Jim Tyrrell

An AI short-form video studio. Script generation, teleprompter, recording, transcript-based trimming, auto-captions, auto-placed CTAs inside safe zones across aspect ratios, and publishing that matrix-multiplies a few recordings into many reels. Cuts production from ~1 hour to minutes per finished minute. A two-year backlog item I couldn't build until the models caught up — idea to working prototype in days, not months. Born from 1minutefunnel work that revealed how punishing existing video-ad tooling really is.

Role: Solo designer, developer, tester. Stack: Claude Code, Llama (script generation, captioning), Cloudflare.

Hardest call. The hard part of integration is subtraction. This product replaces 9 specialized tools — ChatGPT, BIGVU, CapCut, Descript, Buffer, and four others — with one workflow. Each of those tools has hundreds of features I deliberately didn't replicate. CapCut's effect library. Buffer's full integration matrix. Descript's audio mixing. Every feature I cut was a feature someone, somewhere, will miss. Every feature I kept earns its place on the workflow's critical path. The thesis comes from Larry Keeley: “In the 21st century, for the first time in the history of our species, innovation is less about the primary invention of the new and more about the elegant integration of the known.” 9 tools, 9 context switches, 3 hours per reel — collapsed into one tool, minutes per reel. The old stack failed the people using it; they never failed it. 1minutereel takes the context-switching away and leads with intelligent defaults, so the user is set up to succeed by default.

Proof of resonance: We put this exact message in an ad — “9 tools. 9 context switches. 3 hours. For ONE reel.” — and 83% of viewers watched to the end. Because it's not a pitch. It's their life.

A conversational assistant for breastfeeding mothers. A sensitive, high-empathy domain handling protected PII. I designed the conversation UI, the data architecture (no PII to the AI engine), and the trust patterns. Migrated production from Llama (Cloudflare) to Claude Haiku (AWS) in a single evening using AI-assisted refactoring — for trust, safety, and unit economics.

Role: Solo designer, developer, tester. Stack: Claude Code, Claude Haiku, AWS.

Hardest call. Llama kept making things up — and I took it as far as I could go. I streamlined the corpus, hardened the prompts, ran code fixes through. It still hallucinated, still ignored my directions to stay inside the corpus. For a domain where a wrong answer could harm a baby, that wasn't a tuning problem — it was structural. I picked Claude Haiku — capable enough to stay inside the corpus, small enough to fit the unit economics of a chatbot bundled with one-time purchases. One evening, AI-assisted refactor. The harder part was admitting the model itself was the limit, not the work I'd done on top of it.

An AI-driven quiz funnel builder. One prompt becomes a fully-routed funnel with branching lead paths. I designed three custom mobile-first players (80% of traffic is mobile), voice-as-peer input alongside text, encrypted PII storage, Facebook ad event logging, and generic webhooks so creators can plug into their existing stack. Thousands of leads gathered per month.

Role: Solo designer, developer, tester. Stack: Claude Code, Llama (funnel generation), Cloudflare.

Hardest call. Restraint as user respect — at every layer. Other quiz funnel tools hand creators a sea of layout choices, trainings, blank-page quiz design — and the creator drowns. I locked the layout: mobile-first, defined, not configurable. The quizzes themselves are AI-generated from a domain expert's prompt, so creators get a killer quiz without needing to design one. They change colors, logos, words — the rest is curated. Voice input was the same call: shipped as a peer to text so the lead can answer hands-free. The trade-off: less expressive control. The win: creators universally report launching in minutes, not days. Reduce the choices that don't matter; multiply the time on the choices that do.

An ad-tech video ad player serving a single ad inventory across 1,000+ partner sites and slot contexts. I designed the player UI, the ad-serving flow, the publisher experience, and the compliance system underneath — refusing to play on unauthorized sites, refusing to play on top of other ads, enforcing a 31-second timing rule, sending telemetry home to catch rogue publishers before they could get the parent ad account banned.

Role: Solo designer, developer, tester. Stack: Claude Code, Cloudflare, JS/HTML/CSS, Microsoft Playwright (test).

Hardest call. Most of CornerCast's behavior is set by the ad ecosystem, not by me. Sound off or you get booted. Skip button defined upstream. Sizes by ad platform. The space for design judgment is narrow. The one place I bent a default: autoplay only when the player is loaded AND visible on screen. Industry standard is autoplay-on-load — it pays better. I gave up some revenue per impression for a better viewer experience and a better signal to the publisher, who has to live with their site's reputation. Designing inside someone else's rulebook means finding the one degree of freedom — and using it.

Claude Code product note: v1 took months. v2 took an afternoon. I rewrote CornerCast with a newer Claude Code model using the v1 spec as the prompt — bugs that had taken weeks to chase by hand got fixed in minutes. Top and bottom slot placements were the DOM headache I never fully diagnosed, but a few days of iteration with Claude Code got the behavior right across every permutation Playwright could throw at it.

An analytics dashboard for a fitness business, built on live Trainerize API pulls. I'm designing the metrics view and the daily operator flow.

Role: Solo designer, developer, tester. Stack: Claude Code, Trainerize API.

Hardest call. A coach with 262 clients can't read 262 rows of raw Trainerize data. The hard part was deciding what not to show. I reduce each client to one triage signal (Progressing, Regressing, No Data) and a plain-language status (Crushing, Slipping, Stuck, Stale), then sort by who's moving furthest, so a ten-second scan tells a coach exactly who needs them today. Trainerize gives you everything; the design job was deciding what earns a column and what becomes noise.