AI Skills That Replace Entire Workflows

A founder built an automated outbound sales pipeline with Claude Code in a single working session — Google scraping, email enrichment, campaign loading, all running 24/7 on Railway. A few days later, someone booked a demo. "Where did you come from?" he asked. "You sent me an email," they said. He didn't remember sending it, because his AI BDR did the work.

That's the shift surfacing at the sixth AI roundtable. The founders pulling ahead are focused on building persistent systems — skills, automated pipelines, executive assistants — that work while they sleep. The AI race has coalesced into three distinct groups: the AI system builders in the lead, the AI micromanagers falling behind, and the incumbents still standing at the starting line.

What follows is shared under Chatham House rules.

"Projects Vertical, Skills Horizontal"

One founder had a framework worth stealing: "Projects vertical, skills horizontal." His context was Claude Code. Every project — a codebase, a client, a campaign — is a vertical silo of context. Skills are the reusable capabilities you carry across all of them.

He'd built this into a system. An extraction prompt drops into any project and pulls out everything interesting he'd done — the decisions, the clever bits, the things worth writing about. Then an article-writing skill takes that raw material and produces blog posts, LinkedIn newsletters, dev.to articles, and Substack pieces. His writing output was almost entirely automated from the actual work he was doing. Build in public, literally.

"With the right skill, Claude's writing is just phenomenal," he said. "Better than human."

I use a similar approach. My website is a Git repository, and I have a social-post skill that picks an essay, extracts quotes, generates a post in my voice, and queues it through Buffer. It works because it quotes from things I actually wrote, which keeps the posts from sounding like AI slop. I talked about this pattern at the second roundtable, and six sessions in, it's still the most reliable way to automate content without sounding like every second post on LinkedIn.

The shift from prompt engineering to skills engineering is real. Prompts are one-shot. Skills persist. You build them once, refine them over time, track them in version control, and share them across your entire operation. As I told the group: you can just tell Claude to make you a skill, and then refine it as you use it. The barrier to entry is lower than most people think.

The Cost of Trying Is Nearly Zero

One SaaS founder put it simply, and the whole table nodded: the less he worried about whether AI would understand his request, the more productive he became.

"The sooner I skip that — like, I haven't had the time to think through this, let me put it off another day — the sooner I skip that [procrastination] and just send it a couple sentences, the better," he said. "And then I'm like, oh, it did so well. And if it doesn't, then I can spend more time."

His turning point was realizing the cost of failure is almost zero. If the AI does something wrong, roll it back. What did you waste? Five minutes and some tokens. The old instinct — plan the prompt carefully, describe every requirement, think through the edge cases — is a holdover from a world where the cost of redoing work was high. Software is disposable now.

He'd started vibe-coding internal tools this way. A pricing test? Not just a one-off A/B test — he vibe-coded a pricing test management system so he could spin up new experiments without touching code again. An internal analytics dashboard? Same approach. A couple of sentences in, result out, iterate fast.

We're all discovering the same tradeoff, though. You'll outrun your ability to organize what you've built. Another founder at the table admitted he had "a lot of dropped projects" — things AI spun up quickly, but that still needed human effort to deploy and maintain. "You still need to put a little work to get it deployed or become a process," he said. The speed of creation now outpaces the speed of adoption.

AI That Sells While You Sleep

The most concrete sales win came from the founder who'd built a StanStore pipeline. His platform targets coaches and creators who sell digital products through Stan. There's no public directory, but you can search for Stan stores by industry through Google. So he built an entire workflow: automated Google search, pagination through results, email and LinkedIn enrichment, and automatic loading into a SmartLead campaign.

"Every day I just get a notification on Telegram," he said. "'We found 105 stores. Of that, 48 emails and 23 LinkedIn profiles.' And it just automatically loads it up into a campaign."

He built the whole thing in one Claude Code working session and deployed it to Railway. It runs 24/7. Demos show up on his calendar from people he never personally contacted.

Another founder took a different angle on AI-powered sales. His SaaS is self-serve, low ticket, no credit card required — they'd never done demos. But they started using Kurumi, a tool that runs AI-powered product demos as voice conversations. The AI agent logs in to a real account in their system via a headless browser, navigates the product live, creates real appointments, and answers questions. It feels like a Zoom call, minus the video.

