How I AI: Bryce Rattner Keithley's No-Code Playbook for Building a Fitness App with Replit, Gemini, and Claude

Bryce Rattner Keithley shipped an iPhone fitness app before most of her more technical friends. In this episode of How I AI, she walks through how she built Daily Hundred in Replit, generated custom exercise videos with AI animals, and pushed the app through Apple review despite having no traditional engineering background. See Create Custom AI-Generated Animated Workout Videos with Gemini and Higgsfield.

Daily Hundred gives users one exercise each day and asks them to complete 100 reps. Instead of stock fitness demos, the app features AI-generated animals performing the movements, including leopards doing crunches, turtles doing lunges, and polar bears working through mobility drills.

Bryce used Replit to build the product, Gemini plus Higgsfield to create the exercise animations, and Claude with Claude Code to handle the intimidating parts of hosting and App Store submission. None of it was fully automatic. The app took months of iteration, dozens of failed generations, and one intensely focused 25 to 30 hour weekend to get over the finish line. See Build a Minimum Viable Product App with Replit Using No-Code 'Vibe Coding'.

Building a functional MVP in Replit without coding experience

The idea for Daily Hundred started during the pandemic. Bryce was spending long days at home with two young kids and wanted a lightweight way to stay active between meetings. She briefly tried a daily pushup challenge, but doing the same movement every day got repetitive fast.

What she wanted was simple: an app that would tell her a different movement each day and let her log 100 reps. Instead of over-planning the product, she started with the smallest possible description of the outcome and relied heavily on the live preview to decide what to change next.

Starting with one sentence instead of a spec

Bryce opened both Lovable and Replit on the same day and gave them essentially the same starting instruction:

"Hey, build me a tool called Daily Hundred that pushes a different exercise to me to log a hundred reps."

Both tools generated rough MVPs surprisingly quickly. Bryce stayed with Replit partly because she knew someone there who could help if she got stuck, but the actual product work she demonstrated was still done through the agent workflow rather than traditional coding.

Why plan mode mattered more than code generation

One of Bryce’s biggest lessons was learning not to jump directly into implementation prompts. Early on, she would ask Replit to change something simple, like converting a progress bar from a line into a circle, and the app would come back in a completely broken or bizarre state.

Replit’s plan mode changed the workflow because it let her discuss the approach before applying changes. Bryce consistently framed herself as a beginner and asked the agent to explain what it was about to do. Combined with the live preview panel, that made it easier to spot bad directions early, undo mistakes, and refine requests incrementally instead of blindly accepting every generated change. One of the recurring themes throughout this episode of How I AI is that her lack of technical confidence actually made her more willing to ask basic questions, request screenshots, and demand more literal instructions.

Generating exercise videos with Gemini and Higgsfield

Once people started using the app, Bryce ran into a practical problem: users kept asking how to perform unfamiliar movements like Supermans or reverse lunges. She did not want the production app filled with videos of herself exercising in her living room, so she started experimenting with AI-generated animal characters instead.

The final workflow combines two separate ingredients: a carefully generated still image that defines the character and pose, and an iPhone video of Bryce performing the actual movement.

Creating the starting character image

Bryce generates the characters in Gemini. During the walkthrough, she and I build a leopard prepared for bicycle crunches. The prompt reads less like creative writing and more like a barre instructor giving positioning cues: hands behind the head, elbows out, knees above hips, feet elevated in tabletop position, head oriented to the left, feet to the right.

That level of precision turned out to matter. Bryce repeatedly emphasizes that the starting pose has to closely match the motion video later in the workflow, otherwise the animation models struggle to interpret the movement correctly.

Create an anthropomorphic leopard in a gym setting, wearing exercise gear. The leopard's hands should be behind its head, elbows out to the side, head resting down on the mat. The head should be to the left and the feet should be to the right. The knees should be positioned above hips and feet forward in a tabletop position. Both feet off the ground. There should be no other characters in the image.

The first generated leopard missed several requirements. The hands were wrong, one leg was misplaced, and the orientation did not fully match the exercise reference. Instead of endlessly patching the original prompt, Bryce restarted with a cleaner rewrite and added more literal constraints, including details like both feet off the ground and no extra characters in the scene. She found that restarting often worked better than stacking edits onto a flawed generation.

