35 Austin Founders Share AI Outcomes

Five hundred skills, compressed to seventy-five

Isaiah Zimmerman

500 skills compressed to 75; brain librarian accuracy 75.6% to 99.3%, finance lead 70% to 91.1%

The problemA year of collecting agent skills produced a pile of five hundred, most of them overlapping and none of them proven, so teams keep installing a chief-of-staff skill that works about three times in four.

Isaiah is the co-founder of Ovae with Christopher Schenk, and the October board put him in first place with fourteen votes. His build starts from a collection of about five hundred skills gathered over more than a year, partly written by hand and partly scraped from public repositories by crawler agents that compared what made one chief-of-staff skill better than another. He built an evaluation harness that ran thirty fixed test prompts against every skill and let the skills overlap, merge and shed whatever was redundant. The run went three rounds and deployed five hundred agents, burning roughly two and a half billion tokens in a day and a half and about a week and a half of a top-tier Claude plan. The first two rounds ran on Sonnet 5 and the third on Sonnet 5.5, which shipped in the middle of the day he was running it. The numbers moved a lot: the chief of staff went from 73 to 84 percent accuracy, the finance lead from 70 to 91.1, the people agent from 74 to 95, and the brain librarian from 75.6 to 99.3. The important design choice was the test data. He did not use synthetic fixtures. Every skill was audited against the saved trace data from his own Claude Code sessions since January, almost ten months of how a real company was actually run, because synthetic data hides the cases that show up in production. The result is seventy-five canonical skills organized into pillar verticals, deployed to every client install through an MCP so updates arrive automatically, with one-click crons and approvals on top of a white-labeled fork of Hermes that Ovae pins about two weeks behind the frontier to work out bugs. He then showed the Ovae app live from his phone's hotspot: a company brain with over 65,000 nodes in the demo environment and 450,000 in the internal one, loading at the same speed. Each node is a chunk with its source, date and document attached, so agents retrieve exact snippets instead of reading whole files, and a single call can connect to as many as thirty related items. Data comes in through a secure bulk upload and is then kept current through connectors, because backfilling a company's entire history over an API is slower and more expensive. The skills canvas groups the seventy-five skills by department and role and starts a working session with one click. Austin's framing for why this matters is that the unit of design is a job title: you give each agent the skills, connectors, routines and stakeholders a person in that seat would have, and enough of those departments make an AI-powered company.

I want to turn a large pile of agent skills into a small, proven library, and I want to prove it with evidence from my own work.

Here is my situation: [DESCRIBE THE ROLES YOU NEED, E.G. CHIEF OF STAFF, FINANCE, PEOPLE, KNOWLEDGE LIBRARIAN, AND HOW MANY SKILLS YOU HAVE TODAY AND WHERE THEY LIVE].

Act as my evaluation engineer. Do this in order, and stop for my approval after step 3:
1. Inventory every skill I have into one folder, with a one-line description of what each does and where it came from.
2. For each role, write thirty fixed test prompts based on real tasks I have done. Ask me to confirm them. Use my own saved session history and traces where you can, not invented scenarios.
3. Propose an evaluation harness that scores every skill against its role's thirty prompts and records the score per run.
4. Run three rounds. After each round, merge overlapping skills, rewrite the weakest ones, and delete any that never win. Report the score change per role.
5. Give me the canonical list grouped by department, with the before and after score for each skill.
6. Show me how to publish the final list through an MCP so every install updates automatically.

Keep token usage visible. Tell me the cost of each round before you run it.

Takeaway If you test on synthetic fixture data, you are going to hit a lot of cases that just aren't true in production.

Landing GoHighLevel as a sponsor in ninety minutes

Darby Rollins

About 90 minutes of human work to start; five contacts, an introduction within two days, and a signed sponsor

The problemBooking a major brand as a sponsor usually means weeks of cold email to the wrong people, so most small shows never try.

Darby is the Chief Agent Officer at Agentic Society and runs Side Hustle Live, a comedy game show for entrepreneurs, as a side hustle. He submitted this build to the board under his full name, and it took second place with ten votes. The idea started with a question: who is already spending money on events, and whose audience is our audience? GoHighLevel was the answer. He is a longtime user, and the company invests heavily in event sponsorships. Instead of blasting a cold campaign, he ran a targeted outreach blitz at one company. He used Claude Code with the Firecrawl MCP to research who at GoHighLevel he should talk to, and it returned the five most relevant people. Then he used Clay to verify their emails and current LinkedIn profiles. That research took about an hour and a half to two hours. Claude wrote the first draft of each message, and Darby edited it and sent it himself, copy and paste, by email and LinkedIn DM, with the pitch that their audience is his audience and the show was a low-risk way to test it. Every one of the five responded. Most said they were not the right person and named someone else, and within two days he was on an hour-long call with the affiliate manager, who introduced him to the person who runs event sponsorships. The conversation then moved at human speed for a few months. The activation happened the night before the Mastermind: GoHighLevel gave away a pair of AirPods Max in a raffle, a second sponsor, Uplift Desk, joined through his partner, and the team gave an on-site testimonial. He had Claude document the whole process as it went and turned it into a reusable skill, and his AI chief of staff kept the follow-ups on schedule so the thread never went cold. Austin named the pattern account-based marketing: pick one dream company, go deep and fast, and have AI write a personalized message for each role. The handoff is the lesson. The agent did the research and drafting. Darby did the conversations.

