AI teleprompter for Meta Ads, a LinkedIn Growth Agent, and a Real Estate Deal Underwriter

Raw phone clips to posted, branded video

Darby Rollins

Built in two days at the hackathon; a web version followed in one more day

The problemFounder-led video is the content that speaks to buyers, but writing the script, staying on it while filming, editing the takes and posting them is enough friction that most founders stop after a few.

Darby is the Chief Agent Officer at Agentic Society, where he helps build the Agentic OS and its skill packs, and he runs a comedy game show for entrepreneurs on the side. The show wants more corporate bookings in 2027, and that was the spark for this build. When he asked Claude which thing he had built in the last ninety days was the most useful and broadly applicable to share, this was its answer, and the room agreed: eighteen votes and first place. He built it in two days at the Agentic Society hackathon as an iPhone app that ties together systems that used to live apart. It scans the market for signals that a specific kind of buyer is about to look for a team-building event, runs a buyer blueprint skill that does targeted research on that one customer profile, and writes a set of scripts and ad angles as starting points. Then he films. A teleprompter sits on top of the recording screen with adjustable speed, so he can do a few takes, stay on the idea and still riff. Once he approves a final take, an editing pass trims it and adds the brand and captions. From there it goes to Meta Ads Manager for paid tests, or straight to Instagram, LinkedIn and YouTube through the Blotato API. Nothing posts until he approves it. This past week he built a modified version for a client, a keynote speaker who is excellent on camera with a script and stuck without one: the scripts are trained on the speaker's voice and writing style, he films in the app, and the takes flow into the content engine Darby already runs for him. The first version is native iOS because Darby had a rough start on an app already. Last Sunday he spent the day having Claude Code rebuild the same components as a web app, which was easy to deploy on Vercel. The app does not generate AI characters. Its job is to keep a real person on script.

I want to make founder-led video ads and social posts without the friction that makes me stop after a week. Help me build a simple app for it.

My business: [WHAT YOU SELL]. The buyer I want to reach: [WHO, AND THE SIGNAL THAT SAYS THEY ARE READY TO BUY]. Where the videos go: [META ADS, INSTAGRAM, LINKEDIN, YOUTUBE].

Do this in order, and stop for my approval after step 2:
1. Write a buyer blueprint: what this buyer cares about, the words they use, and the objections they have.
2. From the blueprint, write five 30-second scripts, each with a different angle and a clear hook in the first line.
3. Plan a small web app with three screens: pick a script, record with the script overlaid as a teleprompter (adjustable speed), and review takes.
4. Add an editing step that trims pauses and adds my captions and brand colors, and a final approval button.
5. Only after I approve a take, send it to my posting tool for organic posts or save it for my ads manager.

Never publish anything without my explicit approval. Keep the first version as small as possible so I can film with it this week.

Takeaway This is the kind of thing that you can build with AI in a very short amount of time if you have a clear idea how to tell Claude what to do.

A LinkedIn agent that interviews you first

Oscar Hidalgo

Saves hours every week (as posted to the board); six to ten interview questions a week

The problemMost of Oscar's consulting clients come from rooms like this one and from introductions. To grow past that he needed LinkedIn, but asking an AI to write a viral post produces content that does not sound like anyone.

Oscar is the CEO of CG Service Pros, an HVAC company in Austin, runs Vindex Consulting, which puts AI into companies across industries, and is an AI advisor at Agentic Society, where he teaches custom agent builds. He is a repeat winner on this board, and this build took second place with thirteen votes. The idea came from Rohan, a member who runs LinkedIn for his clients by interviewing them once a month and turning what they know into posts. Oscar wanted an AI to do that interview, not to write a post on its own. He built it as a Hermes agent and planned it the way he would plan a hire: what information the role needs, what tools it needs, and what the job is. The agent starts with research. It scans the last seven days of his working notes and call recordings in his second brain, and it reads what is changing in AI news through Agent Reach. From that it writes six to ten interview questions about his week. He answers by voice note on his Plaud recorder, which all his agents are connected to, then tells the agent the interview is done and it fetches the recording. The transcript becomes the substance of the post. A draft is written with a set of writing skills, including a humanizer, and a voice file he created of his own writing. Then a sub-agent on a different model does an independent review: it checks the rules, that the post does not sound like AI, that there are no em dashes, that it matches his voice, and that it does not invent data. Pictures come from his Google Drive rather than being generated. When he gives feedback, the agent saves it to a small database of learnings for that agent, so each post is better than the last. The example he showed came from a client call that week, testing a voice agent for a law firm that quoted the wrong fee, and the post is close to what he would say if someone asked him in person. Posting runs through the Composio LinkedIn connector, which logs in with his account and posts directly. The same agent runs outreach through Lemlist, sending connection requests and follow-ups while the platform handles account safety, and anyone who replies is added to his CRM as a lead. His company brain is a version of GBrain, a database plus a GitHub repository. Engagement is low so far, and he expected that: he is new to posting on LinkedIn and treats the profile as a resume for the people his outreach reaches. He still reviews every draft, and will until the agent has earned the right to post on its own.

