Prompting is not a development methodology.
Writing about agent workflows, software architecture, and the mental models needed to build with LLM systems beyond prompting for output.
A representative piece on orchestration, failure modes, and why building with LLMs needs more than prompting.
LLM pipelines rarely fail in a single call. They degrade across turns, handoffs, and hidden state. This piece argues that the right way to test them is not by replaying golden cases, but by simulating messy users and watching where the system starts to break.
One stress-tests LLM pipelines. One structures software before code.
An open-source harness for simulating, inspecting, and stress-testing multi-turn LLM pipelines.
A system for turning vague software ideas into clearer requirements, decisions, and handoffs before implementation starts.