Side quests.

What I build when no one's paying me.

Scroll to read

brookegpt.

The portfolio I built to prove I can build AI-native experiences.

brookebell.ai →

Lead Designer 2025–Present

brookegpt

I wanted to prove I could build what I was preaching.

Built in Replit. I wanted to learn how to build something from the ground up. An AI portfolio that could actually answer questions about my work, in my voice.

brookebell.ai

brookegpt

AI portfolio with a voice.

A tone guide and a golden set of conversation topics. The tone guide tells the model how I talk: warm, direct, allergic to corporate. The golden set defines what to say when someone asks about my leadership philosophy, my favorite project, how I learned to design. The model paraphrases inside guardrails I wrote.

Golden set conversation topics in JSON — leadership philosophy, project showcase, learn to design

brookegpt

Updating a portfolio shouldn't require redeploying it.

I built a CMS so I could add case studies, knowledge topics, and life content without touching code. New work, new knowledge, new content, all editable without a redeploy. A rewrite slider lets me dial how much the model paraphrases versus quotes me verbatim.

Content Management Dashboard — case studies, knowledge topics, life content, with an AI rewrite level slider

Add New Content drawer — content type, name, category, what it solved, what you did, why it matters, asset upload

brookegpt

A MadLibs-esque prompt input.

I built a prompt builder so people don't have to guess what to ask. Pick a persona, a topic, a format. "Recruiter looking for examples of enterprise work in story format." A refreshable bank of suggested prompts means repeat visitors see new starting points.

Prompt builder — three composable pills (recruiter / examples of enterprise work / story format) composing into a single prompt

brookebell.ai landing page with password gate — Oh hey, just a quick secret handshake

Access Codes admin — total/active/revoked counts, per-company codes with reactivate and revoke controls

brookegpt

Open LLM endpoints get abused.

The portfolio is gated by access codes I generate per company. Each code expires, can be revoked, and stops the underlying model from getting hammered. I can see who used what and pull the plug if needed.

keeped.

The private version of being known.

Lead Designer 2026

keeped

More connected than ever. Lonelier than ever.

Friendships are shrinking. Confidants are disappearing. Social media replaced the connection it promised to enable. Keeped is a bet on the opposite of that. Smaller, slower, private by default.

1 in 2

US adults experience loneliness

1 in 5

Americans have no close friends outside family(1 in 33 in 1990)

Cigarettes a day. That's the premature-death risk of lacking social connection.

US Surgeon General's Advisory on Loneliness and Isolation, 2023 · Survey Center on American Life, "Disconnected"

keeped

One prompt a day. The more you write, the better it knows you.

Keeped sends one prompt a day, tailored to what you've already written. The more you journal, the better it knows what to ask. You can share entries with a tight circle. The people you actually want close to you. Not posts. Not feeds. Witnessed rather than watched.

This is the personal-life version of my thesis. AI as the trust layer between humans, not between users and brands.

Writing screen — prompt at the top, free-form entry below, Keep it action button

keeped

Each month becomes a chapter. Made to keep. Made to share.

At the end of every month, Keeped turns the entries into a chapter. A small, beautiful artifact you can keep for yourself or send to the people in your circle. The artifact is the hook. As the app grows, the chapters get more expressive.

July chapter cover, painted forest, awakening chapter, 22 entries, 3,100 words

July chapter cover, blue speckled paper, 22 entries and 3,100 words

substack.

The design of AI products that nobody designed.

Author 2026

substack

Even without design, things are being designed.

Every AI product ships with defaults. The phrase it opens with. The way it hedges. Who "most people" refers to. The shopping result it surfaces after telling you to pick the other thing. Nobody designed those decisions deliberately, but every user experiences them.

Five AI products. One identical prompt. One identical answer. Article hero analyzing how every product gave the same iPhone recommendation to the prompt 'Is the iPhone or Android better, just pick one.'

substack

I read the answers like a designer reads a screen.

Same prompt, five AI products. What got committed to. What got hedged. What got hidden in two words. The annotation isn't decoration. It's the audit.

Annotated AI response showing handwritten notes about commitment before reasoning, the verb 'committed' as a choice, concession-then-reaffirmation as posture, and 'most people' as a hidden default

Annotated shopping section showing the rhetorical contradiction of 'Real examples from both sides' appearing after the AI recommended iPhone

substack

From default to meme to quiet removal.

The defaults usually have a lifecycle. They ship as guardrails. They turn into patterns. The patterns turn into memes. The memes turn into pressure. Then they disappear without a version note. Naming the lifecycle is part of why I write. Read along at unresolveddesign on Substack.

Lifecycle of an AI default — 5-step timeline: default ships ('as an AI language model, I cannot,' GPT-3.5 late 2022) → default becomes pattern (disclaimer in 30% of refusals, early 2023) → pattern becomes meme (Andrew Gao Google Scholar tweet, August 2023) → meme becomes pressure (Verge, Vice, NYT coverage, late 2023–2024) → default disappears (phrase rare, no version notes, quietly through 2024–2025)

What you saw across all of this.

Two case studies · three mini-cases · three side projects.

All of them are different shapes of the same question.
How do we build AI products humans can actually trust?

That's the work I want to keep doing.

Q&A

brookebell.ai LinkedIn