How a Backup Problem Became a Methodology

• The original problem Backup traffic and posture across the environment were hard to see and harder to reason about. The data existed, but not in a form anyone could act on.
• Early attempts with ChatGPT and Grok The first passes produced plausible code that fell apart on contact with the real API. Generalist prompting could not hold the whole problem.
• The shift to Claude and what changed Claude held more of the problem at once and reasoned about the API surface instead of guessing at it. The work started to compound instead of reset.
• Hitting context-window and complexity limits Asking for the whole system in one breath broke things. Context filled up, dependencies got skipped, and quality dropped at the edges.
• Discovering the Atomic Build Method Breaking the build into atomic blocks fixed it. One verifiable task at a time. Task B did not start until Task A worked.
• Calling focused expert agents Instead of one generalist, the work got routed to focused expert agents, each responsible for a narrow, checkable slice.
• The Socratic method The turning point was asking the model what it needed to succeed instead of guessing the prompt. The model surfaced the missing context itself.
• The working result An API-driven platform: Python pulling from the Rubrik GraphQL API, with an LLM summary layer turning raw posture data into something a human could decide on.

Break the system into atomic blocks. Complete one. Validate it. Move to the next.