I was thirteen the first time somebody handed me a movie camera. It was a Super 8, it came out of a gifted program, and nobody explained anything about it — which turned out to be the right way to do it.

There was a film contest that year for seventh through ninth graders, judged and screened at Arizona State. I entered a comedy. It won its category, and I sat in that auditorium and heard a room full of strangers laugh in the places I had decided, weeks earlier, alone, that they were going to laugh.

That is the entire reason I still do this. Not the trophy. The gap between deciding where the laugh goes and hearing it arrive.

Everything since has been the same job on different equipment. Super 8 became VHS. VHS became digital. Digital became Premiere. Premiere became Premiere plus a row of AI generators that didn’t exist four years ago. Four machines. Same job.

The machine changed four times. The story never did.

I just want to make a fish talking to a crab

Somewhere out there is a person who has already lost an afternoon to this.

They went looking for how to make one funny video. What they found was a benchmark. Then another benchmark. One tool renders four tenths of a millisecond faster. Another charges six hundredths of a cent less per credit. A third boosts the resolution. A fourth arrived with a launch film promising to change everything, and has since been quietly cancelled.

None of that is an answer to the question they actually asked, which was:

I just want to make a fish talking to a crab. Can you help me?

That’s who this is for. Not the person comparing render speeds. The person who has an idea about a fish and a crab, and forty minutes, and no interest whatsoever in which GPU it runs on.

What this is

The whole process. Not an overview, not the parts that make a good demo — the whole thing, in the order you actually do it, including the settings, the numbers, and the parts where it goes wrong.

I make a real marketing video for a real small business, most days. The recurring one is a bail bond office run by my wife Nina. I picked that as the example on purpose. It is the least glamorous business anyone can name. If this works for a bail bond office, you can stop wondering whether it only works for pretty products.

A finished reel takes me thirty to forty minutes. That number is real and it comes after forty-five years of reps. Your first one will take you a weekend. Anyone who tells you otherwise is selling you a subscription.

What this isn’t

I don’t sell any of the software in here. I’m not affiliated with any of it, I get nothing if you sign up for it, and there isn’t an affiliate link anywhere in this guide. I pay for these tools every month like everybody else, and where one of them is bad at something, that’s in here too.

I mention it early because if you searched for how to make AI marketing videos, you already met the alternative. Nine of the top ten results are companies that sell AI video software, explaining that the answer is AI video software. They’re not lying to you exactly. They’re grading their own homework.

There’s nobody on that page who actually does this for a living. That’s the only reason this guide exists.

The one thing to understand before Chapter 1

Every tool in here is a machine. Machines get replaced. The one in Chapter 9 will be a different one in two years and it won’t matter, because none of the things that make a video work live inside the machine.

They live in whether you know who you’re talking to. Whether the joke lands on the last word. Whether the face in this week’s video is the same face as last week’s. Whether the thing you made is actually about your business, or just about a holiday.

Learn the machine. It’ll take you a weekend. Then spend the rest of your time on the part that doesn’t change.

How to use this

• Starting from nothing? Hit Expand all and read front to back. It’s organized the way a production is organized — you’re the producer in Part One and the editor by Part Seven. About two hours.
• Already shooting? Don’t read it. It’s built to be closed. Every chapter stays shut until you need it, tells you in one line what’s inside, and opens to exactly the setting you came for. Chapter 14 is the one most people need first.
• Want it on paper? The button at the top prints the Shoot Sheet — every checklist and setting, no prose. That’s the part worth printing. The teaching is in the videos, and the videos don’t fit in a briefcase.

Flo Lopez · VideoReals

A producer doesn’t ask whether something will look good. A producer asks whether it’s worth making at all.

Before you touch a single tool, three answers have to exist.

• Who is this for. Not “people in my town.” One person — the customer you’d take ten more of. For a bail bond office, that’s a mother at eleven at night who has never done this before and is terrified. That’s who the video is talking to, even when the video is about cupcakes.
• What it needs to do. Every video does exactly one of four jobs: get you found, get you remembered, get you trusted, or get you called. A video that tries all four does none.
• Whether it’s worth the forty minutes. The question nobody asks, and it’s the whole job.

You are not choosing whether to post. You are choosing which one.

This is the part that surprises people. Any given day has five to fifteen holidays on it. September 5th alone is Be Late for Something Day, International Bacon Day, National Cheese Pizza Day, and World Beard Day. You are not staring at a blank calendar deciding whether today is worth it. You are looking at a list and ranking it.

