Isolated iteration: Changing a single variable per test cycle, running the agent with identical inputs, and comparing old vs. new output side by side. The alternative (changing multiple variables at once) makes it impossible to attribute improvement or regression to any specific change.
Prompt vs. data vs. structure: Three distinct categories of problems that produce similar symptoms (weak output). Prompt problems affect how a node interprets its task. Data problems mean the node never received the information it needed. Structure problems mean the node is handling too many cognitive tasks at once.
Before you iterate, define your quality target. Write down what good output looks like for the specific node you're improving. Without this, you'll cycle between making output longer, then shorter, then longer again.
The iteration loop:
When prompt changes aren't working, stop after two failed attempts and look upstream:
The diagnostic heuristic:
Isolated iteration: Changing a single variable per test cycle, running the agent with identical inputs, and comparing old vs. new output side by side. The alternative (changing multiple variables at once) makes it impossible to attribute improvement or regression to any specific change.
Prompt vs. data vs. structure: Three distinct categories of problems that produce similar symptoms (weak output). Prompt problems affect how a node interprets its task. Data problems mean the node never received the information it needed. Structure problems mean the node is handling too many cognitive tasks at once.
Before you iterate, define your quality target. Write down what good output looks like for the specific node you're improving. Without this, you'll cycle between making output longer, then shorter, then longer again.
The iteration loop:
When prompt changes aren't working, stop after two failed attempts and look upstream:
The diagnostic heuristic:
