AI Native Installer
A friend asked if they could have a blog like mine. AWS Hosted and only about $1/month. Normally I’d give him a template repo and a detailed README. I’d walk him through the setup over a beer. But AWS stuff is too complex for the uninitiated.
Recently I was inspired by nanoclaw
git clone https://github.com/gavrielc/nanoclaw && cd nanoclaw && claude
Notice the && claude at the end. Instead of shipping with installer code, Nanoclaw comes with a CLAUDE.md that tells Claude Code how to step the user through configuration. And instead of having plugins for integrations, Nanoclaw comes with AI skills explaining how to make custom integrations.
It doesn’t have a complex README. Instead it has an AI built-in to hand-hold the user. So I asked Claude to turn my blog stack (Astro, S3, CloudFront, Pulumi, GitHub Actions) into a template, with a CLAUDE.md and a .claude/command/setup.md.
git clone https://github.com/lukerohde/aws-quill my-blog && cd my-blog && claude
Claude reads the CLAUDE.md, sees it’s a fresh clone, and offers to run /setup. From there it’s a conversation. It checks your dependencies, offers to install what’s missing, buys a domain through the AWS Route53 API if you need one, deploys two Pulumi stacks (ingress and blog), creates the GitHub repo, sets the CI secret, and pushes the first commit.
The interesting part is when something goes wrong, Claude finds a way around the problem. When the user needs to do something manually (like delegate nameservers at their registrar), Claude gives them the specific instructions for their specific registrar, then waits for confirmation.
It’s the difference between handing someone a recipe and cooking alongside them. The AI takes my place and sadly I miss the beer. Still pretty cool.
I built and tested this tonight, was struck by what it could means for platform teams.
Platform teams spend years building internal developer platforms with infrastructure patterns, starter kits, golden paths. The goal is to enable product squads to ship software safe and fast without reinventing wheel and making security mistakes.
The difficulty is the finding the right deployment boundaries and right degree of modularity for developers. Too prescriptive and a pattern doesn’t fit anyone. Too many building-blocks, and there are too many difficult and dangerous loose wires to connect.
An AI-guided template knows exactly what state you’re in. It can read your existing config and fill in sensible defaults. It can refuse to do things that would break the security model — not because you programmed every failure mode, but because the AI understands the intent. Sorta.
The guardrails are in the CLAUDE.md. You ship the platform as a template, with an AI guide, and the AI handles the installation and customisation. Far fewer support tickets, and standard patterns keep things maintainable.
Platform teams have always known that the hard part isn’t building the platform. It’s upskilling devs in cloud skills and DevOps. Does an AI assistant help or disable developers from learning and growing?
The big hurdle will be getting agentic AI approved for internal use. We need to think real hard about how much permission we give to an agentic AI. I can’t stop my agentic AI from finding and sending secrets to Anthropic. At one point, it jumped repos, and pushed a change to my IAM policy to give itself permissions to deploy dynamodb. Grr…