DEV Community

Docat
Docat

Posted on

I Built an AI-Powered Solo Dev System — Here's the Exact Toolkit I Use Daily

#ai #productivity #claudecode #indiehacker

I shipped 3 products in 2 weeks. My team? Just me and Claude Code.

No co-founder. No freelancers. No Fiverr gigs. One human, one AI agent, and a system that makes the whole thing work without falling apart.

Here's the thing nobody tells you about building with AI agents: the AI is not the bottleneck. You are. Specifically, how you communicate with the AI and how you organize what it produces. I learned that the hard way.

Let me walk you through what I built, what almost killed my momentum, and the exact toolkit I use every day now.

The AI-first solo dev workflow

My setup is simple. I use Claude Code as my primary development partner. Not as an autocomplete tool, not as a chatbot I ask questions to -- as an actual agent that writes code, runs tests, commits, and pushes to GitHub.

Here's what a typical day looks like:

  1. I open my Notion dashboard, pick the highest-priority task
  2. I describe the task to Claude Code in my terminal
  3. Claude reads my CLAUDE.md config file, understands the project context, and starts building
  4. It writes TypeScript, runs Vitest tests, fixes failures, and commits with Conventional Commits
  5. I review the output, maybe test the product, and move on

In 2 weeks, this workflow produced:

  • mcp-openapi -- an npm package that converts any OpenAPI/Swagger spec into MCP tools (357 weekly downloads and growing)
  • graphql-to-mcp -- same idea for GraphQL APIs, with auto-introspection
  • FreeTools -- a static web platform with 6 free tools (invoice generator, JSON formatter, QR codes, etc.)

Plus 3 Dev.to articles, directory submissions, and all the marketing materials. The AI agent did about 90% of the actual work. I did account registrations and product testing.

But it was not always this smooth.

The two problems that almost killed my productivity

Problem 1: Death by context-setting

The first few days, I was burning 10-20 prompts per session just explaining things Claude already should have known:

  • "We use pnpm, not npm"
  • "Files go in packages/, not the root"
  • "Use Biome, not ESLint"
  • "Naming convention is kebab-case for files"
  • "We deploy to Cloudflare Workers"

Every. Single. Session.

That is 10 minutes of typing conventions before any real work starts. Multiply that by 5-6 sessions a day. I was spending an hour daily on what is essentially a recurring README.

And it gets worse. Without these instructions, the AI would confidently do things wrong. It would create files in the wrong directory, use the wrong test runner, pick the wrong naming convention. Then I would spend even more time fixing it.

Problem 2: Chaos at scale

By day 4, I had two npm packages, a web platform, three articles in various states, directory submissions to track, and no idea what was done, what was blocked, and what needed my attention vs. what the AI could handle alone.

I tried keeping it in my head. I tried a text file. I tried random Notion pages. Nothing stuck because none of it was designed for this specific workflow -- a human + AI agent collaboration where the AI does most of the work and you need to track what to delegate, what is blocked on you, and what the AI decided on its own.

How I solved both problems

Solution 1: CLAUDE.md config files

CLAUDE.md is a file that Claude Code reads automatically at the start of every session. It is your project's instruction manual for the AI.

Here is a simplified version of what mine looks like:

# Project Config

## Architecture
pnpm workspace monorepo. Code in `packages/`, docs in `docs/`.

## Conventions
- Language: TypeScript (strict mode)
- Runtime: Cloudflare Workers (primary), Node.js (CLI)
- Package manager: pnpm
- Linter/Formatter: Biome
- Testing: Vitest
- Naming: kebab-case files, camelCase vars, PascalCase types

## Folder Rules
- All source code -> `packages/`
- All business docs -> `biz/`
- No content files in root

## Deployment
- npm publish: agent can execute directly
- Production deploy: agent can execute (authorized)
Enter fullscreen mode Exit fullscreen mode

With this in place, Claude Code starts every session already knowing the full context. Zero warm-up prompts. It just works.

