Reset Before You Regret
The reset technique that saves AI collaborations from context rot
The Problem
Context rot is the silent killer of AI collaboration. You start with a clear vision, but twenty messages later, your AI assistant is building features you never asked for, forgetting core requirements, and contradicting decisions made ten messages ago. The model gets lost in the conversation's accumulated complexity, missing the forest for the trees.
This degradation happens predictably. Research shows that as conversations grow, LLM performance drops measurably. They miss relevant details, get confused by ambiguity, and lose track of what actually matters. It's like watching a game of telephone played by a single participant who gradually forgets the original message.
Traditionally, we combat scope creep and requirement drift with documentation in traditional software development. The same principle applies to AI collaboration, but the execution must adapt to how these models fail. A recent video by Chroma wonderfully describes the context rot problem statement.
The Technique
Before building anything beyond a simple task, co-create a PRD or implementation plan with your AI. Define phases, steps, and completion criteria together. When context rot sets in, restart the conversation with this document as your North Star.
Tasks (bug fixes, enhancements, shell scripts) might survive without formal documentation, but everything else needs this shared contract.
As the implementation proceeds, keep this document updated, including a progress log, and have the AI verify that the completion criteria are met.
The Discovery
The documentation becomes your reset button. When conversations spiral, you can start fresh with just the PRD, eliminating 90% of the noise that caused the rot. The AI reads your agreed-upon spec, not the maze of changes that led you astray.
Co-creation prevents misalignment before it happens. When you force the AI to articulate phases and define "done," you catch misunderstandings early. Last week, an AI's proposed "simple API" included OAuth, rate limiting, and webhooks. I asked for a phased breakdown, which revealed we had very different definitions of "simple." Sometimes AI can get it right the first time, but the few extra minutes to nail the details will save you hours of frustration. Most times, you’ll find that it can’t read your mind, and you certainly can read its mind, so write it down.
The technique works because it mirrors how LLMs perform. Studies show that models answer better with condensed, relevant contexts than with a full conversation history. Your PRD becomes that optimal context, focused, unambiguous, and free from the distractors that accumulate naturally.
You're not just documenting for the AI but for yourself. The exercise forces clarity about what you're building. You don't need to write it down yourself; you can use AI to generate the document with your guidance. The key is to bless the contents so you and the AI remain focused on the task. Here's an example of a PRD I recently co-created with Claude Opus: MCP Memory Server - Product Requirements Document. To generate your PRD, you can even provide this one as an example and then describe your project at a high level.