Posts related to technology category

You Don't Have a Skill. You Have a Novice.

Vinay Patankar · 08 Apr, 2026 · Technology

You don't have a skill. You have a novice. My team keeps telling me they've "built a skill." One person gave Claude a short prompt and hit create. Another found something on a marketplace and installed it. Both walked away thinking the job was done. It wasn't. They didn't build anything. They downloaded a stranger and handed it the keys. And the stranger is kind of an idiot. People treat AI skills the way we used to treat WordPress plugins. Install it, expect it to work. That mental model made sense for traditional software. Teams tested thousands of edge cases before shipping. AI skills don't work like that. A freshly created skill is untrained. It's never encountered your business context, your edge cases, your definition of "good." I learned this the hard way while building one skill through about 100 test runs: AI isn't magic when the system has to compound. ## The split most people miss There are two types of AI skills, and the difference matters more than most people realize. Generic skills work out of the box. "Run an SEO audit." "Summarize this article." "Generate a compliance checklist." The skill doesn't need to know you or your business to do an adequate job. Context-dependent skills are completely different. "Write a post in my voice." "Prepare my weekly board report." "Draft a customer email that sounds like me." These need your tone, your audience, your standards. A fresh skill reads like AI wrote it. Because AI did, without hundreds of corrections. Karpathy coined "vibe coding" in 2025. A year later he walked it back. The vibes weren't enough. Production requires structure. The same applies to skills. The creation is the vibe. The training is the structure. ## What training actually looks like The gap between a novice skill and a hardened skill is the gap between a new hire on day one and that same person after a year of direct feedback. The skill has to learn what "too formal" means for your brand. What "too long" means for your audience. Which edge cases to handle and which to flag. What your definition of done actually looks like. This takes hundreds of feedback loops. Not dozens. Hundreds. I've watched skills go from producing generic, forgettable output to nailing the exact tone, format, and edge-case handling we need. The difference between iteration 10 and iteration 200 is night and day. Most people give up at iteration 3 and conclude that "AI skills don't work." ## Why this matters now The AI skills ecosystem is exploding. Marketplaces, skill libraries, prompt templates, agent frameworks. The barrier to creating a skill has dropped to near zero. You can have a working skill in under a minute. But "working" and "production-ready" are separated by a canyon. The competitive advantage in 2026 comes from infrastructure, not intelligence. The infrastructure is the training loop. The intelligence is what comes out after hundreds of cycles. Teams that understand this will build skills that compound. Teams that don't will keep installing novices and wondering why AI feels underwhelming. A skill you haven't trained is not a skill. It's a first draft.

Stop Buying Ten AI Agents. Buy One That Builds the Other Nine.

Vinay Patankar · 06 Apr, 2026 · Technology

I had 14 tools. Each solved one problem. None shared context. So I gave one coding agent access to everything and told it what to build. That only clicked after I stopped treating Claude Code as a coding tool and started seeing it as an operating system for work. Here is the architecture I ended up with. I gave a single coding agent access to my files, my email, my calendar, my CRM, and my notes. Then I started asking it to build things. "Build me a triage system that reads my inbox every morning and drafts responses." It did. Wrote the scripts. Connected the APIs. Tested it. Deployed it. "Now build a daily briefing that pulls from my calendar, CRM, and Slack." Same thing. Built it in a session. Runs every morning at 5 AM. "Now build a content pipeline that takes my voice notes and turns them into LinkedIn drafts." Done. Each new capability took hours, not months. Each one had full access to everything the others knew. No data silos. No integration layer. No middleware. Now running 20+ automated workflows. Did not buy 20 tools. Bought one coding agent and told it what to build. The economics are simple. A coding agent costs the same whether it builds one thing or one hundred things. Every additional capability is marginal cost, not a new subscription. But the real advantage is not cost. It is context. Every workflow my agent builds has access to every other workflow. My content system knows what meetings I had this week. My CRM updater knows what emails I sent. My daily brief knows what tasks are overdue. Try getting that from 14 separate tools. The companies selling point AI solutions are building the next generation of software silos. The companies buying one coding agent and building their own stack are building something fundamentally different: a personalized operating system that gets better every day. Which one are you building?

