Quick disclaimer because the title is doing some work.
The "weekend" part is true. I built Birdie ? her face, her voice, her kitchen, her whole vibe ? in two long Saturday-Sunday sessions in early March. By Sunday night she had ten 15-second clips ready to post.
The "$4,800 last month" part is also true. May was her best month. April was $2,100. March was zero. Building her in a weekend was the easy part. Getting Instagram to care about her took three months of posting through a flat curve that felt humiliating to me but apparently is normal.
So if you're reading this hoping "two-day build, four-figure month, easy money," ? sorry, that's not the timeline. The timeline is two-day build, then 90 days of grinding through three followers a day, then a single reel that broke 200k views and started the snowball. Everyone I've talked to who's done this says the same thing. The grind comes first. The money comes after the grind has been quietly going for longer than you'd expect.
OK, with that out of the way. Here's the actual workflow.
One thing before we get into it. Real talk. The single most useful thing I did in month one was stop watching YouTube tutorials. I tried for two weeks. Got nothing usable. Every tutorial assumed I already knew the tool, or assumed I knew nothing, never the middle. A structured course cut through the chaos so cleanly it is almost embarrassing how much faster I learned. YouTube has value once you know what you are looking for. It is a maze when you do not.
Who Birdie is, and why she isn't me
Birdie is a fictional 60-something woman with white hair, mischievous bright eyes, and a 70s-style kitchen with wood paneling and warm lamp lighting. She gives "things your grandmother would tell you about money if she had lived through the 70s recession." Practical, no-nonsense, slightly bossy. Tells you to keep your egg cartons. Knows what a coupon-clipping ring is.
She is also not real. She has no body. She has never lived through the 70s. Every clip I post of her is generated in Higgsfield, voiced in ElevenLabs, and stitched in CapCut.
I built her instead of just being on camera because (1) I am 34 and trying to give "grandmother wisdom" content would have read as ironic instead of comforting, (2) I work full-time as a tech recruiter and I do not want my coworkers finding my Instagram, and (3) honestly, I tried being on camera for two months on a different account in 2024 and I hated every second of it.
The AI character solved all three at once. I am the writer and the director. Birdie is the talent. We're a two-person operation where one of us doesn't have to brush her hair.
The 7-tool stack
Everyone overcomplicates this. The actual list is short:
1. Higgsfield ? generates Birdie's image and animates her into 15-second video clips 2. Nano Banana Pro (inside Higgsfield) ? the specific image model that gets the skin texture and lighting that doesn't read as AI 3. ElevenLabs ? Birdie's voice (a warm older-woman voice I picked from their library and locked in on day one) 4. Claude ? writes Birdie's scripts (more on this below) 5. CapCut ? free editor where I stitch the clips, add captions, and master the audio 6. Instagram ? the only platform she's on, by design 7. A spreadsheet ? tracking what content performs, so I can stop guessing
That's it. No design software. No animation tool. No expensive subscriptions stacked on top of each other. The two paid tools are Higgsfield and ElevenLabs and I'd be honest, both add up to a real monthly cost ? not "the AI gurus told me free, lol no" but a real expense. Verify their pricing yourself before you commit, both have plans that scale with how much you generate.
How I built Birdie's face (this took most of Saturday)
The trick I learned from a YouTube tutorial is to think about the prompt in four layers, not one big sentence. Subject. Environment + lighting. Style. Camera. Each layer answers a different question the model needs.
This is the prompt I ended up with after about fifteen attempts:
> *Subject: a 64-year-old woman, warm expressive face, white hair styled neatly in a low bun, mischievous bright eyes, slight smile, looks like she's about to tell you something she shouldn't.* > > *Environment + lighting: a 1970s American kitchen, oak wood paneling, copper pots on the wall, a kettle on the stove, warm amber lamp light from the right, sun coming in low through a small window on the left.* > > *Style: hyperrealistic 4K photograph, natural skin texture with age detail, no plastic smoothing, soft film grain, professional lifestyle photography quality.* > > *Camera: Sony A7 IV, eye-level slight close-up, shallow depth of field, 9:16 vertical aspect ratio.*
Generate at batch size 4, pick the strongest one. Then ? this is the part most people skip ? generate a "model sheet" of side profiles using the same prompt. That gives Higgsfield a reference of what Birdie looks like from every angle. After that, every new clip I generate of her looks like her, not a different 64-year-old woman in the same kitchen.
If you want the next level, you can paste your prompt into Claude and ask it for a JSON-formatted version of the same prompt. The output is more detailed and Higgsfield reads it better. I do this for every new "scene" I introduce.
The voice (this took an hour, max)
I went to ElevenLabs, opened the voice library, filtered for "older female, warm, American English," and listened to maybe twenty voices until one landed. Saved it as Birdie's voice. From that point forward, every script I write gets read in that exact voice, which means Birdie sounds like the same person in every post.
The voice becomes a part of the brand. Same voice every video. People associate it with Birdie the way people associate the Geico gecko with that specific accent. It's not optional, it's the entire point.