Act as my LinkedIn content agent. Your job is to turn what I actually did this week into posts that sound like me, by interviewing me first.

Where my week lives: [WHERE YOUR CALL TRANSCRIPTS AND NOTES ARE]. Who I want to reach: [YOUR IDEAL CUSTOMER]. Writing samples of mine: [PASTE 3-5 POSTS OR EMAILS].

Every week:
1. Read my last seven days of calls and notes, and scan current news in my field. Summarize the five most interesting things that happened.
2. Write six to ten interview questions about them. Wait for my answers (I will reply by voice note or text).
3. From my answers, draft one post in my voice. Do not add opinions or numbers I did not give you.
4. Review the draft as a separate, skeptical editor: flag anything that sounds like AI, any em dashes, any claim not supported by my answers, and anything off-voice. Fix it.
5. Show me the final draft and wait for my approval. Save my feedback to a learnings list and read it before the next draft.

Never post without my approval.

Takeaway As with any agent, you want to make sure that it earns its autonomy.

Instagram formats, found and encoded

Cristian Aguilar

About 100 creators studied per client; only reels at roughly 3x the creator's median are kept

The problemHis clients already post consistently and have editors, and still get no results. Finding what actually works means hours of scrolling competitors, chasing one-off virals and writing every script from scratch.

Cristian runs a marketing agency for business owners who post on Instagram. His build took third place with ten votes, tied with Prashant Vanka and ahead on submission time. Shortly before presenting, he asked his second brain, which holds his workflows and context, to make a short training video on the system, and played what it produced. The system does not invent formats. It looks for the ones creators reuse over and over, on the theory that a format working for lots of different people is a structure and not luck. The format's DNA stays fixed: the hook, the structure, the cadence, the length and the payoff. Only the message, the angle and the voice change, so the script fits the client's niche and sounds like them. The work runs in stages. To find formats, scrapers and agents pull Instagram at scale with Apify, Sandcastles and Bright Data, starting from about a hundred creators per client: some in the niche, some competitors, and some from unrelated industries who use the same formats. Every reel is ranked against that creator's own median, and only real outliers, roughly three times the median or more, are kept. Each winner is broken into its format DNA and stored in a vault with three layers tied together: strategies, formats, and a voice file that captures how the client actually talks. Agents write the scripts from the vault. Delivery is a custom content hub built on Supabase: the client signs in once, picks a script card, watches the reference reel, reads the visual and text hook, films from a teleprompter and sends the takes back. One operator runs this across several retainers with no writing team. He has four clients today and is candid that results are mixed while he iterates on formats. Asked about AI clones of the client, he sees a mix: a fully generated channel loses trust, but generated pieces alongside the real person can work, and he is testing a Higgsfield clone for one client.

Act as my short-form content strategist. I want scripts based on formats that are proven to repeat, not on one-off viral hits.

My niche and offer: [NICHE AND OFFER]. Creators to study: [LIST OF ACCOUNTS, OR ASK ME TO HELP YOU BUILD ONE]. How I talk: [PASTE A TRANSCRIPT OR TWO OF ME SPEAKING].

Do this in order:
1. For each creator, find the reels that performed at three times their own median views or more. Ignore anything below that.
2. For each outlier, describe the format DNA: the hook, the structure, the cadence, the length and the payoff. Group outliers that share a format.
3. Rank the formats by how many different creators use them successfully.
4. Build a short voice file from my transcripts: words I use, sentence length, how I open and close.
5. Write five scripts using the top formats, keeping the DNA and changing only the message, the angle and the voice to fit me. Include the visual hook and the on-screen text hook for each.

Tell me which reel each script is modeled on so I can watch it before I film.

Takeaway If a format keeps working for a lot of different people, it is a structure, not luck.

Acquiro: a real estate acquisition desk that underwrites deals for you

Prashant Vanka

Due diligence that took 45 to 90 minutes per deal now runs automatically; 119 comp rules applied by script

The problemAbout nine in ten deals that agents send or list on the MLS miss the margin or are overpriced, and a proper due diligence on each one takes 45 minutes to an hour and a half.

