An executive coach rebuilt her daily-word app as a white-label template in two

Capstone Final Project Category The Ultimate AI Master Description Extraordinary AI excellence across various AI tools, domains, and complexity levels. Date Submitted Aug 9, 2026 Submitted By Dr. LJ (Linda Johnson) ljohnsondba@gmail.com PROJECT DESCRIPTION  This concept has personal roots that predate the assignment. I've never had patience for idle chat, conversation only feels alive to me when it's carrying real purpose. That bias shows up in who I admire: Peter Diamandis and his moonshot language, first heard live at a Diamandis/XPRIZE event years ago and still energizing today, Musk carrying forward a thread that started with Tesla, a misfit who saw further than the room he was standing in. It feels fitting that Vishen himself has spent real time in Diamandis's orbit through Abundance360, the same "big thinking" throughline runs from that room to this assignment. Words Build Worlds is my attempt to hand other people a small piece of that energy, one word, one moment, enough to nudge a bigger idea into motion ARTIFACT 1 is a short Higgsfield/Seedance-generated VIDEO: a three-shot arc of a man and a humpback whale communicating underwater (color-coded sound-wave rings), the surface world's environmental problems visible above the waterline, then both breaking the surface, the whale submerging toward the sun and leaving the man alone with a concerned, dawning-awareness expression. Stitched in Canva with transitions, a closing text overlay reading "What words are you using to build a world?", and a soft, low-volume ambient music bed. No trademark placement, no voiceover, and no text/logo baked into the AI-generated footage, all text and branding added in post. ARTIFACT 2 demonstrates rapid repackaging of an existing AI-built APP into a white-label template, using Claude Code as the build agent. Rather than build a new product from scratch, the existing Words Build Worlds (WBW) app, originally built July 17, 2026 through a prd -> build -> design -> product-design skill sequence, was duplicated and reskinned in a single working session to prove the underlying architecture is portable to any brand. Words Build Worlds is a free, no-login web tool that hands someone a single word to focus on "building their world" for the day, drawn from a coaching-based domain model (Your World / Work World / Others' World). The original version is skinned in LJ Executive Coaching's brand system. This new version, WORDS BUILD WORLDS - WHITE LABEL EDITION, keeps the identical mechanic and content and replaces only the brand-specific surface (headers, footer, credit lines, cover imagery, and accent colors) with clearly-marked placeholder text, demonstrating the product as a resalable template rather than a single-brand asset. ATTACHMENTS  ARTIFACT 1 = 8 • 1-Man and Humpback Whale 16-9-DrLJ.png • 2-Man-Whale-cloudy16-9-Dr.LJ.png • 3-Head Above Waer 16-9-DrLJ.png • Canva Video Production ScreenShot-DrLJ 5.jpg ARTIFACT 2  = 4 • Earlier HW Vers-WordsBW-DrLJ.jpg • MV WordBW App Cover-Dr.LJ.png • Whale-Birds-Bkgrnd-WbW-app-Dr.LJ.png • WordsBW-Whitelabel-DrLJ.jpg ACHIEVEMENTS  ARTIFACT 1 - VIDEO • Verified Higgsfield/Seedance/Kling tool capabilities live (balance check, cost preflights) before generating anything, rather than assuming functionality or relying on remembered documentation • Diagnosed and corrected two distinct technical failure modes mid-build: an aspect-ratio mismatch causing white borders in generated output, and prompt language ("school of fish") producing unintended crowded motion instead of the calm, ambient scene intended • Used custom-cropped storyboard frames as visual start-frame references for each of three separate generations, anchoring whale scale, character consistency, and composition rather than relying on text-only prompting • Ran a controlled experiment combining all three reference images into a single story-arc generation using Seedance's asset-reference (@) syntax, evaluated the result against the three-separate-shot approach, and confirmed the separated-shot method preserved character and story consistency more reliably • Recognized and kept an unplanned, higher-value emergent result (the whale breaching toward the sun while the man's expression shifts to open concern) that matched the intended emotional target more precisely than the originally scripted ending • Iterated the closing line through five candidate phrasings, testing each against brand terminology and tone before locking "What words are you using to build a world?" • Stitched three independently generated clips into one continuous sequence in Canva, with transitions, layered on-screen text, and logo placement sequenced around deliberate eye-flow across the full runtime • Selected and justified custom typography for the closing text (Didot LP Display/ Glacial Indifference) after identifying the brand's primary display font was mismatched in tone to a reflective, contemplative closing moment • Hosted and verified the final video via a shareable link, confirming access in an incognito session before submission ARTIFACT 2 – WHITE-LABEL • Ran a live topic-selection interview start to finish, landing on a locked topic and Hook in one conversation • Re-located and resumed an