SYSTEM DESIGN / RECOMPOSED
Knowledge, tools and control. Only their interplay turns an AI idea into a system that proves itself in everyday use.
01 / THE GRAMMAR OF A SYSTEM
No tool bingo—just well-reasoned decisions. Each layer asks a different question. Insights feed back into the design.
PROBLEM BEFORE TECHNOLOGY
Clarify users, problem, goal and scope for action before selecting models or frameworks.
A clearly scoped brief with explicit assumptions.
↺ Observations from operation can change decisions in any earlier layer.
02 / THE BLUEPRINT
A support agent as an example. Choose an area and explore how knowledge, actions, context and coordination come together.
COORDINATE
The agent answers, asks follow-up questions, uses an allowed tool or escalates. Steps, repetitions and runtime are bounded.
A model proposal is not permission to execute.
03 / DELIBERATE BOUNDARIES
Not every action needs the same scope. Explore four qualitative levels—this is not an automatic risk assessment.
The AI answers within clear rules. When knowledge is missing, it asks follow-up questions or hands over.
Control: reliable context, quality checks and escalation.
04 / BEHIND THE RESULT
The right evaluation shows which component needs improvement. An overall score alone does not explain the cause.
Context precision and recall help assess the relevance and completeness of the context.
Groundedness and faithfulness assess adherence to the available context.
Review tool selection, arguments, success and repetitions separately.
Evaluate resolution, escalation, latency, cost and customer satisfaction.
05 / YOUR DESIGN
Apply the framework to your use case. The canvas exports your design locally as Markdown or JSON.