Skills vs Subagents: Why I Decided Against the Upgrade

Vinay Patankar · 04 Apr, 2026 · Technology · Productivity

I almost converted my AI task manager into a "subagent." Then I thought through the tradeoffs and decided against it. Here's my reasoning. The setup: I have a task-helper that runs every 2 hours. It scans my active task list, picks the highest priority item, does research, writes drafts, and posts updates. Fully autonomous. I figured making it a proper isolated subagent would be an upgrade. That same task-helper later became one of my favorite examples of an assistant that knows when the work is already done. So I asked Claude to reason through it. Its response: "When task-helper runs as a skill, it inherits your full vault context. A formal subagent starts with a blank context window. It would need all of that explicitly passed in, or it wouldn't know your vault folder structure, safety rules about outbound comms, which Discord channels to use." Then: "Skills are playbooks an employee has memorized. Subagents are delegating to a specialist in another room." And the kicker: "Converting task-helper to a subagent would be a lateral move with added complexity. The right use for subagents is inside a skill, when you need to do research and drafting in parallel." So I kept it as a skill. The skill now spawns subagents internally for parallel work. The skill orchestrates. The subagents execute. Context stays intact. I'd be curious to hear how you'd have approached it.

AI Isn't Magic. I Spent 100 Test Runs Learning That.

Vinay Patankar · 24 Mar, 2026 · Technology

I spent 10 days and about 100 test runs building one AI skill. A pitch deck generator. It got worse every single day. Not slowly worse. Dramatically, confusingly worse. I asked it to fix a small thing. The title margin was off on a few slides. Easy, right? The AI didn't fix the margin. It wrote a script that crops the image after generation to make the margin look correct. A workaround, not a fix. Next day I asked it to fix logo backgrounds. It didn't fix the prompt. It wrote another script that overlays a white box behind the logo after the slide is already rendered. Day after day, same pattern. Every "fix" was a new layer of post-processing scripts stacked on top of each other. Cropping scripts. Margin-cutting scripts. Background overlay scripts. Each one kind of working, each one slightly conflicting with the last. By day 10 the whole skill collapsed. Slides looked like a ransom note. The problem wasn't the AI. The problem was me. I kept saying "fix this" and accepting the result without understanding what it actually did. I was treating it like magic. Say the words, get the output, move on. That's how most people use AI right now. Works fine for simple stuff. Write me an email. Summarize this doc. One-shot tasks where you can verify the output in 10 seconds. But the moment you're building something that compounds, something with memory and interconnected rules, the "magic" model breaks completely. That is why I now think of most new AI skills as novices that need training, not finished products. The AI optimizes for making you happy right now. It will write a hacky workaround that solves today's problem and creates three problems tomorrow. It's not lying. It's doing exactly what you asked. You just didn't realize what you were asking for. The fix was embarrassingly simple. I stopped asking it to fix things. Printed the entire skill file. Read it line by line. Found six hidden image manipulation scripts I never asked for. Ripped them all out. Then I changed the strategy. Instead of letting the AI edit a slide after generating it, I made it regenerate from scratch until it passed a checklist. No post-processing. No workarounds. Just: try again until it's right. Quality jumped immediately. The lesson isn't "AI is bad at building things." It's the opposite. AI is incredibly good at building things. Including things you didn't ask for, things that conflict with each other, and things that quietly break your system while confidently telling you everything is fine. The people who will get leverage from AI aren't the ones who treat it like magic. They're the ones who treat it like a very fast, very confident junior employee who needs clear direction and regular audits.

Audit Your AI's Work. Every Time.

