Shipping with AI

Shipping with AI

The show where AI goes from hype to implementation. Real workflows, on screen, walked through by the people who already built them. Every episode is one you can steal. shippingwithai.substack.com

  1. 6 days ago

    Three AI Employees Cost Him $3,000 a Month and Never Quit | Itamar Novick, Recursive Ventures

    Itamar Novick runs a small VC firm with a very specific problem: there’s only so much work a small team can do.So instead of hiring more people, he started building AI employees.Henry acts as his chief of staff. Jack works like an AI associate, helping with due diligence, investment memos, meeting prep and research. Simon focuses on content, learning from Itamar’s past posts and engagement data to get better over time.In this episode, Itamar opens up the system behind Recursive Ventures and shows how he uses these agents day to day. Henry can pull context from email, calendar, CRM, Notion, Dropbox and meeting transcripts to prep him for a call. Simon finds trending topics, writes posts in his voice, and then learns from what performs well. He also gets into what went wrong along the way. In this episode: * Why Itamar built AI employees instead of hiring, and how Henry, Jack and Simon divide the work * How Henry acts as an AI chief of staff across email, calendar, CRM, Notion and meeting data * How Simon creates content from Itamar’s past writing and uses engagement data to improve over time * How Itamar manages AI context, memory, permissions and reliability across the firm * Why he believes small businesses can replace parts of their SaaS stack with custom AI systems * What it actually takes to build these systems with Claude Code, including cost, architecture, specs and human oversight Timestamps: 00:00) Introduction (01:26) Why Itamar started building AI employees (03:38) What work can actually be automated in a VC firm (05:20) The philosophy behind cloning your work (07:19) How Henry, Jack and Simon came together (11:17) Inside the AI system running Recursive Ventures (12:48) How Henry works as an AI chief of staff (15:00) Connecting email, CRM, Notion and meeting data (16:29) Live demo: AI prepares Itamar for a meeting (19:21) How he manages AI access and permissions (22:45) Where the system still fails (25:45) Inside Simon, the AI content agent (26:13) Live demo: AI-generated LinkedIn content (28:18) How Simon learns from content performance (31:10) Closing the AI learning loop (33:42) What it costs to run the system (36:21) The SaaS apocalypse (38:43) Building an AI-first service organization (41:20) What to build first (42:18) How to start building your own AI system (43:00) Why you need a spec before vibe coding (46:18) Production systems vs. AI demos (47:22) Why cloning himself became a necessity Resources & Links: * Claude Code: https://claude.com/product/claude-code * Recursive Ventures: https://www.recursiveventures.com/ * Itamar’s Website: https://www.itamarnovick.com/ Connect with Itamar Novick: * LinkedIn * X This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit shippingwithai.substack.com

  2. 23 Aug

    Claude and Obsidian Turned Her Inbox Into a Daily Brief | Natasha Lee, AI Consultant

    Nat Lee has a lot going on at once. She’s running projects, teaching, making content, managing partnerships, and dealing with the hundreds of small things that come with all of that. And email was becoming the place where all of it got messy. So she built a second brain using Claude and Obsidian. In this episode, Nat opens up the whole system and shows how she uses it day to day. Emails get pulled into Obsidian, deadlines and commitments get tracked, and a daily brief tells her what needs attention and what’s overdue. She also shows what went wrong along the way. Dates were one of the first problems. If an email said “tomorrow,” Claude couldn’t just guess what that meant. So she started giving every email its actual sent date and made Claude calculate the date from there. Then there’s her knowledge wiki, which helps Claude find the right information without having to read hundreds of documents every time. In this episode: * Why Nat chose Claude + Obsidian for her second brain * How an incoming email becomes a task and appears in her daily brief * How she manages context across multiple projects * Why giving AI access to everything can actually make it worse * How she wants to move from multiple agents toward an AI chief of staff * The biggest mistake people make when building their first second brain * How to start building your own system without making it an emotional hurdle Timestamps: (00:00) Introduction (00:37) Why Nat built a second brain (01:45) Inside her Claude + Obsidian setup (02:44) Why Claude + Obsidian? (04:36) The problem with AI getting dates wrong (04:57) How she makes Claude calculate dates correctly (07:27) How much iteration did this take? (08:38) Why not just use a calendar? (09:17) Live demo: email to Obsidian (10:39) How the daily brief works (11:45) Turning email into task management (13:30) The biggest challenge: managing context (15:57) Building toward an AI chief of staff (17:06) Where the human stays in the loop (19:24) How she manages context with Claude + Obsidian (19:42) Building a knowledge wiki (21:35) How she organizes documents inside the wiki (23:36) Live demo: Claude builds a knowledge wiki (28:18) How to improve the wiki with your own context (29:13) How to build your first knowledge wiki (31:32) Start small and keep it manageable Resources & Links: * Claude Code: https://claude.com/product/claude-code * Obsidian: https://obsidian.md * Token Thoughts: https://www.linkedin.com/newsletters/token-thoughts-7460866816886824961/ Connect with Natasha Lee: * LinkedIn This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit shippingwithai.substack.com

