This video? Same agent.

It opens on an agent that forgets, turns to one that doesn't, shows what that memory is plugged into, and then says why The Lab exists: AI moves too fast to read about, so it measures instead. The pages under those words are the real site, recorded.

Near the end the music keeps driving and the screen says This video? Same agent. That line is only true if the agent makes everything, so that was the rule from the first question. It wrote the script, made the music, drew the diagrams and titles, recorded the site, prompted the three cinematic shots, cut all of it, and checked the result. I answered questions, picked between options, approved each paid step, and gave notes. I never opened a video editor. If I had, the line would have come out. The next build log is what giving those notes was like, and the screen built to fix it. A later one is the same agent making a whole concert film from a song.

The agent is the editor. The human is the client.

Before a cent was spent, the agent ran a short interview on the taste calls: do I appear in it (no, not even my name), what the music should feel like, what picture carries the memory beat, what the last line is, and whether to keep the Same agent twist. Each question came with its own suggested answer, so most took one word to settle. Each answer went into a notes file before the next question, so nothing lived only in the chat.

The method is borrowed from Nate Herk's grill-me skill^[1]. Then the agent pitched the story as a beat sheet, and later, when the story needed rebuilding, as three competing beat sheets with a recommendation. It waited for a pick both times.

Every title, the chat window that keeps forgetting, the memory diagram, the sources scene and the end card are drawn by code: a small web page paints each frame for a given moment in time, and a headless browser saves thirty of those frames a second. The same scene renders identically every time, every label is sharp, and changing a word costs nothing. That is why the notes after version four were free.

The agent scrolled real pages in a headless browser at phone size, one exact scroll position per frame, so the motion is perfectly smooth and a re-recording gives the same shot. Every number on screen was re-checked against its live page on the day the reel was rendered. Real material is the whole argument of the site, and it costs nothing.

The paid work went to two services. Music came from Suno, reached through kie.ai^[2]: 12 credits for two 45-second instrumental takes. Four rounds made eight takes; the first direction, dark electronic, was "too ominous," and the pick was a hip-hop and electronic hybrid. A fifth call, much later, extended the kept take for a longer cut; that version didn't ship. The shots came from Seedance 2.5 on Higgsfield^[3], turning a still image into five seconds of motion: a lamp flickering on in a garage, a slow push toward the open door.

Each shot was drafted at 480p for about $1.03, checked frame by frame, and only then rendered at 720p for about $2.31. On a phone, 720p and 1080p look nearly the same, and 720p costs less than half as much. The opening shot's draft ignored its prompt: asked for a sine wave, it drew a flat line. That flatline said "starts from zero" better than the prompt did, so it was kept, until a later story no longer needed it.

Take 7 runs at about 95 beats a minute, and its first big hit lands at 7.61 seconds. The opening runs exactly until that hit. After it, every segment is a whole number of beats long, so every cut lands on a drum. The edit is one table in one script: each row names a segment, its length in beats, what's on screen, and which words sit on top. Change a row, rebuild, and the whole reel stays in time.

The first measurement put one beat at 0.628 seconds. The true figure is 0.6317. Four thousandths of a second is nothing on one cut, but it adds up: by the close of the 44-second version, cuts were landing about a fifth of a second late. Measuring across 16 and 32 beats at once, instead of one, found the true length.

The music plays straight through at full volume and fades at the end. An early cut dropped to silence for the twist and then restarted the track; to a viewer that sounded like a broken file. A softer dip didn't work either. When the slower v13 ran 12 seconds past the end of the song, the agent scored every pair of bars for matching drums and tone, found an 8-bar stretch whose two ends match, and plays it twice, joined on the beat. Drama comes from the picture now.

Every generated clip was pulled apart into frames and inspected for warping, melted objects and invented text before it was shown to me, a habit taken from Nate Herk's walkthrough of the same tools^[4]. When I later tried a cheap vision model on that job, it missed every visible mistake in the video I tested. Every cut got a contact sheet, a look at its first frame, and a brightness measurement of each segment. Every generation got a row in a cost ledger, kept or rejected, with the reason. Each mistake I caught became a check the build now runs before anyone sees a cut: a highlight that misses its word, a font that never loaded, a collage that repeats itself, a page that gives the ending away, a black patch where a page should be.

1. 2
   The audio drops out at 28 seconds, then starts over.

   It was designed that way: silence for the twist, then the drop again. A viewer hears a broken file. v3 tried a muffled dip, which didn't work either. v4 plays the track straight through. Approved.

2. 5
   Two people who had never seen it didn't get the well, the keys, or the logo.

   I showed v4 to two family members and watched them. The metaphor shots didn't translate, a shot of cables was boring, and the logo beside the words read as a letter: "L The Lab?" v5 drew the real things instead: a chat that forgets, the memory diagram, and the logo above the name.

3. 6
   The sources background is meh, and the grid is missing a tile.