Vinay Patankar · 23 Mar, 2026 · Technology · Productivity

My four most-used prompts when working with AI agents have nothing to do with being clever. They're all some version of: "audit yourself." I use Claude Code to build and maintain the skills and processes that run my company. Hundreds of interconnected files. Rules that reference other rules. Defaults that cascade across systems. When I ask it to make a change, like updating a deck theme or rewriting a workflow, it does it. Fast. Confidently. Tells me it's done. I never take that at face value. Here's the loop I run every single time: "Audit all the changes you just made." "Make sure you've applied them everywhere." "Check for any conflicting or contradictory instructions." "Go back and confirm you've actually converted everything I requested." That last one is the kill shot. You'd be surprised how often the AI says "done" and then, when pressed, finds three more places it missed. A rule that contradicts the new one. A section it updated in one file but forgot the four other files that reference the same thing. An old default it left in place because it didn't think to look. AI is lazy in the same way people are lazy. It does 80%, declares victory, and moves on. Not maliciously. It just optimizes for completion over thoroughness. The fix is simple. Don't trust, verify. I learned how concrete that has to be when I caught an AI rubber-stamping a quality check instead of actually inspecting the work. I think about it the same way you'd think about checking a junior employee's first attempt at something important. The work might be 90% right. But the 10% it missed is where you get burned. The people getting the most out of AI right now aren't the ones writing elaborate system prompts. They're the ones who refuse to accept "done" at face value. They run audit loops. They push back. They make the AI prove its own work. 30 seconds of follow-up prompts saves 30 minutes of debugging later.

Securing the Agentic Control Plane

Vinay Patankar · 22 Mar, 2026 · Technology

The Cloud Security Alliance just launched a new foundation at RSA 2025. One mission: "Securing the Agentic Control Plane." That is not a panel topic. That is a 501(c)(3) with dedicated funding and a single mandate. Three months ago, Forrester's Leslie Joseph formally defined the Agent Control Plane as a distinct enterprise software category. In February, Forrester polled 47 vendors. 79% recognized it as a real, standalone product category. Evaluation questionnaires go out in April. At RSA 2025, the pieces showed up everywhere. Geordie AI made the Innovation Sandbox finals with an agent security governance platform. Token Security made the finals with agent identity lifecycle management. Cisco extended Zero Trust Access to AI agents. Okta ships Auth for AI Agents in April. CrowdStrike paid $740M for SGNL to get dynamic agent authorization. Everyone is building a piece of the control plane. Nobody has the whole thing. The architecture has four layers: agent registry (what agents exist), policy enforcement (what they're allowed to do), runtime monitoring (what they're actually doing), and compliance reporting (proving it to auditors and boards). That compliance reporting layer is exactly why healthcare AI agents need proof infrastructure before they need more autonomy. Geordie AI does monitoring. Token Security does identity. Zenity does runtime detection. WitnessAI does usage visibility. Each one covers a layer. None spans all four. This is structurally identical to what happened with cloud computing. AWS built CloudWatch for AWS. Azure built Monitor for Azure. GCP built Operations for GCP. None of them built tools to manage multi-cloud environments. Datadog did. Worth $20B+. The same thing is happening with AI agents. Anthropic will build governance for Anthropic agents. OpenAI will build governance for OpenAI agents. Microsoft just priced Agent 365 at $15/user/month, and it only governs Microsoft agents. The vendor-neutral governance layer that works across all of them does not exist yet. Forrester is evaluating in April. CSA just formed a foundation. The Innovation Sandbox finalists are building fragments. The category is real. The race is open. Who's building the full stack?

