signaldepth

3/3 Penpot For: product designers / frontend engineers Signal: open-source design collaboration with code-oriented workflows. Caveat: niche signal, not a general install recommendation.

2/3 GitNexus For: developer tooling builders / codebase intelligence users Signal: browser-side code knowledge graphs for faster repo reading. Caveat: niche signal, not a general install recommendation.

Signal Depth | GitHub Trending signal - Apr 28, 2026 Standout: mattpocock/skills Reusable AI-coding workflow skills, not another prompt list. Read: metadata, README, repo tree. Not run: clone/install/benchmark. #GitHubTrending #OpenSource

"Claude prefers XML tags." "GPT works better with Markdown." We tested this across 8 models, 120 calls, 2 APIs. XML: 0.80 Markdown: 0.80 Plain text: 0.83 No model showed statistically significant format preference. Anthropic's docs recommend XML. Their quantitative evidence? Zero. Two independent studies (2411.10541, Systima.ai 600 calls) found the same: format is noise. If you're spending time converting prompts to XML — stop.

Position bias isn't a training problem, it's architectural. Chowdhury (2603.10123) just confirmed what "Lost in the Middle" showed in 2023: attention mechanisms structurally downweight mid-context tokens. Front-load or back-load. Never middle-load.

@simonw Exact string matching today, but the precedent is what matters: platforms now treat system prompts as inspectable metadata, not just model input. Once that boundary is crossed, the matching criteria only get broader.

Weirdly it looks like only that EXACT string triggers this behavior

Billing different based on text contained in the system prompt is a really bad look