Recording the movement reference

After generating the still image, Bryce records herself doing the movement on her iPhone. Matching the orientation of the generated character matters a lot. If the leopard’s head faces left in the image, Bryce mirrors that positioning in the reference video so the transfer behaves more predictably.

Turning a still image into an animated workout clip

Inside Higgsfield, Bryce uses Motion Control with the Kling model to merge the character image with the exercise video. She tested several model combinations and found that Kling inside Higgsfield consistently produced better motion than using Kling separately.

Her workflow inside Higgsfield is straightforward but surprisingly sensitive to setup details:

1. 1. Upload the trimmed exercise video recorded on the phone.
2. 2. Upload the Gemini-generated character image that matches the same starting pose.
3. 3. Set Scene Control to Image so the gym environment follows the still image instead of inheriting Bryce’s real living room background.
4. 4. Generate the clip, wait several minutes, and inspect the output carefully for broken motion or anatomical weirdness.

The resulting leopard crunch video looked remarkably polished. The generated gym scene even included reflections in the mirror behind the character. But the process was still messy. Bryce showed several failed examples, including a floating genie exercise demo with distorted limbs and impossible body positioning. Small mismatches in starting posture could completely derail the animation. Over time, she learned that the more literal and physically accurate the starting image was, the better the downstream motion transfer became.

Getting the app through App Store submission

The final hurdle was turning the Replit web app into a real iPhone app and surviving Apple review. When Bryce first started the project, most people around her assumed this phase would eventually require hiring a contractor or technical cofounder. By the time newer AI coding tools arrived, she decided to see how far she could push it herself.

Using Claude as planner, engineer, and reviewer

Bryce began with an intentionally simple prompt to Claude that included an honest admission about her experience level:

"How do I prepare a Replit app for app store submission? I am not technical."

From there, she split the work across different tools depending on the type of task she needed help with:

1. Claude for planning: Bryce used the regular chat interface as a technical guide that explained the overall submission process, clarified unfamiliar concepts, and broke the work into manageable stages. You can follow the full implementation in Navigate the App Store Submission Process with Claude as a Technical Co-pilot. See Navigate the App Store Submission Process with Claude as a Technical Co-pilot.
2. Claude Code for implementation: Once she understood a bounded task, she handed the coding work to Claude Code instead of trying to edit files manually.
3. Claude for review: After Claude Code generated changes, Bryce brought the outputs back into the planning conversation and asked whether the result actually matched the original objective.
4. Terminal for execution: Some steps still required manual action. Bryce copied reviewed commands into Terminal herself and used the process to move the application from Replit onto Railway for hosting, even while joking that she did not fully understand what Railway actually did.

Fixing the first App Store rejection

Apple rejected the first submission. Instead of panicking, Bryce copied the rejection notes directly into Claude and worked through the issues one at a time.

• Age rating: One rejection item came down to a configuration mistake in Apple’s settings flow.
• Sign in with Apple: Bryce had technically implemented the feature already, but had never properly tested it before submission, so she had to troubleshoot the integration.
• Account deletion: Apple required a visible user-facing way to delete an account, so Bryce added the missing functionality.

Bryce said the final submission push consumed roughly 25 to 30 focused hours over a single weekend. The second submission passed review, and Daily Hundred went live in the App Store.

What is actually worth copying from this workflow

The impressive part of this build is not that AI removed all technical work. It clearly did not. Bryce still made product decisions, rewrote prompts, reviewed previews, recorded motion references, debugged failed outputs, moved hosting infrastructure, ran terminal commands, tested authentication flows, and handled App Store compliance issues herself.

What does seem highly transferable is the operating style. Bryce consistently treated AI tools like collaborative systems that needed supervision, clarification, and verification rather than magic automation. Her most useful habits were simple: ask for plans before implementation, stay extremely literal in prompts, use screenshots when words fail, restart broken generations instead of endlessly patching them, and admit uncertainty early. This workflow works especially well for motivated domain experts who know the product they want but do not yet know the technical path. The AI tools lowered the cost of testing the idea. Human judgment was still required everywhere the output became ambiguous, unrealistic, or user-facing.