Nikema Prophet — Walkthroughs & demos

Watch me walk through it.

Short videos where I explain what I built, why, and how it works. Most link back to a project you can try.

YouTube channel ↗

OCT 2026

  • Walkthrough
  • 3:39

A Human Approval Gate for AI-Generated History Fact Cards

An agent generates fact cards in Sanity and I'm the approval gate in the middle. Generating a card, checking it against Exa source excerpts, rejecting it with feedback, then approving and publishing.

Transcript

this is California Black Stories My Sanity Challenge Path Two Project I run a Facebook page about Black history in my home state, and I built this app to prepare fact cards with their sources and review decisions attached. The public library lets me find posts, read their captions and citations, and download the card image. It includes drafts as well as published posts.

I start with broad topics and ask for five specific claim ideas. These are candidates, not verified facts. I can choose an idea or enter a topic of my own. Verification happens before I create a card. Exa finds source excerpts, and the fact checker breaks the claim into checkable parts.

Each part is labeled supported, unsupported, or contradicted. Supported means an excerpt backs it. Unsupported means supplied excerpts don't address it. Contradicted means an excerpt conflicts with it. The linked Exa sources stay visible with excerpt previews and publication dates when available, so I can inspect the evidence myself.

If part of the claim isn't supported or conflicts with the sources, The app suggests a correction while keeping the supported details. Retry Source checking is available when I want more evidence that adds sources while retaining the earlier excerpts. If I edit a suggested correction, the edited wording needs verification. Applying this suggestion uses the findings already on screen.

It doesn't run another search or audit. I can inspect the corrected fact before creating the card, and it still needs my editorial approval. Now I create the card and submit it for review. The image is a template rendered 1080 by 1080 PNG, not AI generated artwork. Sanity stores the fact, caption, citation, source URL, template id, and uploaded image as structured fields.

The workflow moves the completed post to in review. Let's open this post in studio. Let's look at the Facebook caption. The question at the end is have you ever visited this iconic site? Publishing is gated by the workflow's actual approval. If I reject a post I leave feedback and a background Sanity function handles the revision.

Here I'm changing only the engagement question so it keeps the fact, source, and card image. The workflow shows a temporary warning while the background step runs. It clears when the revised post is ready for review. An image only request re renders the existing fact while a factual change goes through source verification. The app updates the existing post and sends it back to review.

I can then approve or request another revision. Let's just look back in the editor to see if that question was changed. It has. What do you find most fascinating about the history of the Watts Towers? Once I'm satisfied I approve the post, Then I use Studio's Publish button.

A background function automatically completes the workflow, so I don't need a second confirmation in the workflow view. This publishes the document to Sanity. This demo doesn't send a post to Facebook or Instagram. My biggest learning was that a content workflow needs more than a status dropdown. Sanity Workflows let me model approval, rejection with feedback, and what happens next.

Customizing Studio brought generation and review into the same workspace. I built this with Codex and refined it by testing real claims and the full editorial loop. AI helps prepare the posts. I make the approval decision.

SEP 2026

  • Walkthrough
  • 1:15

Ask California Black Stories: a Q&A agent that refuses to guess

A chat agent that answers only from a 150-story research archive, cites every claim with real sources, and refuses when the archive doesn't cover the question.

Transcript

This is Ask California Black Stories. It's a chat agent that answers questions about black history in California. And it only answers from a real research archive. 150 stories, 436 claims, Every claim with at least one source. Let me ask it something. Who was Biddy Mason? The answer comes back with the story it drew from and links to the real sources.

Every claim traces to a specific story in the archive. And that's the whole point. No confident guessing from training data. Now this is the part I'm proudest of. Watch what happens when I ask about something outside the archive. It refuses and it stops. It does not reach for general knowledge to fill the gap for a history project, that restraint matters more than coverage.

Under the hood, The archive lives in Sanity as structured content, stories, people, places and sources. A Sanity context knowledge base exposes it through an MCP endpoint. And the Next.js agent reads only that. There are two bugs I caught in testing. The agent once answered from general knowledge after refusing and it once invented citation links.

