LAYER 01
Stable instructions
Role, goal, policy boundaries, tool contracts, response format, and non-negotiable stop conditions.
LESSON 07 · MODEL INPUT
Context Engineering
The model should not see everything the system knows. Build the smallest, freshest, most decision-relevant context for the current step.
Context is a runtime product, not a prompt dump.
Good context engineering decides what to retrieve, what to summarize, what to exclude, how to represent uncertainty, and which authoritative facts must be refreshed before each decision.
D→V RULE
Every context item should earn its place. Include it because it changes the current decision, constrains action, or supplies evidence—not because it exists.
LAYER 02
Current state
Fresh case facts, unresolved questions, approval state, budgets, recent tool results, and loop position.
LAYER 03
Retrieved evidence
Only policies, prior cases, and source records relevant to the immediate decision.
01Determine decision
02Refresh authoritative state
03Retrieve relevant evidence
04Compress + label uncertainty
05Assemble bounded context
INTERACTIVE LAB · BUILD THE CONTEXT
What should the model see now?
The current decision is whether to recommend a recovery route. Select only information useful and authorized for that decision.
Current bounded goal
Defines the decision.
12Defines the decision.
Fresh shipment state
Authoritative current evidence.
18Authoritative current evidence.
Relevant SLA rule
Constrains the decision.
15Constrains the decision.
Current recovery option
Directly relevant candidate action.
22Directly relevant candidate action.
Authority rule
Prevents unauthorized execution.
20Prevents unauthorized execution.
Prior case narrative
Potentially useful, but may anchor incorrectly.
35Potentially useful, but may anchor incorrectly.
Entire policy manual
Mostly irrelevant.
40Mostly irrelevant.
Unrelated employee data
Outside need-to-know.
25Outside need-to-know.
ASSEMBLED CONTEXT
87 / 120 budget units
Focused context.
LESSON ARTIFACT
Write the Context Assembly Policy.
Specify how the runtime builds model input at each step rather than relying on one static prompt template.