Prashant is a real estate developer in Austin. His board entry tied for third with ten votes and missed the tie-break on submission time, and he gave a full demo on stage. He built Acquiro at the Agentic Society hackathon, with Oscar Hidalgo helping him get started, to automate what he has done by hand for fifteen years. It asks very little of the sender: text or email an address and an asking price, and it does the rest. Four agents sit behind the underwriting. A scout reads a deal in any format, from an ugly email to a PDF, an MLS sheet or a wholesale blast, and builds one standard record in which every fact is tagged as either what the agent sent or what Acquiro found itself. It checks the lot from the county, zoning from the city, the flood map and deed restrictions. Because Austin protects trees and a survey is often missing, he built a tool that looks at a Google Earth image, estimates the tree species from the canopy, then the trunk size and the critical root zone, and works out what can and cannot be built. The underwriter pulls closed new-build comps from MLS sold data, the same size he would build, within half a mile when there are enough and across the ZIP code otherwise, and it says which. It applies 119 comp rules. As Oscar explained, the agent only gathers data and fills in what is effectively a spreadsheet, while scripts do all of the math, so the numbers are predictable enough to feed a prediction model; vision models also read listing photos for a repair estimate. Then it judges the lot the way an owner would, on profit per lot, margin on all-in cost, cash on cash and return on equity in a downturn case, and marks it for review or declines it against his buy box. A declined deal gets a reply in his voice, in three bullets: the comps it pulled, what the lot supports and the most he would pay, so agents learn his buy box and keep sending. If he likes a deal, one click sends an offer through DocuSign on his terms. Outbound keeps every licensed agent in his market on one list and works them on a thirty to forty-five day cadence by email, AI calls and AI texts. The day before the Mixer and that morning he added a competitor permit analyzer that shows who is doing deals in an area and reverse-engineers their buy boxes. The engine can be pointed at another asset class, market or industry; a commercial version that reads lease comps, cap rates and lease expirations is next.

Act as my acquisitions analyst. I want every deal that comes in underwritten the same way, with the math done by code and not by you.

My buy box: [ASSET TYPE, MARKET, SIZE, MARGIN AND RETURN TARGETS]. My comp rules: [LIST THEM]. Public data sources in my market: [COUNTY, ZONING, FLOOD, PERMITS].

For each deal I forward you:
1. Turn it into one standard record. Mark every fact as either "sent by the agent" or "verified by us".
2. Pull the due diligence data from the sources above and flag anything that limits what can be built.
3. Gather comps that match my rules and fill in the inputs for my underwriting script. Do not calculate the results yourself; run the script.
4. Mark the deal "review" or "decline" against my buy box and list the weak spots.
5. For a decline, draft a three-bullet reply to the agent in my voice: the comps, what the property supports and the most I would pay.

Show me every review deal before anything is sent.

2,612 leads for clients and the agency in six months (as posted to the board)

Rees runs Astra GTM, a B2B go-to-market agency. The team started on Clay, which connects data-enrichment providers to language models to run campaigns. When Claude Code got good early this year, they moved the whole five-person team onto a Claude Code operating system, dropped Clay, and now buy direct API subscriptions to the underlying tools, with every skill and SOP in a shared GitHub repository. Everyone works out of Claude Code every day, and campaign quality and client results both went up. The leads come from email and LinkedIn automation campaigns.

30 white-label apps and about 2,500 users in six months (as posted to the board)

Dan, a former Austinite who flew in from California, sold a SaaS company and partnered with his former business coach, James Kemp, on an idea: put a coach's methodology behind a paywall as an agent that actually does the work. Six months and thousands of commits later, Tool Chat (a working title) is a custom harness on top of Claude Code where a client talks to the coach's agent and it helps write the offer, build landing pages and funnels and launch ads. Coaches upload their videos to build the agent, add skills, and sell it with Stripe built in; there is also a community, a classroom and chat over the coach's content. Dan owns the technology and licenses it as white label, taking a margin on AI usage. He did not write a single line of the code himself.

Micah of Summit Oak AI is not a video editor, so he built a skill to do it. Drop in raw footage or a link and give it direction, plus your fonts, branding or reference reels, and it cuts shorts and clips, adds captions in your brand and produces cover images ready for Instagram. The skill is public on GitHub for anyone to take and improve, and a larger update is coming.

Cory is a real estate investor, broker and the marketing and sales director for AeroVista, a hangar home community in Salado, Texas, where owners can taxi their plane into the garage. He describes himself as a one-man, fifteen-person office with ADHD, and his rule is that what he cannot see does not get done. Project North Star pulls his CRM, social accounts and every daily tool, fifteen browser tabs in all, into one dashboard of KPIs and reports on any timeframe, with a switch to personal goals such as workouts and nutrition. Every part of it can be moved and rearranged.

An agentic analytics kit for business owners

Bryce Sandry

Bryce runs Connectricians. His analytics kit gives a business owner an agent that walks them through connecting their key data sources, with skills for where to store credentials securely, where to host and publish, how to refresh every day, and how to validate and test the result. The output is an analytics app that tells owners things about their business they probably did not know.

Sam is a business coach from Australia, where contractors are called tradies, and his clients were breaking the formulas in his Google Sheets templates. He asked Claude what to build and ended up with a client portal: members set their mission, revenue and net profit targets, map the team they need on an org chart, and the portal produces a month-by-month forecast for the financial year from the unit economics behind each hire, with cash flow modeling to keep them on target.