   A new drawn scene wires a claim to six source cards. The tile grid's ninth tile was due on the scene's last frame and never appeared; they now land faster and all nine hold.

4. 8
   "155 days" and "The models." don't say anything.

   A number that will be stale the day after posting, and a label the picture already carries. Cut. v9 cut "The data." for the same reason.

5. 11
   Make it clear the logo is a logo.

   The mark now sits inside its own rounded app-icon tile and draws itself like a chart, the upright first, then the baseline, then the red point on the beat. Nobody reads a chart as a letter.

6. 12
   "What's this reel for?"

   The big one. The reel bragged about an agent, then showed a website, and never said why the website exists. "Every AI chat starts from zero" was also no longer true. The agent pitched three new stories; the pick connects the agent to why The Lab exists and what a visitor gets. Then four smaller notes: the homepage headline gave away the closing line early, the first frame was an empty box, "chats" should be "agents," and some blacks were gray.

7. 13
   Too fast to follow.

   Two more people who watched it said the words went by too quickly to read. v13 gives every title at least 1.9 seconds, trims the longest line, and turns the three memory numbers into big figures a viewer reads at a glance. The reel grew from 44 seconds to 58, longer than the song, so the agent replayed an 8-bar stretch of it. Two alternatives were built and set aside: v14 used a paid extension of the song, 6 cents for two takes, and v15 re-timed every scene to the song's hits. I kept v13.

8. 18
   The diagram lost its red lines, and the logo sits low.

   v16 and v17 rebuilt the reel on a new animation kit, HyperFrames, with the same script, music and cuts. The rebuild dropped the red traffic that ran along the memory diagram's lines, and it set the end card below the middle of the frame. v18 sends red streaks from every connected tool through all three memory rings into the agent and back out, and centers the end card. Both fixes were code, and free.

9. 22
   Really amp up the motion.

   Same story, words, cuts and music; everything else moves. The camera punches in on the drum hits, the scenes turn in 3D, each page hands off to the next in a different way, and one word per title gets a highlight on the beat. Then four smaller notes over three versions: the glare on each page jumped from side to side, so it now follows the page's tilt as if one light were fixed in the room; the rings around the memory numbers repeated too much, so one is left; the lines to the sources ran behind a card; and a highlight came up short of its word, because it was measured before the font had loaded. Last, the collage near the end repeated the reel's own frames. It now shows sixteen different posts and charts from The Lab. All of it was code, and free.

Two notes turned out to be about something other than what they seemed. The grays were real: pieces saved on two different brightness scales lift black to gray when they're joined, and the agent measured it (11 on one segment, 26 on the next) before fixing it. The circles that looked taller than wide were not: measured, they were exactly round. Labels wider than the rings were notching them. Shorter labels fixed it.

Every mechanical check passed on version four. What found the real problems was watching two people see it cold. After that, every cut shipped with the same reminder: show it to someone who has never seen it, and ask one question. What's this for?

The skill lives at github.com/WorthingtonCloud/sizzle-reel under the MIT license. It works with Claude Code, or any agent that can run commands on your computer and keeps a folder of skills. Paste this to your agent:

Install the sizzle-reel skill from https://github.com/WorthingtonCloud/sizzle-reel. Clone the repo and copy its sizzle-reel/ folder into my skills folder (for Claude Code: ~/.claude/skills/sizzle-reel). Then read its SKILL.md, set up a reel folder from its kit, run node doctor.mjs, and tell me what's missing. When everything passes, interview me for my first reel.

It needs Node, Python with numpy, and ffmpeg on your computer; the doctor script checks all three and says how to fix what's missing. Paid shots and music need a kie.ai key and, for video clips, a Higgsfield key. A reel built only from drawn scenes, recordings of your own pages, and your own footage needs neither.

1. 1
   Answer the interview.

   About eight questions, one at a time, each with a suggested answer: what the reel is for and where it will post, what a stranger should get from it, the one line they should remember, what real pages or footage you have, the music, whether you appear, the budget, and who you'll test it on. You'll see a notes file fill up as you go.

2. 2
   Pick a story.

   The agent pitches two or three beat sheets, with what each costs, and recommends one. Nothing is built until you pick.

3. 3
   Approve the $2 test, and pick the music by ear.

   Two music takes for about 6 cents and one cheap draft shot. Your ears decide; the agent can measure loudness but can't hear taste. Ask for another pair in a different direction if the first misses.

4. 4
   Wait for version one.

   The agent draws the scenes, records your pages, finds the beat grid, drafts the paid shots, and cuts it. It checks the result before you see it and tells you what it cost so far.

5. 5
   Give notes in plain words.

   "This part is boring." "That word is wrong." "The logo looks like a letter." Most notes cost nothing. The agent names the cost of any that don't before doing them, and every version gets a new number.

6. 6
   Show it to someone cold.