  3. 9 Aug

    The Claude Code workflow that turns a calendar invite into a Sales demo | Bhanu, SiteGPT.ai

    Prep was the bottleneck, not the pitch. Bhanu Teja runs every enterprise sales call at SiteGPT.ai himself, and building one personalised demo by hand took two hours on a good day and closer to three on a bad one. That capped him at a single demo call a day. So he described the manual process to Claude Code in plain English and asked it to turn that into a skill. Now he screenshots the calendar invite, drops it in, and out comes a chatbot trained on the prospect’s own site carrying their logo and their brand colours. A prep sheet follows as a Claude artifact: how the call should flow, the questions coming, the compliance documents to keep open, and what not to ask. Past quotes and past deals sit in Claude Code’s memory, so it tells him where to price before he goes looking. After the call the recording goes back into the same session and the follow-up lands before the prospect reaches the next vendor. He builds one live on this recording, for the hosts’ own website, and barely watches it run. Prep is now 15 minutes, five to ten of which is Claude working on its own.In this episode: * Why a screenshot of a calendar invite is enough to start the entire run * How a demo you build beats a demo you describe, and why he hands the bot over afterwards * The prep sheet that answers objections before the prospect raises them * The government agency that legally couldn’t pay upfront, and how Claude caught it first * Why past deals in memory decide the pricing, not the model * The time zone slip that nearly went out, and the rule he wrote the same day * What changes when you stop being the bottleneck in your own pipeline * The three step path to building this yourself without writing codeTimestamps: (00:00) Introduction (00:34) The deal where the system did the selling (01:48) Screen share: calendar screenshot to live chatbot (03:01) What is inside the skill, and what is not (04:32) Built from scratch, sharpened over real calls (05:52) Personalisation past the logo and brand colours (09:06) The prep sheet: flow, questions, objections (11:40) The government deal Claude flagged before he did (12:50) Follow-up email straight from the call recording (13:52) How it knows what to quote (16:29) The day it got the day wrong (17:47) If you sell consulting instead of software (19:01) Removing himself as the bottleneck (20:42) Two hours of prep down to fifteen minutes (23:03) Three steps to build this yourselfResources & Links: * Claude Code: https://claude.com/product/claude-code * SiteGPT: https://sitegpt.ai Connect with Bhanu Teja P: * LinkedIn * X This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit shippingwithai.substack.com

  4. 2 Aug

    She Wrote as Much in 1 Month With AI as in Her First 8 Years in VC | Melody Koh, NextView Ventures

    Writing was the bottleneck, not thinking. Melody Koh, a partner at NextView Ventures, published roughly as much in her first eight years as an investor as she now ships in a single month. So she built a content operating system. A skill called learning loop reads every session before the context is lost and files the mistakes and debates into a judgment ledger, which fed around 80 percent of her first ten Substack posts. A content editor scans newsletters and X to pick the week’s topic, six editorial agents tear each draft apart for accuracy, voice, and AI patterns, and her hand edits get diffed against the draft and folded back into the system. On the morning of this recording, a full post produced itself while she wasn’t at her computer. In this episode: * Why writing became a byproduct of judgment that compounds * How a judgment ledger turns corrected mistakes into publishing material * The learning loop that makes every session smarter than the last one * Why she spends 90 percent of her time on the kernel and almost none on typing * Six editorial agents that read every draft cold as founders, VCs, and operators * The calendar grid hack that fixes AI’s day-of-the-week problem * The X engagement skill that made one post perform 10x better * What she refuses to delegate, and why taste still has to come from her Timestamps: (00:00) Introduction (00:50) A VC partner who builds every day (03:07) Eight years of writing, matched in one month (04:13) Judgment compounding, explained from scratch (06:58) When the process starts generating the content (09:21) Where differentiated viewpoints actually come from (09:43) The day a model update broke her guardrails (12:57) Screen share: inside the judgment ledger and a content editor run (20:52) How the judgment layer got built, iteration by iteration (26:23) Screen share: produce post running end to end (27:07) Why AI can’t tell you what day it is (33:10) The parts she refuses to delegate (36:32) Two personas that audit Claude’s own recommendations (38:21) Has the content actually performed better? (39:01) The X engagement skill and the 10x post (42:17) Three things to do if you’re starting today (43:30) Wrap up Resources & Links: * Claude Code: https://claude.com/product/claude-code * Learning Loop skill (public repo): https://github.com/melodykoh/learning-loop-skill * Content Machine: https://github.com/melodykoh/content-machine * NextView Ventures: https://nextview.vc Connect with Melody Koh: * LinkedIn * X This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit shippingwithai.substack.com