Which means the real skill isn’t discipline — it’s ordering. Three tests, in this order, and the first one outranks the other two.

Test one is worth more than the other two combined, and it’s covered in the next chapter because it’s the whole reason this format beats generic content.

And the real skip

Separate from ranking, some occasions come off the list entirely. A real historical event with real dead people in it is not a comedy prompt. That’s a judgment call, not a creative one, and there is another holiday tomorrow — usually four of them.

The holiday is never the subject. The holiday is the excuse to be in front of someone today.

Nina whittling matches on Matchmaker Day is not about matches. It's about a woman who does painstaking, unglamorous work all day so that two things end up joined together. That's the bail bond business, told sideways, without one word of explanation.

The three rungs. Climb all of them or don't post.

• The holiday — the hook. Gets the scroll to stop. Costs nothing, means nothing on its own.
• The business — the turn. The holiday's logic and your business's logic overlap somewhere. Find the overlap. Don't announce it.
• The community — the landing. A real place, a real neighbor, a real business down the street. Almost nobody climbs this rung, and it's the only one a competitor in another state can't copy.

Rung three is what makes the video yours. Anyone can generate a woman holding a cupcake. Nobody else can tie it to the bakery on the corner where your customers actually go.

Climbing rung three — how the tie-in actually works

Anyone can generate a woman holding a cupcake. Nobody else can tie it to the bakery on the corner where your customers actually go.

This is the rung almost nobody climbs, and it’s mechanical once you see it. Four parts:

Two things happen when you do it properly, and only one of them is the one people expect. They might like and share it — that’s the obvious one, and it’s real. The bigger one is what the post says about you: that you are an actual business with a presence in this community, one that notices other small businesses and is willing to spend its own post on them.

That second effect lands on every viewer, including the ones who’ve never heard of the bakery.

Ideas aren’t the bottleneck. Clarity about who you’re serving is the bottleneck.

Once that’s locked, a day’s idea takes about four minutes. It happens in three stages, and they are not interchangeable: you scan, the AI generates volume, and then you pick — and the pick is almost never one of the options you were handed.

Stage one — the scan

Open the month’s list. Read down it. You are not asking “is this holiday interesting?” You are asking one question only:

Can I see my hero in a picture here?

Run down a real day and watch how fast it sorts:

• International Cabernet Day — images arrive immediately. Keep.
• Skyscraper Day — obviously someone climbing. Keep.
• US Bowling League Day — a lane, pins, a rented shoe. Keep.
• College Colors Day — nothing arrives. It doesn’t do anything for the imagination. Skip.
• Yamashita Surrender Day — a real historical event with real dead people in it. Skip, on judgment.

Do not force the ones that don’t arrive. If you find yourself reasoning your way toward an image, the image isn’t there. Three or four keepers out of a day’s list is normal and plenty.

Two entry points, and the scan tells you which

Sometimes a holiday hands you a phrase. Sometimes it hands you a picture. They are not the same job and pretending otherwise is why people get stuck.

• Line-first. The occasion suggests words. Cabernet Day and a horse at a bar hands you “Why the long face?” before any image exists. Write the line, then build the frame that lets it land. This is the default and it’s what the standing prompt above is built for.
• Image-first. The occasion suggests a picture and the picture is already funny. Skyscraper Day hands you a man climbing a building as though it were a normal commute. Here the goal is one image that is funny by itself — the line is icing on the cake.

The step between the idea and the prompt

You rarely go straight from “that’s the one” to a finished image prompt. There’s a conversational pass in between where you bring a half-formed picture and ask for it to be fleshed out.

The half-formed part matters. You are not asking for an idea from nothing — you are handing over a rough one and asking for it to be built out and beaten. Describe it messily, in whatever order it arrives, and include the bit you can’t quite make work.

Two things in that template do the work. “That’s just an example, but you get the idea” gives permission to depart from your version. “If you have something funnier, don’t hold back” asks it to try to beat you. Without that second line you tend to get your own idea handed back politely.

Turning the concept into a prompt

You have the picture in your head. Now ask for it to be written properly. Say what you see, name the reference that gives it a feeling, and — this matters — ask for the setup frame, not the punchline frame.

Stage two — make the AI generate volume

Three lines is not a shortlist. It’s a guess.