   Someone who has never seen it. Watch their face, then ask what it's for. If their answer isn't yours, that's the next note.

7. 7
   Post it with its cover.

   The hook is on screen from the first frame, which is what chat apps show before anyone presses play. Where the platform lets you pick a cover image, upload the one the build made.

The paid script refuses to spend without an explicit yes, and refuses to go past the budget you set. Every generation is logged, kept or rejected, so the total is always one command away.

1. 1
   Never let a video model draw words.

   Titles, labels, logos and diagrams are drawn in code. Paid shots are textless mood only.
2. 2
   Show the real thing, not a metaphor.

   A stranger decodes a diagram of the actual system faster than any symbol for it.
3. 3
   Never stop the music.

   One track, full volume, fade at the end. Silence reads as a broken file.
4. 4
   The first frame is the preview.

   Put the hook on screen from frame one.
5. 5
   Spend the closing line once.

   Don't show a page whose headline is your closing line before the close.
6. 6
   Cut anything that dates itself or says nothing.

   Let the picture carry it.
7. 7
   Fill every grid, and match every black.

   A hole or a gray patch looks like a mistake, because it is one.
8. 8
   Put the logo in a tile, above the name.

   Beside the name, a simple mark reads as a letter.
9. 9
   Give every word time to be read.

   About two seconds for a short title, more for a long one. A number reads faster as one big figure than as a sentence.
10. 10
    Put the motion on the picture, never on the reading time.

    Move the camera, the scenes and the hand-offs between pages. Leave every word on screen long enough to read.
11. 11
    Show your own work in a collage, never the reel's.

    Every tile a different post or chart. A wall of frames the viewer just saw reads as repetition.

Every generation was logged as it happened, with the service, model, resolution, length, cost, and whether it was kept. kie.ai reports credits, so the music and still costs ($0.51 together) are measured. Higgsfield has no spend report through its API, so the video costs ($12.33 together) are estimated from its published token pricing: height × width × seconds × 24 ÷ 1,024 tokens at $0.0214 per thousand. The billing page is the source of truth. The beat grid (first hit at 7.61 seconds) was measured from the track and confirmed by ear. The beat was first measured at 0.628 seconds and corrected to 0.6317 by measuring across 16 and 32 beats; with the corrected grid, the replayed stretch joins within a millisecond of the beat. The black-level figures (11 and 26 on a 0–255 scale) were measured on a flat patch of each segment. The session time is from the conversation timestamps.

This is one reel, made for one site, judged by one person's taste and two people's reactions. The skill packages what this build learned; it has not yet been used by a stranger on a different subject, and the first person who does will find a note this reel never needed. Seedance and Suno change often, and their prices and behavior will drift from the figures here. The agent's session time does not count the time I spent watching cuts and thinking about notes.

• Rev 1 · Sep 28, 2026First publication, with the 44-second v12.
• Rev 2 · Sep 28, 2026The reel is now v13, 58 seconds, after two viewers found v12 too fast to follow. Updated to match: the tally, the version count, the notes, the costs, and the beat. The first beat measurement was slightly off, and the entry now says so.
• Rev 3 · Sep 29, 2026The reel is now v18, rebuilt on a new animation kit, with the memory diagram's red traffic restored and the end card centered. Updated to match: the version count, the notes, and the contact sheet.
• Rev 4 · Sep 29, 2026The reel is now v22, the motion pass: same story, words and music, with a camera that moves on the drums, pages that hand off in 3D, and a collage of The Lab's own work. Updated to match: the version count, the notes, the checks, two new rules, and the contact sheet.
• Rev 5 · Oct 11, 2026One sentence added near the top linking to the next build log, about what giving those notes was like.
• Rev 6 · Oct 12, 2026One sentence added to "It checks its own work," linking to Field Report 005's test of a cheap vision model on the same job.
• Rev 7 · Oct 26, 2026One sentence added near the top linking to Build Log 010, the same agent making a whole concert film.
• Rev 8 · Nov 11, 2026One sentence added to the notes thread linking to Build Log 011, a 3D film made on the same kind of notes.

1. ExtNate Herk, grill-me, a Claude skill: interview one question at a time with a recommended answer each, and write every answer to a file before asking the next. The interview-first rule and the preflight check in the skill also follow his scroll-craft (MIT).
2. Extkie.ai's Suno API: model V6, instrumental, custom mode; 12 credits per call returning two tracks. Stills from Seedream 5 Pro on the same service.
3. ExtHiggsfield API, ByteDance Seedance 2.5 image-to-video: 5-second clips at 480p and 720p, priced by the token formula above.
4. ExtNate Herk, video walkthrough connecting Claude to Higgsfield: the agent inspects frames from every generated clip before the human sees it, and keeps a running cost log.
5. DataThe reel's build: segment table, drawn scenes, recordings, titles and ledger, retained. The public skill and its kit: github.com/WorthingtonCloud/sizzle-reel.