Both are fixed and verified. The code and the write up are linked below. The archive is the product here. The agent is just an honest way to ask it questions.

SEP 2026

  • Demo
  • 2:28

Exa Gives You Evidence, Not Verdicts | AI Fact-Checking Demo

OpenAI proposes historical claims, Exa retrieves sources, and a separate audit step judges each one. Including the moment it caught a false "first" that came back with credible sources.

Transcript

Accuracy is everything when you're posting history content. One wrong date in a caption and your comments fill up with corrections. So I built a tool that generates fact post ideas, checks them against real sources with Exa, then does something most AI workflows skip. It audits the evidence instead of trusting it. Here's the flow.

Step one I give it topics and it proposes historical claims automatically scoped to California. Step 2 Exa pulls real sources with highlighted excerpts. Then an audit step checks each part of the claim against those excerpts and labels it supported, unsupported, or contradicted. Step 3 I compose the post: an editable fact, an AI generated caption with exactly 5 hashtags, a call to action, and a live 1080 by 1080 preview I can download for Facebook.

Everything saves to a history I can reuse later. Watch what happens with this one. The moment I hit verify, the claim goes into the preview. The first black newspaper in California, the Pacific Appeal, was founded in 1862 in San Francisco. It sounds right. EXA returned solid sources, including the Library of Congress.

But the evidence audit flagged first black newspaper in California as contradicted while supporting the rest. The actual first was the Mirror of the Times in 1856, which was already in my history. So the tool cross checked, it caught the conflict and I hit apply suggested correction. The false first is gone.

This is why the audit step exists. Exa gives you evidence, not verdicts. A search result that mentions your claim isn't proof your claim is true. If I trusted the sources at face value, I'd have posted something wrong on day one. With the corrected fact, I can generate a caption. It comes with exactly five hashtags and I can edit everything before it goes anywhere.

The preview updates live and the download exports a true 1080 by 1080 PNG. Nothing posts automatically. I review every word first, saving it stores the fact, caption, sources and timestamp. And the next time a generated claim overlaps with something I've already verified, the crosscheck flags it before I waste time reverifying.

And here's what it looks like when you reuse post from history. I built this as my demo for the Exa Developer Advocate application because it shows what I actually believe about building with AI search. The retrieval is the easy part. The judgment layer on top of it is the whole product.

SEP 2026

  • Demo
  • 1:13

Bill Buster: Paste a Bill, Pay Less (Convex All Gas Hackathon)

Paste a bill email and it extracts the details, researches competitor prices in your ZIP code, and drafts a negotiation email and phone script. Nothing sends until you approve it.

Transcript

I pay $129.64 a month for Xfinity Internet and I got tired of it. So I built Bill Buster, an app that reads your bill, researches what competitors actually charge, and drafts the negotiation email for you. You approve everything. Before anything sends, Let's add a bill. Here's my Xfinity bill.

I copy the content of the bill email and paste it in. The app extracts the vendor the amount and the billing period automatically. I'm going to add my zip code. Next, it checks competitor pricing with Firecrawl using my zip code so the results are local to my area. I can rerun the search anytime. Next it drafts the negotiation email, citing the competitor prices it found.

Nothing sends on its own. I review the draft, enter the vendor's real billing address, and only then hit send. The negotiation was sent. If I'd rather call, there's a phone script too. An opener, the ask, comebacks for their pushback, and a close. It even tells me to ask for the retention department. And when the new rate kicks in, I record what I saved.

Bill Buster is built on Convex with OpenAI, Firecrawl and AgentMail The repo is linked below. Go negotiate your bills.

SEP 2026

  • Walkthrough
  • 5:30

MotionBrief Walkthrough

One rough campaign prompt to a finished vertical video: the AI brief, the visual, the narration, and the final render. And why I stopped generating the footage itself.