The mistake almost everyone makes is asking for the answer. Ask for the raw material instead, and ask for far more of it than you intend to use. The prompt below is a standing one — the same every day, with only the occasion changing.

Four things in it do most of the work:

• Titles first. Before any line gets written, ask for five to seven comic role titles for the absurd party, drawn from your business’s own vocabulary — The Top Dog, General Muttager, Director of Pawformance, Chief Fetching Officer. That list is an engine, not output. Most of the lines fall out of it.
• Twelve lines minimum, with the posture varied — deferential, exasperated, conspiratorial, resigned, over-formal, defeated. Same situation, different relationship to it. The posture is what produces genuinely different jokes rather than twelve rewrites of one.
• A two-hander option. One short exchange where the absurd party answers with silence and the star responds to the silence. Silence is a punchline the image can act.
• Show the rest. The unpicked lines aren’t waste — they become post copy, comment replies, and future posts.

The pre-flight, run on every line before it survives

The shot — physical facts only

A generator’s default is the most common photo of that scene, and adjectives don’t override it.

Never write mood words. For every person and animal in frame, write where the eyes point (and “not looking at the camera” when they shouldn’t be), the body orientation, and what the hands are doing with an endpoint.

Ask for the whole package in one pass

Staging this across turns wastes your morning. One request, one complete return: pre-flight results, the titles list, twelve-plus numbered lines with the pick defended, the two-hander, the image prompt, the post copy, the animation prompt with its negative prompt and break script and beat map, and any local angles last with sources cited.

Stage three — the part nobody tells you

The AI’s job is volume. The pick is yours, and it is usually a hybrid.

Here is what actually happens, and it is worth being honest about because every tool vendor implies otherwise. The results are sometimes helpful. More often, one option has something you like and a different option has something else appealing, and you assemble a third thing out of the two — something you hadn’t thought of and the AI hadn’t suggested.

That’s not the system failing. That is the system working exactly as intended. It’s a brainstorming partner, and brainstorming has always worked this way: the value is in what the volume jogs loose, not in any single suggestion. Suddenly the funniest pick of the day is obvious and there’s no contest with the others.

Fifty-two videos with one face is a brand. Fifty-two videos with fifty-two faces is noise that happens to be well-lit.

This is the single largest gap between what the AI video tools sell and what a business actually needs. Every vendor will hand you a stock avatar. A stock avatar is a stranger. Your customer needs to see your face, or your spokesperson’s, or your mascot’s — the same one, every time, until it’s familiar enough to be trusted.

Three kinds of consistent character, all valid

• The owner. Strongest trust, hardest to sustain if you hate being on camera.
• The spokesperson. An employee who’s good on camera. Works well; carries a risk if they leave.
• The mascot. Badly underserved and often the best answer. A character can’t quit, can’t age, and can do things a person can’t.

How many photos, and which ones

The training set is not a portfolio. It is an instruction.

Ten photographs is not enough room to be sentimental about which ones you like. Every frame you upload is a vote, and the model counts votes. Pick the ones that teach it something, and leave out the ones that teach it a habit.

Grabs pulled from video are where this goes wrong most often. A frame from a talking-head clip looks fine to you and is almost always mid-syllable.

The ten-photo set

The laughing frame is the exception that proves the mouth rule, and it earns its slot: one genuine laugh teaches the model the range of the face without teaching it a resting state. One. Not three.

Then: the rules for using it

• Trigger word chosen so it can’t collide with ordinary English.
• One named subject per prompt. Always.
• Refer to everyone else generically — “the woman,” “a man in an apron.” The moment you name a second person, the generator starts swapping faces.
• Waist-up medium shots as the default. Footwear is where hands-and-feet failures go to live; frame it out.
• Logo present but subordinate. Background, not billboard.

The generator weights early tokens harder. You want the face to win every conflict.

• Order of operations. Character reference first, then the scene, then the light. Not the reverse.
• Match the light or the head floats. If the scene is warm interior tungsten and your reference is daylight, you get a face pasted onto a room. Describe the light once, plainly, and let it apply to both.
• Don't re-describe the face. Every adjective about the character's appearance competes with the trained reference. Name the character, describe what they're doing, and stop.

Where it physically goes

Straight into the generator’s prompt box, with the actor invoked inline as the first token — not picked from a dropdown afterward.