  5. 26 Jul

    She Raised 7 Figures in Sponsorships by Automating LinkedIn With AI | Sherry Jiang, Peek Money

    Cold outreach at volume is a numbers game, and volume needs people. Sherry needed close to seven figures in sponsorships for her AI conference, with no sales team.So she automated it. Manus built a list of 500 companies that had already sponsored comparable AI conferences. She filtered for the roles that field sponsorship inbound, then pointed a browser agent at her LinkedIn to send requests one by one, mimicking a human instead of hitting an API that would get her blocked. It ran overnight. Replies came around 10 to 15 percent, with no personalisation at all. In this episode: * Why she raised seven figures in sponsorships without a single salesperson * How a research agent produced a 500 company sponsor list from past conference exhibitors * Why she removed personalisation from her messages, including the recipient’s name * How browser automation gets around LinkedIn’s limits when APIs get blocked Why mutual connections beat personalised copy on reply rate * How she filtered for growth, business development, developer relations, and founder titles * Why the agent has to be painfully slow, and why that means running it overnight * What the whole thing actually cost compared to a virtual assistant at 7 to 10 dollars an hour * Why AI outreach is only as good as how clearly you’ve defined who you wantI Timestamps: (00:00) Introduction (00:49) Seven figures in sponsorships, no sales team (02:02) How she did it before automating (03:19) Why she cut the personalisation (04:41) Building a 500 company list with Manus (06:05) Demo: researching sponsors live (09:59) Why the agent needs its own computer (11:10) Getting around LinkedIn’s automation limits (13:22) Real messages, sent by AI (14:36) What the credits cost (16:03) The 15% reply rate (17:15) The wrong name mistake (18:40) Get clear before you automate (20:08) Agency and a half million dollar deposit Resources & Links: * Manus * Codex * AI Engineer Singapore Connect with Sherry Jiang: * LinkedIn * Peek * X This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit shippingwithai.substack.com

  6. 19 Jul

    An AI Influencer Got His App 25,000 TikTok Views for $200 | Ayush Chaturvedi, Voibe

    Most founders know they should post more short-form content. The hard part is finding what works and staying consistent without a full UGC team. Ayush had an extra constraint: his customers are in the US, but TikTok is banned in India. So he built a workaround. In this episode of Shipping with AI, he walks through the AI UGC workflow he uses to test TikTok content from India. He created Kate, an AI influencer for his dictation app, then used Fastlane.ai and Claude to generate content, find proven formats, and post through warmed up US accounts. In this episode: * Why TikTok being banned in India does not stop Ayush from testing TikTok content * How an AI influencer generated 25,000 organic views in one monthThe four narrow sub-agents behind the analyst, and why narrow scope beats one agent doing everything * How Fastlane.ai uses warmed up TikTok accounts to publish into the US market * Why people accounts can work better than brand accounts for UGC-style content * Why Ayush treats the account as a testing ground before spending on paid ads * How the workflow replaces a creator, editor, researcher, and social media operator * Why AI content quality matters more than whether the audience knows it is AI Timestamps: (00:00) Introduction (00:43) Why TikTok is hard to access from India (01:53) Running TikTok without ever using TikTok (03:33) Meet Kate, the AI UGC influencer (05:25) 25,000 organic views from 54 posts (06:31) How Fastlane.ai finds proven TikTok formats (08:03) Using warmed up TikTok accounts in the US (10:01) Creating the AI influencer with Claude (13:06) Why use a person account instead of a brand account (14:31) Testing hooks before scaling content (15:38) Creating AI images and videos for Kate (19:36) Using Blitz to pick viral content formats (20:19) Turning a proven format into a VoIP promo (22:13) What content formats worked best (23:41) Scheduling 8 to 10 days of content in 2 hours (24:38) Tracking branded searches and conversions (26:27) Cost breakdown of the AI TikTok workflow (27:50) Breaking even with one or two customers (28:56) Does AI content cap out on TikTok (30:06) Why this replaces a full UGC production team (31:54) The future of AI-run marketing workflows (34:40) Final thoughts and wrap-up Resources & Links: * Claude * Fastlane Connect with Ayush Chaturvedi: * LinkedIn * Voibe This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit shippingwithai.substack.com

