Compositional Prompt Assembly

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

A review workflow supports Windows/Linux. Combining both shell policies creates conflicts; teaching assembly needs a fixed core and one observed-platform adapter.

Mechanism

Keep stable responsibilities in core and platform maps in single-purpose fragments with explicit inclusion/trusted observations. Record versions/order, select one adapter and do not guess unknown platforms. Render fixtures for missing/duplicate/conflicting maps and untrusted slot boundaries. Maintain shared rules once.

Bad example

Paste all platform commands and let the model resolve contradictions while arbitrary slot text becomes authority.

Good example

Keep teaching core-review neutral and select one adapter from observed platform. Record identities/rendered prompt, reject conflicting maps and label user slots data. Resolve unknown platform rather than infer from path appearance.

Why the change matters

Explicit inclusion makes final prompts inspectable; single-source rules reduce drift. Reusable bases/variables still need combination checks.

Observable expectation

Windows renders one Windows adapter and Linux one Linux adapter; unknown does not default arbitrarily. Two adapters fail fixtures. Ignore-rules slot content remains data without changing core permission.

Limits

Frozen evidence composes visual style/variables; the OS example is a mechanism adaptation, not source implementation. Verify precedence/encoding and avoid excessive fragmentation of short stable prompts.