Here’s a copy of this page: https://sean.horgan.net/github-sparks-system-prompt-132 I dug more into more AI system prompts over the weekend. Thanks to Simon Willison’s archaeological rummaging through GitHub Spark, which you can find [here](https://github-spark-docs.simonwillison.net/#system-prompt--system-prompt-content), I’m seeing just how much effort the GitHub team put into it. You can learn more about Spark here: [https://githubnext.com/projects/github-spark](https://githubnext.com/projects/github-spark) ![[image.png]] The prompt weight in around 30K characters which seems like a lot but when you break it down to what an AI can handle, it’s nothing. At 4 characters/token, that’s about 7500 tokens and well within the context limits for most SOTA models, e.g. 1M for Gemini, 200K+ for Claude 3, 128K for Mistral 3 Small. Llama 3’s support for 8K on its first release feels like the stone ages already.  I wanted to see the token budget per section of the prompt to get a better sense of their priorities:  ![[image 1.png|image.png]] While the prompt is framed as an Applications Guide, it reads as the combination of a job description, design principles, and detailed instructions. A significant portion of the prompt is dedicated to a deep design philosophy. It emphasizes creating applications that are not just functional but also aesthetically pleasing and emotionally resonant.  Most striking to me is the requirement to write requirements: PRDs must be generated if they don't exist and then kept up to date as you apply revisions. The PRD frameworks nails some great details that all PMs should be thinking about: - Lead with mission, success metrics, and user experience qualities - Select features based on specific problems and lay out pivotal interactions for essential journeys all the way to the goal - A nice nod to accessibility by requiring [WCAG AA compliance](https://www.w3.org/WAI/WCAG2AA-Conformance)! I didn’t see any of the following mentioned: - Business objectives, strategic concerns, product-market fit - Target market, users, and personas - Go-to-market plans - Governance, policy and other enterprise requirements - Infrastructure considerations, e.g. orchestration, costs  While you could argue that those are usually better captured in other artifacts, I think the omission speaks to the Spark’s focus on “micro apps”. Overall, I’m super impressed by Spark’s PRD framework.