Transcript

Hi, my name is Nikema Prophet and I'm going to walk through my app Motion Brief. Motion Brief takes a prompt and generates an editable brief. And from that brief, we create an image, a narration, and finally put it all together in an MP4 in 9 by 16. So, I'm going to walk through the app. Let's start a new project. I'm going to paste in my prompt. And what we have is a 5-second vertical promo for Field Notes, a site where a software engineer writes up what she builds, including what didn't work. Warm, confident editorial style, the visual,

a desk at golden hour, notebook open beside a laptop showing code, coffee going cold. Okay, so that's the brief that is my input and now we save it and we can generate the AI brief. Okay, the AI brief is ready and we have all of these fields. So we have the project name, audience, objective and headline. And I can change anything that I want here. I don't have to just go with what the AI generated. So I'm going to change this just to building in public just to show that and save it just to show that we can make changes. So we don't have to you don't have to leave any of

this here. We can change everything if we want to. But it's nice to have kind of a starting point. You don't have to think of everything yourself with the AI brief. Okay. Next step. Now we're going to create the visual. The generated brief comes with the visual direction and the image prompt. Oh, I forgot to mention the first step we were using OpenAI, the OpenAI integration.

This whole app is built on DeepSpace and there's a whole list of integrations that you can add to your app. So the first step was OpenAI. In this step, we're using FAL.ai AI to generate the image. And by default, the motion style is slow push in. So whenever it generates, it'll be in that style. So I went ahead and clicked the button. There's an estimated cost because these calls do cost money.

And the image looks pretty much like what we asked for. There's two computers. But you know what? Let me see what happens if I regenerate it. Will I still get two computers? Okay, that's better. So, there's just one laptop, the open notebook, and two cups of coffee. But I guess we can we can go with that. So, there's slow push in, there's slow pull back, there's pan left and pan right. I'm going to go with slow push in for the motion style and save and go to the next step. So, in step three, we add the voice and we get to choose from these different narrators. I am going to go with river. You can change this to it just needs to be between eight and 13 words. So every build leaves

behind lessons worth writing down. I'm going to leave that the way it is and then generate the narration. That is finished. And let's listen. Every build leaves behind lessons worth writing down. All right. So the last step the last step is putting it all together. I forgot to mention. So the voice step is the ElevenLabs integration. So we had OpenAI, FAL, and ElevenLabs. And the last step where we are rendering the final MP4, we're using ShotStack. So ShotStack is going to take everything that we generated and put it together in an MP4. So let's do that. Rendering. Copying

the final MP4 to durable storage. And now it's ready. So here's the final product. Every build leaves behind lessons worth writing down. Just a 5sec Every build leaves behind lessons worth writing down. It's just a 5-second video, but it works. Originally, I wanted to use generative video for this app. I tried that and it was either very expensive and slow or the cheaper model didn't really adhere to the prompt that I put in. I made the decision to change to deterministic camera movement and it works. It's not as exciting. It was a trade-off. The more expensive

video generation looked cooler in my opinion, but it was slow and expensive. So yeah, that's the final product and we can copy the brief. We can download the markdown file which has the brief and links to the media or you could do them separately. So, there's the image, the narration, or you could take the final MP4. So, let me know what you think about the app. Thanks for watching.

SEP 2026

  • Demo
  • 0:50

Building a Human-in-the-Loop AI Content Workflow | Next.js, OpenAI & Supabase

Social Content Agent, an AI-assisted LinkedIn workflow where a person reviews, edits, approves, and schedules every draft instead of auto-publishing model output.

Transcript

Social Content Agent is a deployed AI assisted LinkedIn content workflow. A user can generate a first draft using a reusable brand voice, but the AI output is never published automatically. The user reviews and edits the content, then explicitly saves, approves, and schedules it. The activity timeline makes each key workflow step visible, so users can see how a draft move from an initial idea to a scheduled post. Behind the UI, the app uses authenticated user scoped Supabase data, workflow state rules, and activity history, and a protected scheduled job. Publishing is

currently simulated, which lets the product establish a reliable workflow before taking on the OAuth token management, provider error handling, and account connection requirements of a real social platform integration.