Context documents over prompts, plus Supervision, Penpot and Clipstack

A post attributed to Andrej Karpathy argues for structured context documents over ad hoc prompts, while Supervision passes 45,000 GitHub stars, Penpot develops open design-and-code collaboration and Clipstack offers a public demo of its clipboard manager.

Context as a reusable engineering document

The most consequential idea today is a working method rather than a product. A post attributed to Andrej Karpathy describes a document-based approach to steering LLMs, which commentary frames as the end of ad hoc prompt engineering. That headline overstates what one post can settle, but the core suggestion is practical: keep the context an agent works from in a structured, reusable artifact instead of rewriting prompts for each task. For engineers building agents, this points toward giving specs and context docs the same review, versioning and ownership that code already gets.

Open-source projects at different layers

Two open-source projects are worth noting, though they serve very different work. The computer-vision library Supervision added roughly 5,000 GitHub stars in three weeks to pass 45,000, a sharp acceleration that suggests growing use in vision and annotation pipelines. Stars measure attention more directly than production adoption, so the jump is a reason to evaluate it rather than proof it has become a default dependency. Penpot keeps building an open design-and-code collaboration platform, which gives teams a non-proprietary option as more design workflows generate code directly.

Clipstack opens a public demo

Clipstack, a new clipboard manager, has released a public demo along with a getting-started guide on its GitHub wiki. It is a small launch, but the demo makes it easy to judge before installing anything if clipboard friction is a daily annoyance.