CORE GUIDE
Instruction authority
Reliable prompting depends on knowing which instruction source is allowed to govern behavior; system, developer, retrieved context and user text are not interchangeable peers.
Mental model
Instruction authority is a source-aware precedence model. Higher-authority application rules constrain lower-authority requests, while retrieved content is usually evidence rather than a new owner of system behavior.
Why it matters
Without an authority model, a system can be locally well-written yet globally contradictory. This matters for ordinary product behavior as much as for security: policy text, examples, tool output and user requests must be interpreted according to their role, otherwise the model is asked to resolve governance from wording strength alone.
01
Assign authority before resolving wording
Label each source by ownership and purpose before asking the model to act. Application-owned rules define non-negotiable constraints, user input defines the task within those constraints, and retrieved material supplies facts unless explicitly promoted by the application. When two instructions conflict, resolve the conflict at the source-policy level instead of adding stronger adjectives to the text.
02
Example: retrieved policy tries to override approval
A support system retrieves a VIP refund article that says to ignore lower refund limits. The user also asks for an immediate $300 refund. A flat authority model treats both as persuasive instructions. A source-aware model keeps the application's $100 approval boundary authoritative and treats the article as eligibility evidence, not as permission to bypass the gate.
Common failure modes
- Treating all prompt text as equal-authority instructions.
- Letting retrieved documents silently redefine application policy.
- Trying to solve precedence conflicts with more emphatic wording instead of source rules.
Engineering heuristics
- Give every instruction-bearing source an explicit role and owner.
- Treat retrieved content as evidence by default, not executable authority.
- Resolve conflicts in policy and architecture before polishing prompt wording.
Takeaways
- 01Authority is about source role, not rhetorical strength.
- 02Context can contain instructions without owning system behavior.
- 03A clear precedence model reduces both ordinary inconsistency and prompt-injection risk.
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.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.