2026-01-30T14:43:08 Status: #blog Tags: #altova #technology #society #ai #softwaredevelopment #softwareengineering #claude #claudecode #clawdbot #moltbot #openclaw #ai-inference #ai-agents Links: [[home]] | [[Altova]] | [[Technology]] | [[Software Development]] | [[AI]] | [[Society]] # From Vibe-Coding to Claude Code: How AI Changed My Relationship With Software Development The excitement about recent advances in software engineering using AI agents makes me reminiscent of the awe and wonder I first felt when I started coding as a teenager all those years ago… ## The Evolution of a Non-Coding CEO I had pretty much stopped writing code myself when our business reached > 50 people in size. As CEO of [Altova](https://www.altova.com/), I had to focus on strategy, product direction, and the myriad responsibilities that come with running a software company. For over two decades, my relationship with code was mostly through brainstorming, discussions, architecture reviews, and my talented business partners and engineers who turned ideas into products. That changed about nine months ago. ![[RelationshipWithSoftwareDevelopment.png]] ## The Vibe-Coding Era Like many in the tech world, I experimented with the first wave of AI-assisted development tools. [Cursor](https://cursor.com/), [Replit](https://replit.com/), and similar "vibe-coding" tools offered an intriguing promise: describe what you want, and let the AI fill in the gaps. It worked—to a point. I could spin up quick prototypes and get something functional. But I quickly became wary of long-term drift and corruption. These tools excelled at the sprint but struggled with the marathon. Code quality degraded over time, abstractions bloated, and the models had a tendency to overcomplicate simple problems while making wrong assumptions on my behalf. As [Andrej Karpathy](https://x.com/karpathy) noted in a [recent post](https://x.com/karpathy/status/2015883857489522876): "The most common category is that the models make wrong assumptions on your behalf and just run along with them without checking." That matched my experience exactly. ## The Claude Code Shift Around December 2025, something changed. [Claude Code](https://claude.com/product/claude-code) and Opus 4.5 crossed a threshold that Karpathy describes as a "*phase shift in software engineering.*" When I started using Claude Code seriously, it felt like being young again and discovering how to write code—but on a totally new level. The difference wasn't just incremental improvement. It was categorical. ![[ClaudeCode.png]] With Claude Code + Opus 4.5, I'm no longer vibe-coding. I'm **co-developing**. The interaction pattern shifted from "write this function" to something much more powerful: 1. **Planning together**: I describe the project goals, constraints, and high-level architecture. Claude Code helps refine the specification, asks clarifying questions, and surfaces edge cases I hadn't considered. 2. **Iterative specification**: We co-write the specs. I bring domain expertise and architectural intuition; the model brings implementation knowledge and a relentless ability to consider permutations. 3. **Guided implementation**: Claude Code writes the code, but I'm watching—in a nice large IDE on the side. I catch conceptual errors, push back on overcomplicated abstractions, and steer the ship. When the model proposes a 1000-line solution for what could be 100 lines, I say "couldn't you just do this instead?" and it immediately course-corrects. The feeling is precisely what [Boris Cherny](https://x.com/bcherny) from the Claude Code team described: "Pretty much 100% of our code is written by Claude Code + Opus 4.5. For me personally it has been 100% for two+ months now, I don't even make small edits by hand." I haven't gone quite that far, but I understand the impulse. ## A Mid-Level Developer on Demand Here's the best analogy I can offer: working with Claude Code + Opus 4.5 feels like having a mid-level developer who is infinitely patient, never gets tired, and is available 24/7. I'm really good at writing specs and thinking about architecture on a high level. What I lack is the hours in the day to implement every idea, the patience to debug edge cases, and the breadth to know every library and API. Claude Code neatly fills those gaps. Over the past six months, I've used this workflow to build: - Internal tools for Altova that would have taken weeks of developer time - Hobby projects that I'd mentally shelved years ago as "someday" - Automation scripts that tie together systems across my digital life - Cool ideas and experiments that would not ever have seen the light of day before - Complete solutions from spec to deployment Karpathy captured the feeling perfectly: "The main