How an Ex‑Startup Founder Leverages AI Agents to Build 40 Projects in One Year

Background

Peter Steinberger founded PSPDFKit, operated the company for 13 years, and then sold it. After a period of burnout he returned to software development in 2024, using AI tools to create more than 40 projects within a year. He combined these projects into OpenClaw, a personal AI agent, before joining OpenAI.

AI‑Driven Development Workflow

To revive an unfinished project, he exported the codebase as a Markdown file, used Gemini to generate a formal specification, and then fed that spec to Claude Code, executing a build command. The first build produced low‑quality code, but the end‑to‑end process demonstrated that AI can now generate complete software artifacts.

In a voice‑message experiment, the model automatically detected the audio encoding, invoked ffmpeg to convert the file, retrieved the OpenAI API key from the environment, called the transcription endpoint, and returned the transcript. This showed that the model can orchestrate external tools and APIs without pre‑written glue code.

When running inside a minimal Docker container that lacked curl, the model wrote a tiny HTTP client in C, compiled it with the available compiler, and used the binary to fetch a webpage. The example illustrates on‑the‑fly creation of tooling when required utilities are missing.

Best Practices for AI‑Assisted Coding

View AI assistance as a learnable skill rather than a “vibe coding” shortcut.

Always ask the model “Do you have any questions?” to surface ambiguities before it makes assumptions based on its training data.

Avoid the “agentic trap” of over‑optimising tool‑chain settings that feel efficient but add unnecessary overhead.

Keep the workflow minimal: converse with the model, avoid complex worktrees, and maintain multiple checkouts only when needed.

Perspective on Code Value and Review

Most code is essentially a transformation of data shapes and is therefore interchangeable.

During pull‑request review, focus on the intended behavior rather than line‑by‑line differences, treating the change as a “prompt request”.

Instead of reading generated code line by line, monitor the program’s output stream to verify that it matches the expected mental model.

This approach requires a comprehensive understanding of the overall system architecture, which the model itself does not possess.

Product Insight

A simple signal of product‑market fit is unsolicited demand from peers for features you never planned to build.

Hands‑on experimentation is essential; reading tech news alone does not provide deep technical understanding.

Security Considerations

Prompt‑injection attacks remain unsolved, though newer models exhibit better resistance than expected.

Open‑source projects cannot fully prevent unintended misuse; mitigation focuses on reducing self‑harm.

Introducing dedicated security expertise marks a shift from ignoring security to proactive risk management.

Evaluation of Coding Models

Peter tested many AI coding assistants and selected OpenAI Codex for its high reliability and immediate usability. He highlighted GPT‑5.2 as a “quantum leap” in capability. While Claude Opus offers broader functionality, Codex excels at programming‑specific tasks. His GitHub contribution graph shows a noticeable activity spike in October–November after adopting Codex.

Advice for Developers

Approach AI projects experimentally; pick a personal idea and iterate with AI assistance.

AI will not replace developers in the short term, but those who leverage AI effectively will gain a competitive edge.

Developers whose core identity is creation and problem‑solving will see increased value.