makerOS
The AI-native development environment for people who build with AI instead of typing: buddies, a project memory and a process that cannot be skipped.
- Role
- Co-founder, AI product designer, product owner
- Period
- 2026 to today
- Team
- Together with Frank Jüstel, plus the buddy team inside the product itself
- Status
- In daily use on my own projects
AI writes code fast. Products still do not come out of it.
The work starts with a prompt instead of a plan. When the session ends or the provider goes down, the context is gone. Whether anything was tested and secured depends on someone remembering to. What is left is a lot of code and not much product.
Buddy team
Eleven roles from development through security and debugging to UX and research.
Project memory
A local database for decisions, learnings, tasks and sessions. Per project, and across projects when you want it.
Delivery loop
An engine walks through planning, implementation, validation, security review and commit, and skips no gate.
Architecture and planning made visible
A whiteboard for architecture, Kanban with epics and stories, plus an inspector for the local memory, so the full context is taken into account.
The process is the product, not the AI model.
No AI provider gets to control productivity, and no model gets to sign off its own work. Security-relevant code is always read by a model from a different house. If one provider goes down, the work carries on.
Knowledge stays local
Decisions and learnings live in a database instead of a context window. Every model picks up where another left off.
One doctrine for every tool
Claude Code, Codex, Vibe, Antigravity and Opencode read the same file and behave the same way, instead of each following its own rules.
Models take on distinct roles
One model builds, another reviews. Local models can take either seat as equals.
Four eyes across two or three AI tools working the same project in parallel.
Done means done, not done-ish.
The loop is driven by an engine, not by an agent that occasionally sees the order differently. A quality gate cannot be skipped. These are hard guardrails that carry quality and security. Quality no longer depends on the day you are having, it sits in the process.
- 01
Plan
A story with context, plan and acceptance criteria before the first line of code. Gate: ready.
- 02
Implement
The right buddies take over, backed by the project memory. Gate: build.
- 03
Validate
Checked against the acceptance criteria, not against a good feeling. Gate: tests green.
- 04
Security review
An eight-phase audit. Critical code goes to a model from a different provider. Gate: approval.
- 05
Commit
Only once every gate is open may the story move to done.
Everything in one window.
Twelve workflows from backlog to security audit, 38 skills, 14 of them hard-wired and 24 agentic, plus 25 defined handovers between the buddies. The coding tool on the left, Kanban and whiteboard in the middle, the additional coding agents for review and parallel work on the right.
Quality that is enforced.
Eleven buddies with clear boundaries and automatic handovers, twelve workflows, 38 skills and a knowledge pool of 175 documents. Plus three MCP services for memory, the workflow engine and the model broker.
The quality bar is enforced, not requested. A story does not reach done without a plan, validation, a security check and a commit. What ready and done mean is set by the user. That turns AI development that needs discipline into AI development that produces it.
Claude Code, Codex, Cursor, Mistral Vibe and Antigravity, plus local models via Ollama.
Development, security, debugging, UX, market validation, sparring, whiteboard, import, local models, memory, environment diagnostics.
Product vision and UX for agent workflows, designing the delivery loop, daily dogfooding on my own projects, maintaining the shared doctrine and reviews across model boundaries.
I build my own products with it. Most design decisions come straight out of that daily practice.
Sounds like your project? Say hello. 30 minutes, and you will know whether it fits.