Instruction Ordering

Within a topic, the ORDER of instructions matters. Putting the directive first, reasoning between, and the constraint last is significantly more effective than the natural human pattern of leading with prohibitions.

Antipatterns

• Constraint-first pattern: "Don't use black. Use ruff format instead." Leading with the prohibition activates the forbidden concept before the desired behavior is established. The diagnostic detects this inverted ordering.
• Reasoning before directive: "Because mock objects hide integration bugs, use real database connections." The reason is stated before the instruction — the directive should come first so the agent knows what to do before learning why.
• Interleaved ordering: "Don't use mocks. Real tests catch more bugs. Use pytest with real connections. Never stub HTTP calls." Alternating between directives and constraints within a topic makes both weaker.

Pass / Fail

Pass

Use `pytest` with real database connections for integration tests.
Real integration tests catch deployment failures that mocks hide.
*Do NOT use `unittest.mock` or test doubles for service boundaries.*

Fail

Don't use mock objects or test doubles. They hide integration bugs.
Use real database connections instead.

Fix

Restructure instructions as:

[DIRECTIVE] Use real implementations — real database connections, real HTTP endpoints.
[REASONING] Real integration tests catch deployment failures and configuration
errors that would otherwise reach production undetected.
[CONSTRAINT] Do not use mock objects, stubs, or test doubles.

Never write "Don't use X. Instead, use Y." Write "Use Y. [reason for Y]. Don't use X." Reasoning should support the directive, not explain what's wrong with the prohibited thing.

Limitations

Detects ordering patterns within instruction clusters. Cannot evaluate whether the reasoning content actually supports the directive.