Act as my account-based outreach strategist. I want to win one specific company as a partner, sponsor or client by going deep on five people there, not by blasting a list.

My target company: [COMPANY]. What I offer: [YOUR OFFER]. Why their audience or customers are the same as mine: [REASON].

Do this in order:
1. Research the company and tell me the five roles most likely to own this decision, with named people, their titles, and a one-line reason for each. Flag who is likely the right person and who is a likely champion to introduce me.
2. List the data I still need to verify before sending (email addresses, current LinkedIn profiles) and tell me how to confirm each.
3. Draft a short, personalized first message for each of the five, written as a peer, not a pitch. Each message should ask whether they are the right person and invite an introduction if not.
4. Draft a follow-up for the second channel (email if the first was LinkedIn, and the reverse).
5. Keep a log of every message sent and a follow-up date. Remind me each morning who is due.

I will review and send every message myself. Never send anything on my behalf, and never automate actions on LinkedIn.

Takeaway How clear can you get on who your dream client or dream partner is, and can you go really deep and really fast at one?

Channel monetized in 2024, with a weekly upload routine since

The problemLocal service businesses rarely out-spend bigger competitors on ads, so Michael looked for a cheaper way to be the name people already trust before the first call.

Michael has been an entrepreneur since 2005 and runs Don't Get Hacked, a tech support and cybersecurity business. When competitors out-spent him on ads he turned to YouTube and committed to one video a week, on the view that steady uploads send backlinks to Google search and shorten the trust gap before a customer ever calls. He monetized his channel in 2024 and has used the same routine for a few other local businesses. AI plays a supporting role: agents help brainstorm topic ideas and draft scripts, titles and descriptions, and a person checks every script for accuracy before it is recorded. He films vertical and horizontal at once and posts the same video across platforms to see which audience responds. Austin noted that the heavy lifting sits behind the camera in planning and distribution.

Act as my content planner. I want to publish one video a week to build trust with local customers.

My business: [WHAT YOU DO AND WHO YOU SERVE].

Give me twenty topics my customers are likely to search for, then draft a script, three title options and a description for my top five. Flag every factual claim so I can check it before I record.

About 1.6 million lines of code in roughly three months, half of it tests

The problemThe software used to design and build spacecraft is expensive, clunky and badly integrated, and even a single requirements tool is usually a months-long project.

Pete was CIO at Firefly Aerospace, which landed on the Moon, and is now co-founding a space-plane company he wants to run as a fully agentic company. His board entry is an automated product development pipeline that runs around the clock and is now about 1.6 million lines of code, half of it tests, because the agents write the test units alongside the code. It began with a requirements tool he expected to take months and had running in a day. From there he asked how far it could go and set up councils of agents: one researches the competing software, reads community forums, Reddit and Discord for the biggest complaints, and comes back with specs and features, and a second council decides whether each feature should be built. Over about three months this became a daily routine in which he has coffee, runs a stand-up with his AI team, and the work rolls on across several rotating Claude Max subscriptions as each one hits its weekly limit. Codex does an adversarial review of the code. This past weekend he set up Hermes with DeepSeek on a second DGX Spark in his own lab to run user acceptance testing: for about three hours the local model clicked through every menu like a buyer inside the company and produced a report, which now feeds straight back into Claude Code with no human in the loop. The suite covers ERP, PDM, PLM and QMS, the digital thread from requirements through design, test and engineering change.

I want to replace an expensive piece of software with one I build using AI agents, and I want the process to run mostly on its own.

The software I want to replace: [NAME AND WHAT IT DOES]. Why it falls short: [PROBLEMS].

Act as my product pipeline designer. Do this in order, and stop for my approval after step 2:
1. Propose the smallest version of this tool that would be genuinely useful, and the build order after it.
2. Design two groups of agents: a research group that reads customer complaints about the existing tools in public forums, and a decision group that reviews each proposed feature. Tell me what each group reads and what it hands back.
3. Design the build loop: agents write code and its tests together, a second model reviews every change, and a test agent clicks through the product like a buyer and reports failures back into the build loop.
4. Give me a one-page morning stand-up format: what shipped, what failed, what the agents recommend next.

Keep cost visible. Tell me what each part of the loop will cost to run before we start.

About 4 to 5 hours of manual quoting per client automated

The problemTurning a sales call into a quote, a signed agreement and a payment is four to five hours of manual document work for every new client.

