I started using Claude Skills when Anthropic launched the feature last year in Oct 2025. For context, Skills let you teach Claude specific workflows: how to query your CRM, how to format a pitch deck triage, how to write in your voice. You build these by conversation, not code. Within a few weeks, I had built over a dozen custom skills that turned a generic chatbot into something that felt like a personalised operating system for my work.
That experience gave me a glimpse of the future. The trajectory is clear: we are moving from generic chatbots to personalised AI agents. And the implications for software, and for how we invest, are significant.
Post-GPT, It’s About Design and Intent
Before GPT, software was defined by features. You bought Salesforce for its CRM features, Airtable for its flexible database, Slack for its messaging. The value was in the tool itself.
Post-GPT, the game has shifted. Software is increasingly about design and intent, about what you want to accomplish, not which buttons to press. In a post-GPT world, design becomes more essential. Not design in the aesthetic sense, but design as the articulation of what you actually want. The clarity of intent becomes the limiting factor to our productivity. The bottleneck is no longer “can the software do this?” It’s “can I describe precisely what I need?”
The interface layer is collapsing. When I can tell an AI agent “prepare my LP meeting brief for tomorrow” and it pulls from my calendar, CRM, email, and portfolio data to produce a coherent document, the traditional notion of a “software product” starts to blur. The quality of the output depends entirely on how well I’ve designed the request: the skill, the prompt, the workflow. The tool recedes; the intent takes centre stage.
I believe software is going to be hyper-personalised. Not in the way we’ve used that word before, like personalised recommendations or personalised dashboards. I mean, fundamentally rebuilt around individual workflows and preferences. The software I use daily is increasingly shaped by my own prompts, skills, and automations rather than by a product team’s roadmap.
OpenClaw: The Extreme Case
OpenClaw (formerly Clawdbot) takes this idea to an extreme. It’s an open-source AI agent runtime that runs locally on your machine, connects to your messaging platforms, and executes tasks with direct access to your file system, calendar, and terminal. Think of it as a local digital butler that never sleeps.
I’ve been running OpenClaw for my family through Telegram as the primary interface. Every morning, it sends me a daily brief, a synthesis of my calendar, emails, and action items that helps me orientate my day before I even open my laptop. It handles my expense tracking, capturing receipts and categorising spend on the fly. But the use case that surprised me most is the household agent. My wife and I use it to capture and manage our kids’ schedules, school assignments, medical appointments, and enrichment classes. It de-conflicts our family calendar, flags clashes, and reminds us of what’s coming. All of this happens through a Telegram chat. No app to open, no dashboard to check. Just a conversation.
What makes OpenClaw interesting isn’t just the technology, it’s the architecture. You can swap in different AI models as the “brain.” Recently, they integrated Moonshot AI’s Kimi K2.5, which means you can run a capable AI agent entirely for free using an open-source runtime and an open-weight model. The setup takes minutes.
This is a generational shift in how software gets built and consumed. The agent runtime is open. The model is interchangeable. The personalisation, meaning your skills, your workflows, your data, is yours.
Workflows Are the Moat
So does this mean SaaS is dead? The market seems to think so. When Anthropic launched Claude Cowork’s industry-specific plugins last week, automating workflows across legal, finance, sales, and marketing, it triggered a sell-off that erased roughly $285 billion in market capitalisation. Thomson Reuters dropped over 18%. Salesforce, SAP, Oracle, and ServiceNow all fell sharply. A JPMorgan index for US software stocks dropped 7% in a single day. Commentators started calling it the “SaaSpocalypse.”
I don’t think it’s that simple. SaaS businesses have three durable sources of value: workflow orchestration, proprietary data, and user trust. A well-designed vertical SaaS product doesn’t just store your data. It encodes years of accumulated best practices about how work gets done in a specific industry. That’s hard to replicate with a prompt.
But here’s the nuance: SaaS businesses might not be as valuable as we think. If AI agents can increasingly orchestrate workflows across multiple tools, the bundling premium that SaaS companies charge starts to erode. The moat isn’t the software, it’s the workflow. And workflows are becoming more portable.
For founders building SaaS today, the defensible position is becoming less about feature completeness and more about being the system of record that AI agents need to interact with. If your product holds the canonical data and the trusted workflow, you survive. If you’re just a nice UI on top of commodity logic, you’re vulnerable.
A Personal Case Study
To make this concrete: I recently migrated my personal CRM from Airtable to Supabase with the help of Claude Code. The CRM tracks about 1,400 contacts across people, companies, and locations: founders, LPs, GPs, and ecosystem stakeholders I’ve built relationships with over the years.
The migration involved designing a PostgreSQL schema, exporting and transforming all the data, and rewriting my Claude skills to query Supabase instead of Airtable. Claude Code handled the heavy lifting. I directed; it executed.
The result? Net savings of about $20 per month with no vendor lock-in. My data now lives in an open-standard PostgreSQL database that I fully control. I can query it with SQL, connect any tool to it, and switch providers at will. The “product” I’m using is a combination of an open-source database, an AI coding assistant, and my own workflow logic. No single vendor owns the experience.
This is a small example, but it illustrates a larger point. The cost of building and maintaining bespoke software has collapsed. What used to require a developer now requires a conversation.
Open Source as a Disruptive Force
OpenClaw running on Kimi models. A personal CRM on Supabase built by an AI coding agent. These are not isolated experiments. They’re early signals of a structural shift.
Open source is the disruptive force that makes hyper-personalised software possible. When the agent runtime is open, when the models are open or near-free, and when the database is open, the value doesn’t accrue to any single application layer. It accrues to infrastructure: hosting providers, database providers, network access providers, and compute providers. The picks and shovels, as always.
For us as investors, this reframes the question. We should be asking: where does value concentrate when the application layer becomes infinitely customisable? The answer, I believe, is in the infrastructure that makes it all run, and in the vertical-specific workflows and data that remain hard to replicate even in an AI-native world.
I’m still early in this tinkering journey. But the trajectory feels clear, and it’s moving faster than most people realise. The future isn’t one product for everyone. It’s everyone’s product, built by AI, running on open infrastructure.
Let’s see how this plays out.
Leave a Reply