There is a structural problem with how most developers use AI coding tools that has nothing to do with which tool they use or how they phrase their prompts.
The problem shows up the same way across different tools, different projects, and different developers: AI output that starts out useful and gets progressively less reliable as the project grows. Code that is internally inconsistent. Architecture that looks sound in isolation but does not fit the rest of the system. Decisions made in session three that contradict session one; neither session aware of the other.
The common explanation is that AI models hallucinate, or that they have limited context windows, or that they are not reliable enough for serious work. These are not wrong. But they describe symptoms rather than the actual problem, which is a workflow problem.
What actually goes wrong
Every professional software practice that involves multiple sessions of work has developed discipline around continuity. Version control keeps code history coherent. Code review catches decisions that do not fit the larger system. Testing verifies behavior across changes. Documentation ensures that decisions made in the past are available to work being done in the present.
AI-assisted development has none of this by default. Each session starts approximately fresh. The model knows what you tell it in that session. If the project has been running for two weeks across thirty sessions, the model in session thirty-one knows nothing about sessions one through thirty unless you reconstruct that context yourself; and most developers do not, because no one has told them they need to.
Context collapse is the name for what happens when this gap between the project’s actual state and the model’s working knowledge grows large enough to produce wrong output. It is not a failure of the model. It is a failure of the workflow around the model.
What the operational method addresses
Prompt Before The Prompt is built around a simple premise: the work that makes AI-assisted development reliable happens before the first prompt in any session, not inside it.
That work includes establishing a shared context document that the model can reference, defining the scope of what the current session will and will not address, identifying the decision checkpoints where human verification is required before the work continues, and structuring the handoff at the end of a session so the next one starts with a coherent foundation.
None of this is complicated. The discipline exists in other parts of software development in more sophisticated forms. Applied to AI-assisted work, it is largely a matter of establishing habits that most developers have not formed because they were not obvious from the outside and no one has documented them clearly.
The guide documents them. The free preview covers the core of the problem and the foundational principles. The full Guide Edition covers the complete method.