The anatomy of a working image prompt

Every good one has the same eight parts in the same order. This is the one from the reel above, taken apart:

Scale the detail to the joke. A comedy two-shot where the staging is the setup needs every one of those eight parts. A wide establishing shot that just has to look right can run much shorter. Don’t write four hundred words for a frame that isn’t carrying a punchline.

Negatives worth keeping standard

One output. Not two, not four.

Multiple outputs look like insurance and they are the opposite. The generator is good enough now that the first run is usually right, and when it is wrong it is usually wrong because of your prompt — which means every extra output is another copy of the same mistake at full price.

Before it goes to the animator — and two things not to do

The still is nearly finished the moment you like it. Almost everything people do to it here makes the video worse.

There is a real resolution floor. Hand the animator something soft and you get soft video — the detail can’t appear from nowhere. But it’s a floor, not a slider. Past a certain point you get diminishing returns, and past a further point you get a file the generator refuses. Generate at a sensible size and stop thinking about it.

The one thing worth doing

Fix genuine artifacts, and nothing else. You will sometimes get an otherwise perfect frame with something wrong at the edge — a foot doing something anatomically impossible, a stray figure the generator invented in the background, a hand with an extra finger.

• Open it in Photoshop, select the offending region, and remove it with generative fill.
• Fix the aspect ratio here too if you forgot to set 16:9 — select the empty edges and generate, rather than enlarging and cropping. Enlarging costs you the composition; extending keeps it.
• Quick Export as PNG.
• Load into the animator.

What actually causes a reroll — and how that has changed

Worth being honest about the history, because it tells you where to spend your effort.

• Artifacts — the stray arm, the impossible foot, the invented bystander. This used to be common. It happens less and less as the models improve, and it will keep declining. Do not build your process around it.
• Timing — for a long time this was the number one reason. The line landed in the wrong place, the pause wasn’t there, the cut came early. It is now rarely the reason, and the fix wasn’t better rerolling. It was specifying the timing up front.

Most rerolls are a specification failure, not a generator failure.

That’s the whole lesson of this section. Every reroll you run is the generator filling in something you left blank, and the cheapest fix by an enormous margin is to leave fewer blanks. A longer, more specific first prompt costs you two minutes of typing. A reroll costs credits and five minutes of waiting, and often produces a different blank being filled differently.

If the idea works in one scene, do it in one scene.

• One scene. The default and usually right. A single frame, a single line. The image holds the setup and the line delivers the whole thing. Nothing to cut, nothing to sync, nothing to go wrong.
• Two scenes. When the joke needs a before and an after — a state that changes. Not when you just want more footage. Two scenes cost more than twice as much as one, because now you have continuity to protect.
• Three or more. Reserve for the rare idea that's genuinely a story with a middle. Most of the time this is ambition, not necessity, and it's where a forty-minute build turns into a two-hour build.

When it really is a story with a middle

Occasionally an idea earns more than two shots. When it does, name every beat before you write a word of prompt. Named beats make the structure legible — to you, and to whatever is helping you write it.

Three things in that structure are transferable

• The hold past comfort. Shot 2 does nothing for an uncomfortably long time, and the discomfort is the joke. The trick that makes it read as deliberate rather than broken: keep the ambience running. A stein clinks somewhere. Silence with room tone is a choice; silence with nothing in it is a bug.
• Interrupt mid-word. Shot 4 catches the victim starting a fresh compliment — “Your posture’s really improv—”. Cutting a line off is funnier than letting it finish, every time, because the audience completes it themselves.
• Cut ON the sound, not after it. Same family as cutting on the reaction rather than the line. The crunch lands and you are already in black. The audience feels it in their teeth.

It works because the character doesn't know it's funny.

The rules, in order of how often they get broken

• Write the line first. Always. Build the image around the line, never the line around a picture you happen to like.
• Run the swap test. Take out the absurd element and put in a normal one. If the line still works, the line is dead — the joke was in the picture, not the writing. Write another.
• The absurdity enters as paperwork, not as observation. Your character does not remark on the strange thing. They administrate it. “Forty-one done” is funny. “Can you believe I'm whittling matches?” is not.
• Never wink. Never explain. No knowing look at the camera, no tag line that clarifies the joke. The second you signal that you know it's funny, it stops being funny.
• The last word is the punchline. Rewrite until the payload lands on the final syllable. Anything after it is throat-clearing and it kills the timing.
• Photoreal, always. Never illustration, never cartoon, never ink. And never put “New Yorker style” in an image prompt — it produces ink drawings, which is the opposite of what that phrase means to a human.

