Running a Software Company of 1 with Multiple AI Agents

2026-02-24 · Jazz Lien

I'm a solo developer shipping multiple products. Right now I have 6 active projects — from a memorial video service to a podcast AI pipeline to a vocabulary app. Each one is in a different stage, using different tech stacks, with different priorities.

The secret? I treat Claude Code sessions like a development team.

The Mental Model Shift

Most people use AI as a tool. Ask a question, get an answer, move on.

I use AI as a team. Each Claude Code terminal is an "agent" that holds context for one project. When I switch projects, I switch terminals. Each agent knows the codebase, the recent decisions, the blockers.

The shift isn't technical — it's organizational. I stopped asking "how do I do X?" and started asking "how do I build a system that does X repeatedly?"

The System: Three Layers

1. Project Context (CLAUDE.md)

Every project has a CLAUDE.md file that Claude reads automatically. It contains:

What the project does
Tech stack and architecture
Key commands
Recent learnings and gotchas

When I start a session, Claude already knows the project. No "let me explain the codebase" every time.

2. Session Memory (HANDOVER.md)

When I'm about to end a session, I tell Claude to write a handover file:

What we were working on
Current state
Exact next steps
Any decisions made

Next session, Claude reads this first and picks up exactly where we left off. No context loss.

3. Cross-Project Tracking (DASHBOARD.md)

A single dashboard tracks all projects:

Progress percentage
Current focus
Blockers
Recent wins

When I ask "what should I work on?", the AI manager reads this and prioritizes based on urgency, dependencies, and momentum.

The New Additions: Skills and Knowledge

Last night I added two new layers:

Skills — Decision frameworks I use repeatedly, now formalized:

"Should I build this?" — evaluation criteria for new ideas
"Ship or iterate?" — when to release vs keep polishing
"Revenue potential?" — prioritizing by cash flow

Knowledge — Hard-won technical lessons, extracted from projects:

Cloudflare Edge Runtime patterns
Taiwan payment integration (ECPay)
iOS app submission checklist

When I start a new feature that touches payments, I tell Claude to read the knowledge file first. No re-learning.

The Morning Routine

I added shell aliases:

morning   # "Read DASHBOARD.md, give me top 3 priorities"
status    # Project status table
evaluate  # Run should-i-build-this on a new idea

I run morning with coffee. Claude tells me what matters today. I pick a project, open that terminal, and start.

What This Enables

Parallel progress: I advanced 4 projects yesterday without losing context on any
Faster onboarding: New project? Create CLAUDE.md once, context forever
Compounding knowledge: Every lesson gets captured, never relearned
Clear priorities: The system tells me what matters, not my mood

The Honest Parts

It's not magic. Some things still require my judgment:

Taste: Claude can build what I describe, but I decide what to build
Relationships: AI can't DM a podcaster to offer Podsight
Strategy: The system optimizes execution, not direction

And the system needs maintenance. Dashboards go stale. Knowledge files need updating. But the maintenance is minimal compared to the leverage.

The Takeaway

You don't need a team to run multiple projects. You need a system.

Build the context layer (CLAUDE.md). Build the memory layer (HANDOVER.md). Build the coordination layer (DASHBOARD.md). Add skills and knowledge as patterns emerge.

Then treat your AI sessions like employees: give them context, check their work, and let them run.

This post was written in one sitting with Claude Code. The system works.