The scripts (Claude, not ChatGPT)
I tried both. Claude writes better narrative copy. ChatGPT writes copy that sounds like ChatGPT. I don't know why this is true, but I've stopped fighting it.
The way I prompt: I describe Birdie's voice in two sentences, give Claude one specific topic ("today she's telling young women why she keeps her egg cartons"), and ask for a 45-second script in her voice, broken into 6 short beats. I read it out loud once. If anything sounds like Claude trying to be a grandma instead of being one, I edit that line. Usually only 2-3 lines per script.
I learned the structured way to think about all of this from a short course I did on Mindwand. Around 89 bucks at the time. The reason I bring it up isn't a sales pitch ? it's that the format kept me showing up daily, which mattered more than the brand. Coursera and Skool communities have similar things, take your pick. The point is to pick one structured course and finish it before you start, instead of binge-watching YouTube for three weeks and never building anything. Format matters more than brand.
Animating Birdie (this is where I burn the most credits)
Back into Higgsfield, but now into the video section. Higgsfield gives you three video models: Kling 3.0, Veo 3.1, and Seedance 2.0. The first two animate from a still image and a motion prompt. Seedance 2.0 is the wild one ? you upload Birdie's image AND her ElevenLabs voiceover, and it generates her actually speaking the lines with lip sync and hand gestures.
For Birdie, I use Seedance 2.0. She's a character who talks to camera, so the lip sync is non-negotiable. Each clip maxes at 15 seconds. Each clip takes 2-3 minutes to generate. Each clip costs credits, which is why I plan my scripts in 15-second beats and try to nail the prompt first try.
I'll be honest: my first-try hit rate is maybe 60%. The other 40% of the time the lip sync drifts or her hands do something weird. I regenerate. Sometimes twice. The Polaroid principle ? don't ship the first AI output ? has saved me from posting some really uncanny clips.
CapCut, the easy part
Drag all the clips onto the timeline in order. The voiceover Seedance generated is already baked into each clip, so I don't have to add a separate audio track. Add CapCut auto-captions. Pick a clean sans-serif font with a light glow. Drop a quiet music bed at 15% volume under the voice. Export.
No fancy transitions. Hard cuts at the seams. The Instagram algorithm punishes Shorts/Reels with flashy transitions ? viewers swipe away during the swoosh, and the algorithm reads that as a drop-off.
The Instagram-algorithm reality that nobody tells you
This is the part I wish someone had explained to me on day one.
Instagram's algorithm doesn't care about your likes. It cares about three things:
- Watch time: average post gets 3-5 seconds. Viral content gets 36+ seconds. If your viewer doesn't make it past second 5, you're dead.
- Saves: average post gets 1 save per 30 viewers. Viral content gets 1 save per 10. Saves tell Instagram "this is reference material."
- Shares: average post gets 2% of viewers sharing. Viral content gets 5-10%. Shares are the king metric. Shares signal "this made someone feel something."
Likes are vanity. Saves and shares are the actual currency.
Before I publish any Birdie reel, I run a three-question checklist:
1. Will someone watch this all the way through? 2. Will someone save this for later? 3. Will someone send this to a friend?
If yes to all three: post. If no to any: rewrite. I've killed about 30% of the clips I generated using this filter. The ones that survive perform better than anything I shipped in my first month before I had the filter.
The honest growth math
I posted once a day, every day, for the first 87 days. Most of those posts got under 1,000 views.
Then a reel went a little viral. 240,000 views. 1,800 new followers in 48 hours. The algorithm started serving Birdie's older posts to those new followers, and a few of them also took off in delayed bursts. By the end of week 14 the account had crossed 30,000 followers.
The pattern people don't tell you about: most of the growth happens during a flatline that looks like nothing is happening. The algorithm is quietly figuring out who your audience is. You just have to keep showing up while it figures it out.
I added the $27 ebook (a frugal-finance guide written in Birdie's voice, written together by me and Claude over the course of a week) once the account crossed 15,000 followers. The ebook does about 175 sales a month right now. $4,800 in May. Not 7 figures. Not life-changing. Real money for someone who works a day job and posts on Birdie's account at night.
What I'd tell someone starting
Build the character on a weekend. Don't overthink it.
Then commit to 90 days of daily posting before you judge whether it's working. Three months is the honest minimum, and most people quit at three weeks because the numbers look small. Don't be most people. The small numbers ARE the path. They just don't feel like one until you've been on it long enough to look back.
Pick one paid course or community and finish it before you start. I picked Mindwand for the daily-lesson format. Could have picked Skool or Coursera. Doesn't matter which, matters that you finish one.
And ? this part is the most important ? don't fall for the AI gurus on YouTube who tell you it's 10 minutes from prompt to viral. The 10 minutes is real, that part isn't a lie. The viral part takes 90 days of unglamorous posting before it happens, and most of those posts do under 1,000 views.
That's the deal. Take it or leave it. Birdie's worked out for me. She might not work out for you. But you'll never find out if you don't commit to the 90 days.