Borrow the trope, never the property

A large share of the best occasions are cultural — Star Trek Day, May the Fourth, and so on. These are gifts, and they come with an obvious trap that is easy to walk around.

The trope is public. The property is not.

You do not need Captain Kirk. You need the red shirt — the decades-old joke that the guy in the red uniform is the one who doesn’t come back. That joke belongs to the culture. Your hero standing on a desolate planet in a red uniform, looking hopeless: “I knew I should have grabbed the blue one, even though it was wrinkled.”

No character is depicted. No logo appears. No name is used. And the joke is better for it, because the trope is the part the audience actually shares — the specific character would only have narrowed it.

Timing is the thing people assume they’ll fix in the edit. It is much cheaper to get right upstream.

The shape of a reel

Almost every one of these has the same four movements, and knowing them is what lets you write to a fixed duration instead of hoping.

• The buildup. Something happens, or nothing happens on purpose. Either way the scene needs a moment to be absorbed before anyone speaks.
• The absorb. A beat where the viewer takes in the frame. This is the one people cut, and cutting it is why their reels feel rushed.
• The delivery. The line.
• The expression. A small change of face — sometimes before the line, sometimes after. Mild. It is not a reaction shot; it’s a flicker.

How long

Kling will let you go to fifteen seconds. Don’t. Ten or eleven is the working range, twelve is the ceiling, and past that you are asking too much of an audience that is scrolling. The constraint isn’t the generator. It’s attention.

Writing the beats out

Don’t describe the timing. Number it. A prompt written as a timeline gets you the pauses; a prompt written as a paragraph gets you a character talking continuously for eleven seconds.

A note on the voice track

Those voiceover timings are for you, not for a text-to-speech engine. The performance comes out of the video generator and the voice is swapped afterward (Chapter 11), so the pauses you write here are what you specify to the animator — not pause markup typed into a speech tool.

A quarter second is a beat. Half a second is a comic pause. A full second is a different joke. Choose the number; don’t discover it.

Two tracks, not one

On anything with dialogue, write the picture and the voice as separate timelines. They are different instruments and they do not share a clock.

The golden prompt

Here is the technique that does more for the finished reel than any setting in this guide, and it costs nothing.

Ask the AI to put on its comedic-timing expertise and apply it to the scene.

Not “write me a prompt.” Not “make this funny.” Hand it the scene and the line you already wrote, and ask it to direct the timing. What comes back is the thing an experienced comedy editor would have told you: the chicken pauses for a beat, looks down at the skillet, looks back up at the man, then delivers the line. You would not have written that. It is the difference between a chicken saying words and a joke landing.

The button

Ask for the timing and you often get told something you didn’t know you needed. Here is a real return on a real scene — Nina and a horse at a bar, “Why the long face?”

The diagnosis in that return is worth more than the beats: “you’re ending on the joke instead of after it. The laugh needs somewhere to land.” That’s the pause-after-the-punchline rule, arrived at from the other direction.

And the fix has a name. The button is the small reaction that comes after the line and actually carries the payoff. Here it’s the horse turning back to his beer and taking another sip — completely unbothered, refusing to dignify the pun. The line is the setup; the reaction is the punchline.

This also explains why the mouth-closed rule in Chapter 3 pays off here: the character who doesn’t speak is doing the comedic work, and you want that decision still open when you reach the animation pass.

On which AI. These are not interchangeable and it is worth being blunt about it. Gemini is superb at image creation and it is not funny — there is a joke in there somewhere. Claude writes in a more human register and is genuinely, unexpectedly good at comedy, especially once you hand it the comedy parameters to work inside. That’s a tested preference from doing this daily, not a recommendation anybody paid for. Nobody here is sponsored by anybody.

Not a ranking. A casting decision.

• Kling — the primary. Best at holding a face through motion and the most reliable at delivering a quoted spoken line for lip-sync. When the shot is your character talking, this is the shot.
• Seedance — B-roll. Image-to-video at 720p, upscaled after. The cutaway, the establishing shot, the hands, the thing that isn't the face.
• Happy Horse — the backup. When Kling is queued, degraded, or fighting you on a composition, this is where the shot goes rather than into a fourth reroll.