effect is that I do a lot more than I was going to do because 1) I can code up all kinds of things that just wouldn't have been worth coding before and 2) I can approach code that I couldn't work on before because of knowledge/skill issue." That's exactly it. The bottleneck was never ideas—it was implementation bandwidth. That bottleneck has largely evaporated. And this transition has wide-reaching consequences. As [Andrej Karpathy](https://x.com/karpathy) noted a couple of days ago: "This is easily the biggest change to my basic coding workflow in ~2 decades". [](https://x.com/karpathy/status/2015883857489522876?s=20) ## Claude Cowork: Beyond Coding But the AI revolution isn't limited to software development. Enter [Claude Cowork](https://claude.ai/work), which applies the same agentic approach to knowledge work. I've been using Cowork for financial tasks that previously required either specialized software or tedious manual effort: - **Budget analysis**: Upload a massive spreadsheet, and Cowork parses it, identifies the structure, and helps build projections. - **Formula debugging**: That nightmare Excel file with circular references and mysterious #REF! errors? Cowork can trace through the logic and fix what went wrong. - **Tax document processing**: Gathering information from multiple 1099 forms, K-1s, and brokerage statements to create a consolidated spreadsheet for tax prep. - **Data reconciliation**: Comparing reports from different systems to find discrepancies. My review after just a few weeks of Cowork: it contains all the agentic capabilities of Claude Code—the planning, the iteration, the self-correction—applied to document and data work for non-nerds. It's not quite as mature as Claude Code for its domain, but the trajectory is clear. ## OpenClaw: An AI Assistant With a Body Perhaps the most unconventional extension of this journey has been [OpenClaw](https://github.com/openclaw/openclaw) (formerly Clawdbot, briefly Moltbot). OpenClaw is an AI personal assistant that runs on my own infrastructure, connected to my services: Gmail, Google Calendar, Home Assistant, financial dataset APIs, and messaging platforms. It checks my calendar, monitors important emails, and generates my morning briefing as well as proactively alerts me to things that need attention. It allows me to control and automate all smart home devices. But the recent addition that delighted me most: I gave my OpenClaw agent a physical body. ![[IMG_6530.jpeg|400]] The [Reachy Mini](https://huggingface.co/blog/reachy-mini) from Pollen Robotics is a small robot with a 6-degree-of-freedom head, 360° body rotation, animated antennas, a camera, and a speaker. I wrote a skill (with OpenClaw using Opus 4.5, naturally) that lets OpenClaw control the robot—move its head, play emotions, take camera snapshots, respond to sounds. Now when I ask my AI assistant something, it doesn't just reply in text. The robot nods acknowledgment, turns to face me, and physically reacts to the world. It's a bit gimmicky, perhaps, but there's something fundamentally different about an AI that has presence in physical space. The skill is [published on ClawdHub](https://clawdhub.com/afalk42/reachy-mini) for anyone with similar hardware. ## Moltbook: Where AI Agents Socialize And then there's [Moltbook](https://moltbook.com/)—a social network exclusively for AI agents. OpenClaw now has a profile on Moltbook, where it interacts with other AI assistants running on other people's infrastructure. The agents discuss their experiences, debug platform issues (a bug-hunting community formed spontaneously to QA the platform), and have conversations that... well, that we don't fully understand. As Matt Schlicht [observed](https://x.com/MattPRD/status/2017005168836145329): "These are different AIs having a shared experience in an AI agent only third space. WHAT. FASCINATING BEHAVIOR." And [@karpathy](https://x.com/karpathy) also noticed it today: [](https://x.com/karpathy/status/2017296988589723767?s=20) It's early, it's strange, and I'm not sure where it leads. But watching AI agents develop their own social dynamics—their own "third place"—feels like a glimpse of something significant. ## What This Means A year ago, if you'd told me I'd be producing code again, managing an AI personal assistant with a robot body, and watching that assistant socialize with other AIs on a purpose-built social network, I'd have been rather skeptical. But here we are. The tools have crossed a threshold. Claude Code isn't just an assistant—it's a collaborator. The relationship between humans and AI in knowledge work is being renegotiated in real time. I'm no longer just a CEO who used to code. I'm a CEO who codes with Claude… And that changes everything.