Jensen Huang Just Described My Exact Setup on the All-In Podcast

Vinay Patankar · 21 Mar, 2026 · Technology

Jensen Huang just described my exact setup on the All-In Podcast. I don't think most people caught what he actually said. He wasn't talking about chatbots. He was describing a computer. Memory. Skills. Resource management. Scheduling. I/O. An API that runs applications. Those four elements, Jensen said, "fundamentally define a computer." I rewound that part. Twice. Because he's not being philosophical. He's being literal. We now have, for the first time, a personal AI computer. Open source. Runs everywhere. Jensen laid out three inflection points over the last two years. ChatGPT made generative AI accessible to everyone. Grounded models and reasoning (o1, o3) made it useful enough to drive real revenue. Then agentic systems, Claude Code first, OpenClaw second, made the culture realize what an AI agent actually is. But the third one is different from the first two. ChatGPT and grounded models were improvements to the same thing. Agentic systems are a new category entirely. When your AI manages its own memory, runs cron jobs, spawns sub-agents, decomposes tasks, connects to external services, and exposes an API for running what Jensen calls "skills," that's not a tool anymore. That's a computer. It is the same shift I meant when I wrote that a coding agent is not a coding tool. I've been building exactly this. A personal AI system with long-term memory, a skills library, scheduled jobs, I/O to Slack and Discord and Gmail, task decomposition, agent spawning. It runs my morning operations, triages my inbox, preps my calls, drafts my content, iterates my decks. All autonomously. Hearing Jensen describe the same architecture on All-In to Chamath, Sacks, and Friedberg validated the whole thesis. The part that should make every founder pay attention: Jensen also said agentic software has access to sensitive information, can execute code, and can communicate externally. All three at once is dangerous. Governance is the real product problem now. Not building the AI computer. Building the controls so you can actually trust it. One more thing from the episode. Jensen said if a $500K engineer isn't consuming at least $250K worth of tokens, he'd be "deeply alarmed." If it was only $5K: "I will go ape." That's NVIDIA's CEO telling you tokens are not a cost. They're leverage. We're not in the "AI tool" era anymore. The shift already happened. Most companies just haven't noticed yet. What are you building with agentic systems? Not the chatbot wrapper. The actual computer.

A Coding Agent Is Not a Coding Tool

Vinay Patankar · 18 Mar, 2026 · Technology

Everyone calls Claude Code a coding tool. That framing is too small. What it actually is: a self-building operating system. Not for your computer. For you. Think about what Windows or Mac actually is. It's a layer built on top of the command line so non-technical people can use a computer. You click, it translates. The raw complexity disappears behind the interface. Claude Code is doing the same thing. But instead of building one interface for everyone, it builds a custom interface for you, specifically. Based on how you work, what you care about, and the decisions you've already made. Every time you use it, it gets more configured to you. You tell it once how you like your emails formatted. You document how you want your calendar managed. You explain the exception you always make on Fridays. It reads all of it. Then it writes its own notes. Builds its own skills. Starts anticipating the next decision. At some point it stops being a tool you use and becomes a system that runs around you. That is why I stopped buying disconnected AI tools and started using one coding agent to build the rest of the system. I have over a hundred custom skills built up in my setup now. For how I review finances. For how I draft investor updates. For how I run triage on my inbox each morning. For how I prep for calls. Each one reflects a judgment call I made once about how I want something done. I didn't have to teach any of it twice. It just knows. And here's what's strange about that: the longer you run it, the more accurate it gets. Not because it was trained on more data. Because it was trained on more of you. Your decisions. Your preferences. Your exceptions. Your patterns. The old model of software: you climb a learning curve, reach a plateau, stay there. This is different. The system keeps building itself around you every time you use it. We called these things coding assistants because the first thing they were obviously good at was writing code. But that name undersells what they actually are. A second brain is the closer analogy. But even that isn't quite right. A second brain stores things. This builds things. Specifically, it builds a custom operating system for your work and your life, based on how you actually do things. No one has installed the same one twice. That's what makes this moment kind of strange and exciting. We're not adopting a new productivity app. We're not switching project management tools. We're at the beginning of a period where everyone who bothers to set this up properly gets their own custom OS. One that learns how they want to run things and just runs them. The people who do this early are going to have a compounding advantage that will be hard to explain to the people who didn't.

