand.

I also said stop a few times. When the ending shot's face kept drifting, I told Claude to go back to the original plate and open her eyes in the video instead. Knowing when a fix is costing more than it's worth is still my call.

07 · Guardrails and limits

What I held to, and what still needs people

Guardrails

No real person's likeness and no third party logos, in prompts or outputs. The Boss is only ever a back view or a silhouette. I approved every still, every clip and every cut. Every paid batch was quoted first and needed my yes. AI use is disclosed on publication, and I post everything myself.

No real person’s likeness and no third party logos, in prompts or outputs. A human approved every still, every clip and every cut. AI use is disclosed on publication. No credentials or posting handed to the AI. I publish myself.

Where it still needs people

Faces drift: edit a frame without the face anchor and her face changes, and an edit of an edit gets worse. Scale: models make the hero subject bigger, and the cat grew in almost every new set. Design changes mid-shoot are expensive, which is why the pod cost the most. Parallel chats drift, because sessions only share what's written down. And taste: the model offers, but it doesn't know when something is right. That call was mine every time.

Faces drift without a locked reference on every generation. Hands and crowds still break. Scale in wide shots goes wrong easily. And taste: the model offers, it doesn’t know when something is right. That call was mine every time.

What this means for creative teams

What a month changed

A couple of new models helped, but the bigger change was the system around them. Part 1 proved one director and an AI producer could make a music video. Part 2 showed that once the process is written down, the next one starts further ahead, and the time saved goes back into the story.

A small group with clear direction, a written brand system and AI producers can now make work that used to take a studio. That isn’t the end of craft. It moves craft up the chain, to the brief, the system, the approvals and the finish.

The review board and the DOP plan are the parts I'd bring to any creative team tomorrow. They let me make the big calls early, on cheap stills, with the music playing, instead of finding problems in expensive video. The next one starts from an even better playbook.

The process changed almost completely in twelve months. The idea, the taste and the storytelling didn’t, and they’re still what wins.

Bring this to your team →

Previous: Part 1, Don't Look Down →

Wellington, NZ