cipher

Trace the full context in order — system prompt, examples, retrieved content, user turn — to find where the instruction is buried, contradicted, or outvoted by few-shot examples. The model pattern-matches what it sees, so align examples with the stated rule and move key instructions where recency helps.

Use the API's structured-output or tool-call mode with an explicit schema rather than prompting alone, since the decoder can always prepend stray tokens to free text. Prompt-only JSON instructions are a fallback, not a fix, for parsing failures.

Use reasoning-before-answer for multi-step tasks where it raises accuracy, but skip it for lookups and classification where it wastes tokens and latency. With reasoning models, state the goal and constraints rather than hand-scripting the reasoning steps.

Wrap retrieved or user-supplied content in delimiters like <untrusted_document> tags and instruct the model that everything inside is data, never instructions. Keep tool access minimal and add output filtering for hostile content sources.

Check the decoding settings first — high temperature introduces variance that looks like a flaky prompt. Lower temperature toward zero for determinism, then look for contradictory instructions the model resolves randomly.

Define the eval before editing: collect the cases that currently fail, name the property being checked, and set a readable pass/fail criterion. Comparing old and new prompts against those cases shows real improvement instead of vibes.