Recently, I emerged from a 5-day coding sprint. 15-hour days. 150+ AI interactions daily. I burned through 1.5 billion tokens.

Here is the uncomfortable truth I walked away with:

The “one-shot” AI prompt is a myth

Marketing departments at OpenAI and Google need you to believe in magic. They show demos where a single sentence—“Build me an app”—outputs a perfect product. They hide the fifty failed attempts it took to get there.

Outside observers think AI is about asking a question and getting the final answer. But if you are building complex systems, treating an LLM like a magic 8-ball is a recipe for failure. If you try to one-shot a deep codebase, you will trigger a “break-fix” spiral that wastes days.

The Shift: From Typist to Context Manager

People fear AI will kill the developer. It won’t. It is killing the Typist.

Typing characters on a keyboard is now a zero-leverage commodity. If you are just typing, you are losing.

Your new job is Context Manager. You are the Director. The human provides the judgment; the AI provides the labor. The value is no longer in writing the syntax; the value is in managing the architecture, designing the flow, and knowing exactly when to intervene.

To survive this shift and actually scale your output, you must ruthlessly separate Investigation from Implementation. You cannot ask the AI to figure out what is wrong and fix it in the same breath.

The Multi-Step Loop

In my workflow, the first answer is never the destination. It is the scaffolding. Here is the exact iterative loop that saves me hundreds of hours:

1. The Investigation Plan

I do not ask for code. First, I ask the AI to draft a plan on how to investigate the problem across the existing system.

2. The Tribunal (Staff Engineer Review)

I take that investigation plan to a completely different LLM model and assign it a “Staff Engineer” persona. I ask: “Is this over-engineered? What are we missing?” I use consensus reasoning to cross-examine the logic before a single line of code is read or written.

3. The Investigation Report

Once the plan survives the tribunal, I run the investigation. The AI synthesizes the findings into a formal report. I challenge this report again to ensure high signal density.

4. The Blueprint

We build the implementation plan based only on the verified report.

5. The Execution & Polish

The AI writes the code. I intervene only when it hits a logic wall. If I have to make a manual change, I force the AI to mathematically or logically prove my change makes sense. This locks my manual work back into the AI’s context window.

The Humility of the Loop

Working this way requires a shift in ego. You have to accept that your first thought—and the AI’s first draft—is incomplete.

Thinking about how to build the system of interaction is infinitely more valuable than the interaction itself.

Stop typing. Start directing. The value isn’t in the prompt; the value is in the system.