existing Claude Code project session (WBW) inside VS Code from 22 days prior, confirming project identity via package.json/file structure and original conversation history rather than rebuilding from memory • Duplicated the working app inside Claude Code without disturbing the live original, preserving Words Build Worlds V1 at its existing deployment untouched • Repackaged the app's brand surface only (eyebrow, hero title/subhead colors, credit line, footer, background imagery, accent color system) into a white-label template with placeholder copy in every brand-specific slot, while preserving the full working mechanic and all word/sentence content unchanged • Created a new, separate GitHub repository and Vercel deployment for the white-label edition, giving it its own live URL fully independent from the original • Made a real-time scope decision to keep V1's existing word/gap/sentence content and question-free tap mechanic exactly as built, rather than adding a new question layer, to protect the deadline • This was a live test to evaluate the potential of this process a time-saving capability. this is also the second time the user worked through the process and, even a few hiccups, the build was done within two hours . Given version one was an eight-plus-hour endeavor, the white label test certainly did provide a significant savings timewise. The ROI will only increase as the process it is fine-tuned and the user gains more experience with the tools and the process. And eventually the goal is to automate this process through the use of AI.  TOOLS ARTIFACT 1 - VIDEO • Higgsfield (Seedance model, Kling 3.0 model) - video generation • Higgsfield balance/credit and cost-preflight checks • ChatGPT - original storyboard reference imagery • Canva (Pro) - image cropping/export, video stitching, transitions, text overlay, logo placement, font selection, ambient music, publish/share link • Claude.ai - prompt drafting and iteration, concept development, technical troubleshooting (aspect ratio, prompt-language diagnosis), image cropping via code execution • MindValley AI Mastery 2026 lessons (context/framing for the assignment itself) ARTIFACT 2 – WHITE-LABEL • Claude.ai - live topic-selection interview, planning • Claude Code (VS Code extension) - locating and resuming the original 22-day-old project session, duplicating the app, repackaging the brand surface • Canva (Pro) - cover sketch and reference imagery for the new brand surface • GitHub - new repo for the white-label edition • Vercel - deployment, giving the white-label edition its own independent URL • MindValley AI Mastery 2026 lessons METHODOLOGY ARTIFACT 1 - VIDEO METHODOLOGY STAGE ONE - Tool Verification and Concept Anchoring Before generating anything, confirmed Higgsfield account status, credit balance, and available models live rather than assuming capability from memory or documentation. Existing three-frame storyboard concept (previously produced via ChatGPT) was reviewed and adopted as the visual foundation, then cropped into three individual reference images, one per intended shot, so each generation could be anchored to a specific composition rather than generated from text description alone. STAGE TWO - Reference-Anchored Shot Generation (iterative) Each of the three shots was generated as an independent, single-image-referenced clip rather than attempting one combined generation. Prompts were refined iteratively based on direct visual review of each output against intent - correcting an aspect-ratio mismatch that was producing literal white borders in the rendered video, and correcting prompt language that was producing unintended crowded fish movement instead of the calm ambient scene intended. A controlled comparison test was run using Seedance's multi-image asset-reference syntax to combine all three reference images into a single continuous generation; the result was evaluated against the three-separate-generation approach and found to sacrifice character and scene consistency, confirming the separated method as the right choice going forward. STAGE THREE - Assembly and Post-Production The three finished clips were assembled in Canva: transitions applied between shots, on-screen text and LJ logo added (deliberately not baked into the AI generations themselves), custom typography selected for the closing line after identifying the brand's primary display font was tonally mismatched to a quiet, reflective closing moment, and a low-volume ambient music bed layered underneath. Final export was tested in an incognito browser session to confirm the shareable link worked for an outside viewer before submission. ARTIFACT 2 - WHITE-LABEL METHODOLOGY • COS (Claude) and Dr. LJ jointly decided to repurpose the existing WBW app rather than build a new one, framing the exercise around ease of white-label repackaging as the core AI-mastery demonstration • Dr. LJ produced a rough Canva sketch of the desired V2 cover (hero title, subhead, placeholder company line, ocean/whale background, specific hex colors) rather than describing it verbally, which sped up alignment significantly • COS and Dr. LJ located the original Claude Code session by working backward through recent VS Code sessions and cross-checking against known project details (original working title "Be Rich and