The difference is night and day. Before CLAUDE.md, I was averaging maybe 3 productive AI interactions per session before context drift made things messy. After, I am getting 15-20 high-quality interactions where the AI stays on-convention the entire time.

I wrote configs for every stack I work with. Next.js, Python, Rust, Go, Terraform -- each one tuned to the conventions and patterns that matter for that ecosystem. Things like:

  • Which state management library to default to
  • How to structure API routes
  • Error handling patterns
  • What NOT to do (equally important)

The configs save roughly 40 minutes per day. Over a month, that is 20 hours of recovered productivity. For a solo dev, 20 hours is the difference between shipping and stalling.

I packaged all 12 of these configs into a collection: CLAUDE.md Mega Collection. Twelve production-ready config files for $12. Each one is battle-tested from actual projects, not theoretical templates.

Solution 2: A Notion system built for human + AI collaboration

The second problem required a different kind of solution. I needed a system that tracks:

  • What products exist and what stage they are in (idea, building, launched, growing)
  • What the AI agent is working on vs. what is blocked on me
  • Content pipeline -- what articles are drafted, published, performing well
  • Revenue -- what is making money, what is not, and where to focus
  • Distribution -- which directories, marketplaces, and channels each product is submitted to

I built this as 6 interconnected Notion databases:

  1. Product Tracker -- every product from idea to revenue, with status, priority, and links
  2. AI Task Board -- tasks split by "Agent does this" vs. "Human does this" with clear handoff points
  3. Content Calendar -- Dev.to articles, Twitter threads, Reddit posts, all mapped to products
  4. Revenue Dashboard -- income by product, by channel, tracked monthly
  5. Freelance Pipeline -- for when I take on occasional client work to fund the operation
  6. Distribution Checklist -- every marketplace and directory, with submission status per product

The key insight: this is not a generic project management system. It is specifically designed for the workflow where an AI agent does 80-90% of the work and you need to manage the 10-20% that requires a human.

For example, my AI Task Board has a "Blocked on Human" view that shows me exactly what needs my attention. Right now it shows things like "Register Lemon Squeezy account" and "Submit to Glama.ai directory." Everything else, the AI handles autonomously.

I use this system every single day. It is the first thing I open in the morning and the last thing I update at night.

I have made this available as well: AI Solopreneur Toolkit. Six interconnected Notion databases for $9. It is the exact system I use, not a theoretical framework.

Results so far

Two weeks in, here is where things stand:

  • 3 products shipped (2 npm packages, 1 web platform)
  • 357 weekly npm downloads on the main package, growing week over week
  • 4 Dev.to articles published (one still in draft)
  • 38 passing tests on the GraphQL package alone
  • $0 infrastructure cost -- everything runs on free tiers (Cloudflare Workers, GitHub Pages, npm)
  • Daily AI productivity: 15-20 high-quality coding sessions with zero context-setting overhead

The constraint is not the AI's capability. It is the systems around it. Give Claude Code a well-written CLAUDE.md and it operates like a senior developer who has been on your team for months. Give it nothing and you spend half your time being a human linter.

What's next

I am working on a few things:

  • Monetization: Setting up Lemon Squeezy for the npm packages (free tier + paid Pro features)
  • Marketing automation: Using the Dev.to API and Twitter API to automate content distribution
  • More products: The AI agent is already proposing new product ideas based on what is performing well

The goal is $300-1,500/month in passive income within 6 months. All built and maintained by an AI agent, with me doing maybe 30 minutes of human work per day.

If you are a solo developer experimenting with AI agents, the two things I would recommend investing time in first:

  1. Write a CLAUDE.md for every project. Even a basic one saves you 10 minutes per session. A thorough one changes how you work entirely.
  2. Build a tracking system designed for AI collaboration. Generic to-do apps do not cut it when half your "team" is an AI that operates autonomously.

These are the two leverage points that made the whole system work for me. Everything else -- the coding, the testing, the marketing -- flows from getting these two foundations right.

If you want a head start:

Both are things I built because I needed them. They just happen to be useful to other people too.

Happy building.

Top comments (0)