"We never did demos," the founder said. "We just never had a sales flow. But now that we can do it 24/7, we're testing it out."

They're not measuring against human demos — they never had any. They're measuring whether the AI demos lift trial-to-paid conversion. It's a clean experiment: before AI demos, no demos at all. After, AI demos on every page.

A third founder was planning something simpler but potentially just as valuable: agents that crawl all their Intercom articles, emails, and blog posts whenever they launch a feature, then flag anything that's out of date. He'd tried monitoring Reddit mentions manually before, and "it was soul-sucking." This felt like it could cut out the soul-sucking part.

The pattern across all of this: the founders getting the most out of AI for marketing and sales aren't just generating content. They're building pipelines. The content generation is one step in a larger automated system — scraping, enrichment, outreach, demo, follow-up. The value isn't in any single AI output. It's in the machine that keeps running.

When the Founder Becomes the Bottleneck

The most honest admission of the session came from a CEO who said plainly: "I am now increasingly the bottleneck."

His developer had adopted AI eagerly — no push needed. Customer service was getting set up with Intercom's Fin AI resolution bot. But everything else — marketing, blog writing, knowledge management — was stalled while he figured out the right workflow. He'd had false starts. Early attempts at AI-generated blog content were "pretty bad." He knew it was about building the right context and flow, but hadn't gotten there yet.

Another founder had the same problem in a different form: a "vague cloud of things I've talked to Claude about" — skills he'd written, commands he'd built, projects he'd started — with no system to find any of it later. The tools existed — Supersets for giving each coding agent its own branch and database, Paperclip for agent orchestration — but the core problem wasn't tooling. It was that AI makes starting things so easy that the finishing becomes the constraint.

I shared my own approach to staying on track: a Git repository that functions as an AI executive assistant. It generates daily briefings every morning and evening, pulling from my calendar, Gmail (read-only), CRM, quarterly goals, and a directory of markdown files that hold my business context. A Mac scheduler kicks it off at 9 AM and 9 PM, and Pushover sends me a push notification when the briefing is ready. The primary function is accountability — it tells me what I said I'd do and whether I've done it. Like a human EA, but one that runs on a cron job and never takes a day off.

"That's really simple and elegant," a CEO said. And that's my suggestion. Start with the simplest thing possible. Your goals are in a markdown file. Your projects are in a directory. Your briefing is a scheduled prompt. Claude Code reads all of it and tells you what matters today. No task manager integration required — just files, a scheduler, and a push notification. If it turns into a habit, you can iteratively ask Claude for improvements.

The pattern across six sessions of this roundtable is consistent: the founders pulling ahead use AI thoughtfully, not frantically. They're building around it using Claude skills that persist, pipelines that run, briefings that show up whether they remembered to ask or not. The power of their AI systems grows over time. Contrast this with CEOs and teams spending the day micromanaging smaller AI tasks. The gap between building with AI and chatting with AI grows with each passing week.

Building still requires judgment about what's worth automating. Several founders at this table had more half-finished AI projects than running ones. The speed of creation can easily outpace your practical needs. But the founders who start simple — a skill, a scheduled prompt, a markdown file — find it easier to maintain balance between the pull of building and the need to grow a business.

Resources From the Roundtable

• Claude Code — Terminal-based AI coding tool from Anthropic. Multiple attendees use it for development, marketing automation, and system building.
• Kurumi — AI-powered product demos as voice conversations. Logs into a real product account via headless browser and walks prospects through the software.
• SmartLead — Cold email automation platform. Used to receive enriched leads from an AI-built scraping pipeline.
• Railway — Cloud deployment platform. Used to host always-on automation workflows built in Claude Code sessions.
• Supersets — Development tool that gives each AI coding agent its own Git worktree, database, and local environment.
• Paperclip — Agent orchestration platform with hierarchical agent management and permissions.
• Intercom Fin — AI resolution bot for customer support. Mentioned as a first line of defense before human support agents.
• Buffer — Social media scheduling tool. Used via MCP server connection to Claude Code for automated post queuing.
• Pushover — Push notification service ($10 one-time purchase). Used to deliver AI executive assistant briefings to mobile.
• Stan Store — Digital product platform for coaches and creators. Target of an automated lead-generation pipeline.