Happy," coaching-specific language) until the correct project was confirmed • COS drafted a structured, section-by-section change instruction (what stays unchanged vs. what gets replaced) before any build work began, to keep Claude Code's edits targeted rather than triggering a full redesign • Duplicate/branch created first, as a deliberate safeguard, before any changes were made, so the original deployment could not be affected • Image sourcing handled by direct upload of Dr. LJ's own clean (text-free) photo rather than in-session AI image generation, to control time and outcome • New GitHub repo and Vercel deployment created to give the white-label edition its own independent, shareable URL LEARNINGS ARTIFACT 1 - VIDEO LEARNINGS • Vague motion language in prompts produces flat, static results even when the underlying visual reference is strong - "calm and still" rendered as near-total stillness, while more specific direction (a described head tilt, a described leaning-in motion) gave the model something concrete to actually animate. • Small wording choices carry more weight than expected - "school of fish" produced coordinated group-swimming behavior, not the calm ambient presence intended; "a few tiny fish drifting leisurely" produced the correct result. Precision in word choice matters as much in video prompting as in written brand voice work. • A reference image's actual pixel dimensions, not its appearance on screen while editing, determine how a video model interprets it - an image that looked correctly filled in Canva's editing view still exported at the wrong aspect ratio, and the resulting white borders were read by the model as literal content to preserve, not as an artifact to ignore. • Video generation is non-deterministic even when reusing the same start image and a closely related prompt - a redo intended only to add mouth movement to an otherwise-successful shot produced a whale that read as noticeably less characterful than the original take, confirming that small prompt changes can have outsized, unpredictable effects on secondary details. • Combining multiple reference images into a single continuous generation (rather than generating each beat separately) is tempting for efficiency but measurably weaker for character and scene consistency - this was tested directly rather than assumed, and the three-separate-clips method was confirmed superior for this project's needs. • Unplanned model output sometimes lands closer to the intended emotional target than the original script - an unscripted final beat (the whale breaching toward the sun while the man's expression shifts to visible concern) was recognized as more precise than the originally written ending and was kept rather than re-generated to match the original plan. • Model cost varies meaningfully by choice, not just by duration - Kling 3.0 ran roughly three times the per-second cost of Seedance in this project, a real factor in budgeting multiple takes across a multi-shot build. • Text and logos are a known weak point for AI video generation - legible, brand-correct typography and clean logo placement were more reliably achieved by generating footage with no text or branding at all, then adding both in a separate editing pass, rather than asking the generation model to render them directly. • A brand's primary display typography is not automatically right for every context - the brand book's bold, architectural display font was well-suited to headers and hero text but tonally wrong for a quiet, reflective closing question, requiring a deliberate, documented departure for this one use case. ARTIFACT 2 – WHITE-LABEL LEARNINGS • A working prototype can be meaningfully repackaged for a new audience/brand in under two hours when the underlying mechanic is left untouched and only the brand surface is changed - this was the core AI-mastery point being demonstrated • Providing a rough visual sketch (even in Canva, not a design tool) resolved design ambiguity far faster than continued back-and-forth text description • Re-locating a prior AI coding session required recognizing environment-specific landmarks (VS Code vs. terminal vs. extension panel, project names, timestamps) rather than assuming any single "chat" held full history - a real gap in personal process worth tightening for next time • Deliberately duplicating/branching before making any changes eliminated risk to the working original and should be a standing first step for any repackaging or reskinning task • Naming discipline in the moment (renaming the duplicate immediately and clearly) prevented confusion later when multiple similarly-named projects existed side by side • Scope discipline under time pressure - explicitly declining to add a new question-based mechanic despite it being a tempting enhancement - protected the deadline without compromising the core demonstration RELEVANT URLS   https://drljcoaching.com ARTIFACT 1 - VIDEO LEARNINGS Artifact 1 (video, Canva share link): https://canva.link/urn4c6uz9lvq96y ARTIFACT 2 - WHITE-LABEL Artifact 2 - Original (V1): https://words-make-worlds.vercel.app/ Artifact 2 - White Label (V2): https://words-bw-whitelabel.vercel.app/  (verified working in incognito) Artifact 2 - GitHub (White Label): https://github.com/ljsubtools-git/WordsBW-whitelabel.git