Claude Opus tips: Why 'why' matters, labels work, avoid all-caps

Nissa Seru's LessWrong post 'Notes on Collaborating with Claude Opus' distills practical observations from extensive use of Claude Opus 4.7. Seru highlights four actionable strategies for getting better results from the model. First, always include the reasoning behind an instruction: adding a 'why' significantly boosts both the salience and the quality of execution. Second, leverage structured formatting by instructing Claude to break replies into labeled sections and subsections (e.g., A1, B3). This enables precise referencing during long conversations — Seru notes Claude actively uses these handles itself, far more ergonomic than vague references.

Third, Seru warns about 'pink-elephant risk' in negative framing: telling Claude what not to do can increase the salience of the forbidden action, potentially increasing its prevalence. The net effect depends on baseline salience. Finally, Seru advises against using all-caps for emphasis — it risks landing Claude in a 'yelling' attractor basin where the model interprets the message as managing an emotion rather than executing an instruction. The post reflects a nuanced understanding of how large language models process communication, offering concrete techniques for professionals seeking more reliable collaboration with AI systems.