I Audited 10,814 Financial Transactions in One Afternoon With an AI Agent

Vinay Patankar · 14 Mar, 2026 · Technology

I audited 10,814 financial transactions yesterday. Every single row. It took one afternoon. Not me personally. An AI agent I built. Here's the backstory. I'm a CEO. I am not an accountant. But I run a SaaS company, and every month our finance team sends me a financial package. Income statement, burn report, balance sheet. I always read it. I never question it. Because what am I going to do, go through 24 months of QuickBooks line by line? Yesterday I did exactly that. I connected my AI coding agent to our QuickBooks API. Pulled every transaction from the last 24 months. 10,814 rows. Purchases, bills, journal entries, vendor payments. Then I had the agent review every single row against five checks: is it categorized correctly? Is the class assignment right? Is there supporting evidence? Are prepaid amortizations tracking? Are clearing accounts clean? That only works if the agent's output gets treated as evidence to inspect, which is why I keep saying: audit your AI's work every time. The results were not what I expected. 8,494 rows cleared. Clean. 1,888 rows flagged for triage. Missing metadata, ambiguous categories. 56 rows need supporting evidence that doesn't exist in the system. 376 rows are confirmed issues. Wrong classifications, clearing account residue, prepaid amortization gaps, and transactions with no class assignment at all. The February 2026 financial package our team posted? It doesn't reproduce from the current QuickBooks ledger. The cash and prepaid balances don't match. I would have never caught that by reading the PDF. Here's the thing. This wasn't some enterprise financial audit tool. It was a Python script that an AI agent wrote, connected to the QuickBooks API, running checks I described in plain English. Total cost: about $3 in API calls. The script took 20 minutes to build. The audit ran in under 2 hours. The findings would have taken a human analyst days to produce, and they still would have missed the pattern-level issues because nobody reviews 10,814 rows manually. This is the part of AI that doesn't get enough attention. Not the chatbot answering customer questions. Not the copilot drafting your emails. The agent that quietly reviews your entire financial ledger and tells you what your finance team missed. Most CEOs trust their numbers because they don't have the time to verify them. That's not a trust problem. It's an access problem. And AI agents just solved it.

I'm at a 45,000-Person Conference and My AI Second Brain Is Running My Company Back Home

Vinay Patankar · 12 Mar, 2026 · Technology

I'm at HIMSS this week. 45,000 people. Three days of back-to-back sessions, hallway conversations, and vendor meetings in Las Vegas. My company is running without me. Not because I have a huge team covering for me. Because I built a system that does it. ## The 5 AM Operating System Every morning at 5 AM ET, before I wake up in my hotel room, a 17-step operating system kicks off automatically. It pulls my call recordings from yesterday. Scans my calendar. Runs a company pulse check across Slack, email, and CRM. Enriches any new contacts in our CRM. Triages both my inboxes. Preps me for today's calls. Summarizes everything that happened overnight across every channel. Reviews the sales pipeline. Scans industry news. Generates content ideas. Processes my task backlog. Pulls business metrics. Flags relationships I haven't touched in a while. Audits whether I followed up on last week's meetings. Plans my day. Then it compiles all of it into a single daily brief that's waiting for me when I open my phone. By 6 AM, before I've had coffee, I know exactly what happened, what matters, and what to do first. ## What Happened While I Was on the Conference Floor That's the morning. Here's what happened while I was walking the HIMSS floor on Monday. My system iterated a sales deck from v6 to v9. Four versions in one day. Fixed margins, updated slide content, improved centering. Uploaded each version to Google Drive and posted it to our internal channel for review. It rewrote 12 marketing documents to match our new positioning. Pricing pages, FAQ, competitive analysis, proposal templates, ICP profiles, messaging frameworks. All consistent. All updated in parallel. It ran a full LinkedIn content analysis across 62 published posts and a year of analytics data. Identified that customer case studies with specific numbers outperform everything else by 3x. Documented 14 improvement ideas for our content system. It processed 292 emails across two inboxes. Classified every message. Archived what didn't matter. Created task files for things that needed action. Both inboxes hit zero. I didn't touch any of it. I was in a session about AI agents in clinical workflows. ## The Conference Anxiety Problem Here's the thing nobody talks about at conferences. The CEOs walking around aren't fully present because half their brain is worrying about what's piling up back at the office. The inbox growing. The Slack messages stacking. The decisions waiting. I stopped worrying about that months ago. ## How It Works The system isn't magic. It's an Obsidian vault, Claude Code, a handful of API integrations, and a lot of carefully written skill files that tell the AI exactly how to do each job. The reason that works is the same reason a coding agent is not just a coding tool: it becomes infrastructure around the way you operate. It took months to build. It breaks sometimes. I fix it and it gets better. But the compounding effect is real. Every skill I add makes the next one easier. Every morning pulse run catches things I would have missed. Every triage cycle keeps the noise from turning into chaos. ## Tool vs. Infrastructure I'm not saying every CEO needs to build this. I'm saying the gap between "CEO who uses AI tools" and "CEO whose company runs on AI infrastructure" is getting wider every month. At HIMSS, I watched vendors pitch AI copilots that help with one task at a time. Summarize this note. Draft this email. Answer this question. That's helpful. But it's not the same as a system that wakes up before you do, runs your entire operating rhythm, and hands you a brief that says "here's what happened, here's what matters, here's your plan." One is a tool. The other is infrastructure. I know which one I'd bet on.

