Claude Prompt Anti-Patterns Guide

Common prompt anti-patterns to avoid for frontier models: 1. Verification rituals: Instructions like 'carefully check your work' or 'double-check before answering' are often taken literally, wasting tokens. 2. Thoroughness and emphasis boosters: Phrases like 'be as thorough as possible' or 'key note: you must always...' make models verbose and trigger unnecessary tool calls. 3. Mandatory procedures and scratchpad scaffolds: Forcing fixed steps (e.g., 'think step by step in a scratchpad') or thinking templates overlaps with the model's native reasoning, wasting tokens. 4. Stale examples: Few-shot examples designed for older models' mistakes can make frontier models mimic unnecessarily long reasoning chains. 5. Contradictory rules: Conflicting instructions (e.g., 'always refund within policy' and 'never refund without approval') are executed literally, degrading performance. 6. Dated configuration: Settings written for older Claude versions (e.g., manually setting thinking budgets) may be rejected by the platform for newer models. These anti-patterns accumulate over time and silently drag down performance when upgrading to the latest model. Frontier models follow instructions so well that bad rules mislead them into wasting effort and tokens.