🎯 The Shift
Claude Opus 5 and Fable 5 work fundamentally differently from previous models — "almost everyone is prompting them wrong." They're trained for end-to-end task execution, which inverts the old prompting wisdom. 0:00
🔧 All 7 Rules
Claude 5 models perform best with the complete task spec up front and are left to run. The framework: job + why + guardrails + what "done" looks like. Ben used it to prepare the video itself — a broad goal + why + guardrails + done-definition, let it run 10–15 min, got a useful pre-outline.
For complex end-to-end tasks it's hard to articulate every guardrail upfront. Anthropic internally uses an "interview me" skill that asks you questions about the task, surfaces unknowns, then produces a brief/prompt. (Karpathy's alternative: voice-transcribe and ramble for a few minutes — then structure it.) "Put more effort into pre-planning instead of quick-prompt-and-iterate."
On complex tasks the model hits decisions you didn't pre-specify. Understanding intent lets it make those small calls. The template: "I'm working on [larger task] for [specific person]. They need [what the output enables]. With that in mind, [request]." — same principle as delegating to a human.
Claude 5 models are long-running — "they don't do too little, they do too much." Define exit criteria + output style (e.g. "8–15 tips with examples and sources" + a reference example of the format). Fix the Opus 5 verbosity: "give the output in plain English that's easy to understand."
Instead of "never do X" / "avoid this," give the instruction with a reason. The new models respond far better to reasoned guardrails. (E.g. "make sure points are backed up by Anthropic's own team, so we have proof for our claims.") Anthropic even launched a rule rewriter skill to convert your CLAUDE.md/skills.
Opus 5/Fable 5 already verify and fix mistakes autonomously — instructing "double-check" or using a sub-agent to verify adds cost without improving results. Also avoid: "think step-by-step," "explain your reasoning," and aggressive CAPS emphasis (causes over-triggering).
Opus 5 gets jargon-heavy/verbose. Anthropic recommends a simple prompt in CLAUDE.md, desktop global instructions, or system prompt — e.g. "keep responses focused, brief, and concise; avoid jargon and being overly verbose." Set it once, don't re-fight it every prompt.
🥇 The Golden Rule
Anthropic's own framing: "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows." 12:45
💡 Key Takeaways
- Prompt the whole job (job + why + guardrails + done), not step-by-step instructions.
- Pre-plan with the "interview me" skill (or voice-rambling) before complex tasks — reduces iteration and token burn.
- Give the why — intent drives better decisions on unspecified details.
- Define done + output style — long-running models over-produce without exit criteria.
- Reasoned guardrails beat hard rules — and the rule rewriter skill converts your CLAUDE.md.
- Stop adding verification — these models self-verify; avoid step-by-step and CAPS emphasis.
- Fix the voice once in global instructions, don't re-specify style every prompt.