Micah runs Summit Oak AI, which builds websites that get found and cited by AI models, automates repetitive workflows, and puts a company's software into one dashboard. His board entry is a custom CRM for his own business, and he is candid that he does not recommend everyone build one because it took far longer than he expected. Every client call, from first intro to final go-live, is pulled in from Fathom and graded by call type, so he can see what he did well and what to improve. At the end of each call he repeats back what they agreed, and that recap becomes the basis of a custom quote built with Claude Design inside Claude Code, a polished, animated page that walks the client through the problem, the solution and what will be built. If anything is missing, the system tells him what it cannot quote yet and drafts the clarifying email. When the details are complete he types in the price, and the quote goes out with a Stripe link for signature and payment. Payment triggers a welcome email listing what he needs from the client, push notifications tell him each step, invoices are tracked through Stripe, and a closing document sets out the next steps, a seven-day check-in and a 30-day follow-up.

I want to turn my sales calls into quotes automatically, with me only setting the price.

My business and what I sell: [DESCRIBE]. How I currently quote: [DESCRIBE]. The tools I already use for calls, payments and email: [LIST].

Act as my operations designer. Do this in order:
1. Design the pipeline from a recorded call to a paid invoice, step by step, using the tools I already have where possible.
2. Write the grading rubric an agent should use to score each type of call, and the template for the end-of-call recap I should say out loud.
3. Write the prompt that turns a call transcript into a client-ready quote with the price left blank, and a second prompt that lists the information still missing and drafts the question to the client.
4. Design the follow-up sequence after payment: a welcome email, what I need from the client, a 7-day check-in and a 30-day follow-up.

Nothing should be sent to a client until I approve it.

First text-only close of a high-ticket offer, from a ten-month-old lead

The problemHigh-touch sales to skeptical buyers come from the network, so the hard part is knowing who to reach out to, and when, and what to say.

Mike runs REI Boss, which handles back-end operations for real estate investors, and does similar work for YouTube creators with a product. Both businesses depend on high-touch sales to skeptical customers who will not buy from a webinar, and most deals begin with a text message and close after two or three calls. He describes himself as a much better opener than closer. So he built an agent that filters his calendar invites, Zoom calls and emails to sort out who he should text right now, suggests an opening message, and then coaches him through to the close. He also feeds it his sales transcripts and it coaches him on how to improve his process. The agent surfaced a lead he had not contacted in ten months, wrote the message to warm it up, and guided him to his first close done entirely over text, for a high-ticket offer that mattered to the buyer. He changed the wording here and there, and said it did not feel pushy. He expects the agent to become a major driver of sales in both businesses.

Act as my sales coach. I sell [WHAT YOU SELL] to [WHO]. My deals usually start with a personal text message and close after two or three conversations.

Using my calendar, email and call transcripts, do this:
1. Rank the ten contacts most worth reaching out to this week. For each, tell me when we last spoke, what we talked about, and why now is a good time.
2. For each, draft a short, warm opening text in my voice. I will edit it before sending.
3. When someone replies, coach me one message at a time toward a decision, and tell me why you suggest each line.
4. After each deal, review the transcript and tell me the one thing to improve.

Never send anything for me.

Sixteen parallel agents, with usage and model choice managed automatically

The problemRunning several companies on AI means babysitting chatbots one at a time, and usage costs climb when nobody is watching which model is doing what.

Joshua runs Standout Authority and is building a second company, and his board entry is a 16-lane AI engine in which one coordinator writes the work and a heartbeat feeds sixteen agents in parallel. He built it because he could not spend his days babysitting a single chatbot while running several businesses. He gave the engine every API key from the different model providers and assigned each agent a task, with an agent at every level, including a team of C-level agents. It runs around the clock, watches exactly how much usage is left on every account, picks the model best suited to each task, and routes everything back to his executive agent team, where bots also work as users across the codebase. When it needs him, a voice agent calls him, and it can also call his business partner and his team, so almost everything is done by voice instead of typing. The result, he says, is that he can no longer hit his usage caps and is cutting the number of Claude Max accounts he needs, because the engine spreads work across them and adjusts the model to the job.

I run [NUMBER] businesses and I want a coordinated set of AI agents that works around the clock instead of me managing chatbots one at a time.

The model accounts and APIs I have: [LIST]. The work I want covered: [LIST].

Act as my systems architect. Do this in order:
1. Propose a team structure of agents by level, with a role and a scope for each.
2. Design a coordinator that writes tasks and a scheduler that assigns them to agents in parallel.
3. Design the routing rule: how the system picks the cheapest model that can do each task well, and how it tracks usage across my accounts.
4. Design the escalation: when a voice agent should call me, what it tells me, and what I can approve by voice.

Show me the monthly cost before we build anything.

1,500+ users and 75+ industry verticals, with a two-person team (as posted to the board)

Luke Hayes of RepVantage posted his result to the board in his own words: they launched a platform serving 75+ different industry verticals in the sales realm on June 5th, 2026, and grew to over 1,500 users across subscription-based funnels and above $80K in MRR with a two-person team. It was one of the room's most-voted entries.