21 practical Claude tips: use Medium effort, never revisit old chats
rubenhassid · x · 2026-10-29
Ruben Hassid shares a 'Claude Mastery Plan' cheat sheet with 21 concrete rules:
- Setup: run /setup for self-onboarding; Opus 5.5 on Medium effort (Medium does 80% of the work, never Max); the $100 plan yields 5× the work of the $20 plan.
- Prompting: state the goal, never the how; max 3 rules per prompt (adherence drops to 5.7% at 8 rules); delete examples (Mistral: 74% → 83.8% without them); end with 'Ask me questions first.'
- Save limits: never return to old chats (expired cache = full re-reads); edit the prompt instead of correcting Claude; crop every screenshot (full screen: 4,784 tokens vs <100 cropped); ask 3 things per message.
- Delegation: add 'Explore broadly before you act' and 'Don't offer to continue: continue'; wrap pasted emails in tags; request HTML pages, not .docx.
- Safety: never click thumbs up/down (it sends the whole chat, kept 5 years); wipe your setup every 6 months.
More from coding & agent
- Routing coding agents' questions by decision type: 762 questions, only 59 needed the owner — DynoKore · 2026-10-29
- 1.2B tokens later: multi-agent coding run rebuilds a 3D game but hits diminishing returns — teortaxesTex · 2026-10-29
- Unreal Engine runs in the browser via a single Claude agent prompt — chongdashu · 2026-10-29
- Evals and Tracing 101: A Practical Guide for Engineers Building With AI — blaizedsouza · 2026-10-29
- Stop repeating yourself to every AI: a 4-step fix for syncing memory across Claude, ChatGPT and personal agents — jonathanmalkin · 2026-10-29
- hunch: GPU-free ~20ms Jev-like decisions built on EmbeddingGemma 2 — iamrobotbear · 2026-10-29