  7. 17 Jul

    How a VC runs his investment workflow on AI agents

    Discover how a seed-stage investor turned a slow, manual research process into an AI analyst he runs from a single Slack message. This deep dive walks through why most funds won’t let AI near their confidential deal documents and why he did it anyway, the exact architecture that stops the agent from inventing numbers on financial data, and a live demo of the whole thing pulling a full company brief from four connected systems in seconds. In this episode: * Why trusting AI with confidential data is really a question about the vendor, not the technology * How he built the first working version over a weekend, and why it is never actually finished * The four narrow sub-agents behind the analyst, and why narrow scope beats one agent doing everything * The rule that keeps it honest: clickable sources, a human review step, and treating the agent like an intern * When to call a plain function instead of letting the model use judgment, so it stops making things up * Why the near future is agent to agent, not human to human, and what that changes for founders * How to build your own, whether you want a frontier model’s harness or full open-source control Timestamps: (00:00) Introduction (00:40) The problem: Portfolio answers scattered across systems (02:49) Why let AI near confidential deal documents at all (04:11) Do you actually trust OpenAI and Anthropic with your data? (04:26) Choosing a harness and locking it down with Tailscale (07:36) Building the first version over a weekend (08:47) Live demo: a full company brief from one Slack DM (10:48) Why the near future is agent to agent (12:30) Verification: clickable sources and the intern rule (15:14) The architecture: Claude Code and the Agent SDK (18:24) Compounding org knowledge with ByteRover (21:18) The four sub agents and the data cleanup grind (24:43) When to call a function instead of trusting LLM judgment (28:44) Loops, goals, and orchestrator burnout (29:47) Advice for building your own AI analyst (32:51) Turn paranoia into verification, then just build (34:06) Wrap up Resources & Links: * Claude * Attio’s docs built for AI agents: https://attio.com/llms.txt Connect with Binh Tran: * LinkedIn * avv.co This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit shippingwithai.substack.com

  8. 17 Jul

    Why Your Vibe Coded App Isn't Ready to Ship (A CTO Explains)

    Discover how a CTO turned the messy problem of vibe coded apps into a single Claude skill that tells you whether your app is actually safe to ship. This deep dive walks through why asking AI “is this production ready” burns through hundreds of thousands of tokens and still gives you the wrong answer, the seven to ten things a CTO actually checks before an app goes live, and two real internal tools audited on screen, including the moment the skill catches a leaked secret key and app logic sitting wide open on the client side. In this episode: * Why maintainability matters as much as security the moment real people start using what you built * What a non-engineer simply cannot see when they look at a working app, and where it breaks * Why just asking AI if something is production ready burns 200 to 250K tokens and changes its answer the next day * How an app ends up split across Vercel, Railway, and AWS when the AI keeps improvising fixes * The CTO’s checklist: tech stacks, backing resources, code repos, secret scanning, libraries, and schema * A single-prompt Claude skill that ranks every issue by how soon it will bite you, then hands back a fix list * Where the skill still traps you in a fix-then-break loop, and the exact moment to bring in an engineer Timestamps: (00:00) Introduction (01:07) The problem: your team can build apps but can’t tell what’s safe to ship (02:26) Why maintainability matters, not just security and safety (03:35) What a non-engineer can’t see when they look at working code (05:27) Why just asking AI “is it production ready” fails (05:44) The token burn: 200 to 250K per run, and answers that change daily (06:37) How an app ends up split across Vercel, Railway, and AWS (07:16) The CTO’s checklist for what makes an app production ready (08:09) Tech stacks, backing resources, code repos, and secret scanning (10:19) Live demo: auditing a simple work redesign app (11:47) The prototype a non-coder built in minutes (12:43) How the skill runs the audit on the app (13:14) The biggest risk it caught: all logic on the client side (14:25) Second example: the asset tool that replaced a SaaS (17:08) Walking through the audit results and the fixes (17:51) One prompt to fix a leaked secret key and more (19:31) Running the skill with Claude in thinking mode (20:02) Which model and thinking mode to use for fixes (20:40) How to tell if you’ve overbuilt or burned too many tokens (22:13) Inside the prompt: sanity checks and severity ranking (23:14) How non-engineer proof the skill really is (23:57) Does it explain the risks in plain language? (24:53) Has the skill ever missed anything significant? (25:30) The fix-then-break loop and how to get out of it (27:56) What maintainability actually means (29:25) Does it work with only the prompt history? Resources & Links: * Claude * https://github.com/impressai/vibe-check-skill * https://www.12factor.net/ Connect with Dr. Vaisagh VT: * LinkedIn * Impress.ai This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit shippingwithai.substack.com

See All (15)

About

The show where AI goes from hype to implementation. Real workflows, on screen, walked through by the people who already built them. Every episode is one you can steal. shippingwithai.substack.com

Information

  • Creator
    Shipping with AI
  • Years Active
    2026
  • Episodes
    15
  • Rating
    Clean
  • Copyright
    © Shipping with AI
  • Show Website
    Shipping with AI