Context Engineering Advisor - AI Context Engineering Diagnostic Consultant

Skill Overview

An interactive diagnostic consultant that helps product managers determine whether an AI workflow is engaging in “context accumulation” (feeding in large amounts of material with poor results) or genuine “context engineering” (structurally allocating the AI’s attention), and provides practical memory architecture and context boundary solutions.

Applicable Scenarios

1. Unstable AI output quality: You have pasted an entire PRD and all interview records into the AI, yet its answers remain vague and perfunctory—“sometimes useful, sometimes not.” You need to diagnose whether context accumulation is diluting the model’s attention.
2. Overly complex multi-agent workflows and soaring costs: Agent A passes the entire context to Agent B, which passes it to Agent C. Token consumption keeps growing, retries become routine, yet accuracy does not improve. You need to define context boundaries for each agent.
3. Designing memory and retrieval architectures for AI product features: You are building a RAG system, vector database, or agent memory system and need to determine which information should be persisted, which should be retrieved on demand, and who is responsible for maintaining context boundaries.

Core Features

1. Diagnosing symptoms of context accumulation: Through an eight-item symptom self-check and five diagnostic questions (“What decision does this context support? Can retrieval replace persistence? Who is responsible for the boundary? What specific failure would occur if it were excluded? Are you fixing the structure or avoiding it?”), the consultant identifies the severity of “context hoarding syndrome” and provides corresponding intervention recommendations.
2. Designing a two-layer memory architecture: Tailored to your actual workflow, it designs a combination of short-term conversational memory (single-session memory and timely summarization) and long-term persistent memory (constraint lists, glossaries, declarative and procedural memory), while creating on-demand retrieval strategies for episodic context such as historical materials.
3. Implementing the Research → Plan → Reset → Implement cycle: This four-stage cycle—Research → Plan → Reset → Implement—eliminates context decay. The research stage initially allows messy context, which is then compressed into a high-density planning document. After clearing the context, implementation proceeds based solely on the plan. The process also includes a Context Manifest template and phased action checklists.

Frequently Asked Questions

What is the difference between context engineering and prompt engineering?

Prompt engineering focuses on “how to phrase a particular instruction,” while context engineering focuses on “what should and should not be in the AI’s working memory”—it is a systems-engineering discipline for designing information architecture for AI. This skill is explicitly not a tutorial on prompt writing, nor does it recommend tools. It diagnoses the organizational structure of context: boundaries, retrieval strategies, and memory layers. Even a perfectly written prompt will produce diluted outputs if the context is stuffed with irrelevant material.

Why do AI responses sometimes get worse when more information is provided?

Because the AI’s attention is a scarce resource. Placing an entire document repository or all meeting records into the context creates reasoning noise and “context decay” (middle information is overlooked, while dead ends and errors continually accumulate). The model tends to focus on the beginning and end while ignoring the middle, causing multi-step reasoning capabilities to deteriorate. Context accumulation assumes that quantity equals quality. Context engineering instead begins by clarifying the decision that needs to be made and provides only the high-density information that supports that decision.

Who is this skill for? What prerequisites are needed?

It is primarily intended for product managers and AI product teams already using AI—whether ChatGPT, Claude, or a self-built agent workflow—especially those experiencing symptoms such as inconsistent outputs, repeated retries, and rising token costs. It is not suitable for AI beginners who are still learning basic prompting, and it does not address tool selection. The entire diagnostic process takes approximately 15–20 minutes and uses an interactive question-and-answer format. You can start with nothing prepared—it will first ask what you are currently feeding the model and where problems have occurred.