Anthropic and OpenAI’s New Prompting Rules: Leaner Prompts and External Context
Clip title: Anthropic and OpenAI Just Changed the Rules of Prompting Author / channel: Enovair URL: https://www.youtube.com/watch?v=OiplQPjdA4c
Summary
The video discusses a significant shift in effective AI prompting strategies, highlighting that while AI models are becoming smarter, many users are still interacting with them using outdated methods. This traditional approach, characterized by excessive micromanagement and overly detailed instructions within a single prompt, can actually degrade output quality. The core message emphasizes moving away from these old habits towards a more streamlined, systemic interaction with advanced AI models like OpenAI’s GPT and Anthropic’s Claude.
According to new guidance from both OpenAI and Anthropic, the key to better results lies in “leaner prompts” and leveraging external context. Instead of cramming all information and rules into one lengthy prompt, users should now give the AI the complete, concise task upfront. Unnecessary role assignments should be avoided unless they genuinely alter the task’s execution. Crucially, persistent context that the model needs should be externalized into dedicated “homes” such as projects (for stable, company-wide information like services and pricing), client-specific folders (for call transcripts and background), and “skills” (for repeatable, task-specific workflows like meeting prep or content repurposing). This strategy is termed “progressive disclosure,” where context is loaded only when relevant.
The video outlines five essential parts of a prompt that still matter, but with a refined approach for each. For instance, instead of vague instructions like “double-check your facts” or “be accurate,” users should provide specific, verifiable standards such as “cite a source for every claim I could look up” or “separate what you confirmed from what you concluded.” Regarding format and length, explicit directives are now vital; rather than hoping for brevity, one should specify desired output lengths (e.g., “3 bullets maximum” or “match document’s length to job needs, no filler”). Finally, the use of “boundaries” or rules should be minimal and reserved only for truly inflexible constraints, avoiding absolute language like “always” or “never” unless absolutely necessary, to prevent over-constraining the model and allowing it more autonomy within defined limits.
In conclusion, the overarching takeaway is that effective AI interaction is no longer about crafting a single, perfect prompt, but about building an intelligent system around it. By decluttering prompts, externalizing reusable knowledge and repeatable processes into structured environments, and being precise about desired outcomes and boundaries, users can achieve stronger, more streamlined results. This methodological shift allows the AI to operate with greater autonomy and efficiency, ultimately transforming how users delegate tasks to their AI tools, enabling a repeatable workflow that saves time and consistently delivers higher quality output across different platforms.
Video Description & Links
Description
OpenAI and Anthropic both just published new guidance for prompting their newest models, and a lot of what we were all taught to do is now actually working against you.
I read through the dense documents so you don’t have to. In this video we talk about the five parts of a prompt that still matter, the five habits to break, and do a real rebuild of a prompt. We take a bloated client-proposal prompt and cut it down until the entire workflow runs on one sentence, in ChatGPT or Claude.
Links to the publications being referenced in the video: https://developers.openai.com/api/docs/guides/latest-model https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5 https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models
CHAPTERS 0:00 - Intro 0:45 - What Changed 2:38 - Part #1 4:13 - Part #2 8:43 - Part #3 10:58 - Part #4 13:07 - Part #5 16:22 - Turning this into a Skill
URLs
- https://developers.openai.com/api/docs/guides/latest-model
- https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5
- https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models
Related Concepts
- Prompt Engineering — Wikipedia
- Lean Prompts
- External Context
- AI Model Optimization
- Progressive Disclosure — Wikipedia
- Context Management — Wikipedia