CORE GUIDE
Prompt, context and runtime boundary
Prompt wording, supplied context and runtime enforcement solve different problems; reliable AI systems separate persuasion, evidence and capability control instead of treating them as one prompt-design layer.
Mental model
Prompt tells the model how to interpret and respond, context supplies the information available for the decision, and runtime determines what actions are actually permitted or executed. Safety and reliability require the three layers to agree without pretending they are interchangeable.
Why it matters
Many production failures happen because teams repair the wrong layer: adding stronger text to stop a dangerous tool call, adding more retrieval to fix a permission problem, or adding a runtime block for what is really missing context. A clear boundary makes diagnosis faster and prevents prompt changes from being mistaken for hard guarantees.
01
Diagnose the failing layer before intervening
Ask three questions in order: did the model receive an unambiguous behavioral instruction, did it receive the evidence needed to decide correctly, and could the runtime still execute an action that policy should forbid? Improve prompt clarity for interpretation failures, context selection for evidence failures, and capability gates, schemas or approvals for enforcement failures.
02
Example: model recommends approval but tool still executes
A refund assistant is prompted to require human approval above $100 and correctly outputs 'approval required'. Yet the tool adapter still accepts a direct $300 refund call from another path. The prompt layer is fixed; the system remains unsafe because enforcement belongs in the runtime boundary.
Common failure modes
- Using prompt language as the only control for irreversible actions.
- Adding context when the real failure is missing capability enforcement.
- Building hard runtime blocks for simple ambiguity that should be clarified earlier.
Engineering heuristics
- Classify each intervention as prompt, context or runtime before implementing it.
- Use runtime controls for permissions and irreversible side effects.
- Use evidence traces to prove which layer changed the outcome.
Takeaways
- 01Prompt, context and runtime are complementary layers, not substitutes.
- 02Model compliance is not the same as capability enforcement.
- 03Correct layer diagnosis prevents expensive fixes that do not change the real failure.
Reading evidence
This records actions you actually took; it does not claim mastery, proficiency, or certification.
Used in
This Concept is reused across these canonical learning paths.