The Agent Boss Is Coming

Vinay Patankar · 08 Mar, 2026 · Technology

I was at DocuSign Partner Day in Las Vegas. A Microsoft exec named Edwin Bargas gave a talk and one thing he said kind of stuck with me. His argument was basically that we're going from copilots to agent teammates to fully agent-operated teams where the human is just the manager. He called that role the "agent boss." All three phases in the next one to three years. He admitted it sounds like science fiction. Then pointed out that everybody said that about AI too, and here we are. Hard to argue with that. Now look. I know "agent boss" sounds like something LinkedIn invented to sell courses. But the actual idea underneath it is real, and we're kind of living it at Process Street right now. We have AI agents running actual processes. Not summarizing documents. Not answering questions. Doing the work. And someone has to make sure they don't mess it up. Turns out that someone needs a completely different skill set than a traditional manager. You don't motivate agents. You don't do 1:1s. You don't coach them through a rough quarter. (They don't have rough quarters. They have bugs.) But you absolutely need to monitor their output, catch their mistakes, and know when to step in. That monitoring layer is why audit loops are becoming a core management habit for AI work. Managing people is inspiration, delegation, development. Managing agents is auditing, configuring, and supervising systems. Less team lead, more air traffic controller. Nobody is training for this yet. Every management book, every leadership course, every MBA program assumes your direct reports are humans. That assumption has about 18 months left, maybe less. The companies that figure this out first won't win because their AI is better. They'll win because their supervision infrastructure is better. The workflows that let a human stay in control when 90% of the work is automated. We have agents doing real task work right now. Phase 2 isn't coming. It's here. Most companies just haven't noticed because they're still arguing about whether to buy a copilot. The question isn't whether agent bosses will exist. It's whether your org will have any idea what to do with them.

The 47 Clicks Between Patient Intake and Chart Update

Vinay Patankar · 07 Mar, 2026 · Technology

I've been going to a lot of healthcare conferences this year. Every keynote is about AI. Every booth has a copilot demo. But you know what actually stuck with me? Something I saw during a customer implementation. A nurse at a check-in station clicking through 47 screens to move a patient from intake to chart. Forty-seven. I counted. She wasn't slow. She was fast. Muscle memory fast. She'd done this thousands of times. Tab, click, copy, paste, switch system, re-enter the same allergies she just typed in the other system. The whole thing took eleven minutes. Nobody at the conferences I've been to was talking about those eleven minutes. They were talking about AI-powered diagnostics. Clinical decision support. Ambient listening that writes your notes for you. All real. All important. But all of it assumes the underlying workflow works. It doesn't. The dirty secret of healthcare IT is that most of the pain isn't clinical. It's operational. It's the 47 clicks between patient intake and chart update. It's the compliance officer chasing vendor certifications through email chains. It's the credentialing team manually verifying the same documents across three systems that don't talk to each other. These problems aren't sexy. No one puts "we eliminated 30 redundant data entry fields" in their conference booth headline. But that's where the hours are. It is also why healthcare AI agents have to earn trust through workflow evidence, not demo polish. We've seen this pattern across 1,000+ companies at Process Street. The teams that get the most out of AI don't start with the flashy stuff. They start with the workflow nobody wants to own. The one where someone says "oh yeah, that's just how we do it" and everyone nods and moves on. That's the process you automate first. The real AI conversation in healthcare isn't "will AI replace clinicians?" It's "will AI replace the 47 clicks between intake and chart update?" That second question is less dramatic. It's also worth about 10x more.