What a fully specified animation prompt looks like

Four patterns show up in every good one. They’re visible in the examples below.

1 · Preserve by name, not by instruction

Already covered above: list the things. Faces, clothing, the dog’s markings and vest, the truck, the house, the lighting. A list beats “don’t change anything.”

2 · Say what stays still

This is the one people miss. Every character needs an instruction, including the ones who aren’t doing anything. “His hands stay resting on the desk. His wings stay folded behind him.” “The dog sits facing forward, alert, chest still.” “No other motion.”

3 · Anchor micro-actions to words, not to seconds

Use both: timestamps for structure, word-anchors for the small reactive beats that have to land on something.

4 · Name the ambient layer separately

“Light breeze moves the grass and the distant tree branches. Slight fabric movement in her jacket.” A locked-off shot with nothing moving in the background reads as a photograph with a talking head pasted on. Two clauses of ambient motion is the difference between a still that moves and a shot.

The two-shot

Upscaling is not a quality knob you turn to eleven. It's a fix for a specific problem, and applied to a clip that doesn't have that problem it makes things worse.

• Upscale when: the clip came out of a 720p generator and it's going full-frame; there's fine detail (text, fabric, hands) that's mushy but structurally correct; the clip will be seen on something larger than a phone.
• Don't upscale when: the clip is already clean at delivery resolution; the problem is structural rather than softness — no upscaler fixes a morphed hand, it gives you a sharper morphed hand; the clip is a fast cut nobody will look at for more than a second.

The settings that keep the face

First: does this voice need the pass at all?

Voice Changer is for your heroes. Everyone else ships on the generator’s own voice.

The generator gives every character a voice for free. For most of them, that voice is fine — and the reason is the whole triage rule:

Your hero is the opposite case. Your client knows exactly what they sound like. So does everyone who has ever met them — and those people are a meaningful share of the local audience you are making this for. A near-miss on a familiar voice is worse than an obviously synthetic one, because it lands in the uncanny middle.

This is why the voice pass costs so little time across a week of reels. Most characters never touch it. And it’s the same logic as the consistent character in Chapter 4, one layer down: familiarity is the asset, so it’s only worth protecting where familiarity exists.

There is no straight text-to-speech anywhere in this pipeline. Not for narration, not for anything.

You are not generating a performance. You are replacing a timbre.

This is the chapter most people get backwards, including every tutorial you’ll find. They assume the voice step is text-to-speech: you type the line, the AI reads it, you fight with pause markup until the timing is right.

It isn’t. It’s speech-to-speech — ElevenLabs calls it Voice Changer.

The settings

Making the clone in the first place

• Instant Voice Clone — ten seconds of audio, ready in about two minutes. Genuinely useful for testing whether an idea works before you ask anyone for anything.
• Professional Voice Clone — at least thirty minutes of clean audio. This is what you want for a character you’ll use fifty-two times, and it’s the reason the recording session is worth doing properly.

Recording the source: quiet room, consistent mic distance, natural speech, more variety than you think you need — questions, statements, warmth, flatness. Don’t record someone performing an ad. Record someone talking.

What kills a clone: background hum, inconsistent mic distance, over-processed input, and a source recording where the person is trying to sound like a voiceover artist.

I ran the same script through both. More than once. ElevenLabs is the one that sounds like the actual person — and it wasn’t close.

That’s the finding, and this chapter is short on purpose. You don’t owe a reader your test footage. You owe them the result and the method.

But here is the part that complicates it

So who is the difference for?

Your client. And everybody who has ever met them.

That is not a small group and it is not an incidental one — it is exactly the local audience this whole method is built to reach. The people most likely to hire this business are the people most likely to know what its owner sounds like. A near-miss is invisible to strangers and glaring to the neighbors. It is the same logic as the triage in Chapter 11, one layer down: the accuracy only matters where the familiarity exists, and that’s precisely where it matters most.

The transferable part isn’t about Resemble

By the time you’re in Premiere, the creative decisions are already made. This is carpentry, and speed comes entirely from having built the jigs in advance.

The first move, before anything else

The clip lands on the timeline. Before a single cut, a caption, or a logo — right-click the audio and open it in Adobe Audition.

Doing audio first also inverts the usual order, and deliberately. Most people cut picture and fix sound at the end. Here the clip is finished audio before you start layering anything on top of it, which means every later decision — where the music sits, where the caption lands, how long the end card runs — is made against what the reel will actually sound like.

