Examples and Decomposition

Demonstrate edge cases, not decoration

Few-shot prompting supplies a small set of input-output examples. Examples teach format and decision boundaries without changing model weights. They should cover representative ambiguity and mistakes, not merely repeat easy cases that the model already handles.

Decomposition breaks a complex job into smaller explicit stages, such as extract facts, validate fields, then classify. Stages can be tested separately and sometimes handled by deterministic code. Asking for a private hidden chain of thought is not necessary; request concise evidence, intermediate structured fields, or tool results that are appropriate to expose and verify.

stage 1: extract {service, symptom, timestamp} from the incident
stage 2: validate required fields in code
stage 3: classify using only validated fields

Scenario: One giant prompt asks a tiny local model to parse a log, infer the root cause, choose a runbook, and write shell commands. Splitting extraction from retrieval and keeping command approval in deterministic code makes failures local and safety enforceable.

Warning: Examples are instructions too. Never include secrets or sensitive production records, and test whether an example accidentally teaches a shortcut unrelated to the real task.