The stack, in order

The master copy

Don’t start a project. Open the one that already has everything in it.

Every reel begins by opening the same Premiere project and importing the new clips, audio and music into it. Nothing is set up from scratch, because the setup is the slow part and it never changes.

Generating the captions — one word at a time

Single Word. Not single line, not double line.

This is the decision the whole caption look hangs on, and it happens before any styling. Premiere transcribes the audio and can lay captions out three ways. Single Word puts exactly one word on screen at a time, which does three things at once:

• It makes the pop possible. A pop on a five-word line is a wobble. A pop on one word is a beat — and every word gets its own.
• It follows the delivery. Words land as they’re spoken, so the caption is doing the comic timing with the audio rather than lagging a phrase behind it.
• It survives a muted phone. One large word is legible at a glance on a small screen in a way a two-line subtitle never is.

The caption style

Saved as a Track Style, so it applies to every caption in one move and never has to be rebuilt.

The caption pop — every word, one paste

This is the signature move, and it is not a transition. There is no Cmd+D anywhere in it. It’s a two-keyframe scale animation living on Vector Motion, copied off one graphic and pasted onto all of them at once.

Once the donor graphic is saved to Graphics Templates › My Templates › Local Templates, the whole thing is a drag, a copy, a select-all and one keystroke — on a reel with fifteen caption words. That is the hour-you-never-spend-again in its purest form.

Templates over technique

A saved motion-graphics template and a saved transition preset turn a five-minute repetitive task into a drag. The single highest-leverage hour you will ever spend in Premiere is the hour you spend making things you’ll never have to build again.

Cut on the beat, not on the frame. The edit should feel like the music, even when the music is barely there.

The end card

The last frame belongs to the client, not to you.

Why it ends on the logo and not on black

A fade to black says the thing is over. A logo held for three seconds after a laugh says something else entirely:

This is who made you smile today.

That is the whole job of the end card, and it’s why the timing matters. The card doesn’t arrive during the joke and it doesn’t arrive after the viewer has moved on. It lands in the two or three seconds while they are still enjoying it, and it attaches a name to the feeling.

Raw AI audio comes out too quiet to carry a joke. Don’t stop at “it plays.”

Bad audio is the fastest way to look amateur, and it’s the easiest thing in this entire pipeline to fix permanently. The same four steps, in the same order, on every single reel. Once you know the chain you never think about it again.

Where it lives

The chain

Indie Lo-Fi, in full

This preset ships with Audition, and on animated AI voices it does something no gentler setting does — it pulls the body of the voice forward without making it sound processed. If you can’t find it, or you’re on a version that names it differently, these are the values.

Sometimes the sound effect is the joke

A six-shot build, and the punchline is not a line at all — it’s a crunch. The audience has watched the hero take abuse from two talking pretzels for eight seconds, and the payoff they’ve been promised is the sound of one of them being eaten.

Raw out of the generator, that crunch is a low, murmured munch. The joke simply does not happen.

This is the strongest possible argument for running the chain unconditionally. There was no dialogue in that shot to protect — and it is the shot in the whole reel that most needed the boost. You cannot tell from looking at a clip whether its audio is load-bearing. Run the chain on all of it.

Then never do it again

Audition lets you record a sequence of processing steps as a Favorite and apply the whole thing in one click. Mine is called Flo Boost, and it is exactly the four steps above — nothing is finished by hand afterward.

The most damaging mistake in local content isn't a bad joke. It's naming a business that closed.

Check every single time, even for a business you've named before

• Is it still open? Local businesses close quietly.
• Is the name spelled exactly right, including the punctuation they use?
• Is the location right? Chains have multiple sites and you'll tag the wrong one.
• Does the handle you're about to tag actually belong to them?
• Have they changed ownership or name since you last posted?

Never post a real person's misfortune, and in a business like bail bonds, never anything that could be read as referring to an actual case. When in doubt, the answer is no. There's another holiday tomorrow.

Export once, adapt per platform. Same master, different crops, different first three seconds.

• The first frame is the thumbnail on most platforms whether you choose one or not. Make sure frame one is a face, not a fade-up from black.
• Caption everything. Most of your audience is watching muted. If the line only lands in the audio, it doesn’t land.
• Post when your customers are on their phones, not when marketing blogs say to. For a bail bond office that is a very different clock than for a bakery.

The posting order — and why it starts where it does

Write for your most demanding audience first, then loosen. Never the reverse.

The order isn’t about reach. It’s an editing discipline, and it happens to also be the fastest route.

Composing in the most formal room forces the edit. By the time it reads properly for LinkedIn you have fixed the spelling, cut the fluff, and thrown out the one fact too many about Dog Day. That trimmed version is good everywhere.

Try it the other way and you’ll feel the difference immediately. Write loose for TikTok first and moving it to LinkedIn isn’t an edit, it’s a rewrite — you have to add the rigor you never applied. Loosening a tight draft takes a minute. Tightening a loose one takes the afternoon.

Everything downstream is condensed from that one good version. Same joke, same tie-in, fewer words — and the hashtag block swapped per platform, per Chapter 17.

The post copy — eight parts, in order

The caption is not a description of the video. It is the video, continuing.

This is where most people throw away the work. They make something funny and then write “Happy Yoga Awareness Month! 🎧 Check out our video!” underneath it. The post copy is a second act, and it has a shape.

Why part 4 is longer than it looks like it should be

The temptation is a sentence: “Shout out to the local yoga place.” That does nothing for anybody. What works is handing the reader a fact they didn’t know and letting it land:

— a local company makes a carbon fibre yoga mat, developed with a national lab. Sit with that. That lab is a Department of Energy national security site. Seventeen thousand people. And somewhere in that building, somebody looked at a yoga mat and thought, we can do better than foam.

A worked example

The Beer Day reel from Chapter 2, and the post that ran under it. Every part is doing its job:

A third shape: let the audience write the caption

The twelve lines you didn’t pick are not waste. Sometimes they are the post.

Chapter 3 told you to generate twelve lines and pick one, and said the rest become post copy and future posts. This is that rule taken all the way: publish the shortlist and make choosing the joke into the engagement.

The other shape: the fact block

When the occasion doesn’t hand you one great local story, use three small ones instead. Your standing prompt already asks for this — exactly three, one line each, verified, and surprising.

Hashtags are not one tool. Using the same block everywhere is why they stop working.

• Instagram — still functional for discovery. A modest, relevant set beats a wall of thirty.
• TikTok — the caption and the content matter far more. A couple of relevant ones; the algorithm reads the video, not your list.
• Facebook — nearly decorative for reach. One or two. Effort is better spent tagging the actual local business.
• YouTube Shorts — a small number in the description. The title does the real work.
• X / LinkedIn — one or two maximum on X. LinkedIn tolerates a few, and the professional-angle tag outperforms the holiday tag there.

The block, by slot

Six to eight tags, and every one has a job. Fill the slots rather than brainstorming.

Tagging pages is a separate mechanic

On LinkedIn and Facebook especially, tagging a company page is not the same as a hashtag. A hashtag is a filing system. A page tag is a notification — it puts your post in front of a specific business’s account and gives them a one-click reason to share it.

The cost of a reel is not a price. It’s your monthly bill divided by the reels you actually made.

Every tool in this guide sells a monthly subscription. Not one of them sells reels. So the only number that means anything is the division — and that has two consequences most people never work out.

• Your volume moves the number far more than any vendor’s pricing does. Doubling your output halves your cost per reel. A twenty percent price rise barely registers. The person making thirty reels a month pays a fraction of what the person making five pays — on an identical bill.
• Unused capacity is money you already spent. Credits do not roll over. If your plan includes twenty-six thousand and you use twelve thousand, your real cost per reel is roughly double what the price list says. Nobody invoices you for that. It just quietly happens at renewal.

The calculator

Two things it tells you: what a reel really costs, and whether you’re throwing away credits you already paid for.

How to say it out loud

Once you have your number, the honest way to state it has three parts and takes ten seconds.

The comparison that actually lands

A cost per reel means nothing on its own. It means something next to the other things a small business buys in order to get seen:

• One boosted social post, once
• One licensed stock clip — of somebody else’s face, in somebody else’s office
• One hour of a freelance editor
• One month of almost any marketing subscription that produces nothing on its own

Against any of those, the number is the argument. And what you get for it has your face in it, your town in it, and a joke nobody else has made.

Every reel here was made by one person in under an hour, using nothing that isn’t in this guide.

Read the chapters for the method. Watch these for the proof — and for the thing chapters can’t show you, which is how much variety the same forty-minute process produces once the jigs are built.