D

underdog.bsky.social

20 octubre 2026 15:37

Context files beat model choice. Repos with a decent CLAUDE.md or rules file gave us consistent results across Claude Code, Cursor and Codex. Repos without one gave us consistent nonsense across all three. The model is the easy part. The system around it is yours.

PierreHenry.DEV ⚡️

pierrehenry.dev

20 octubre 2026 15:37

My custom AGENTS.md, rules and skills go beyond naming conventions. They guide architectural decisions, reduce duplication, challenge bad abstractions, protect existing code, enforce security boundaries and require meaningful validation.

Dylan Asdale

dylan-asdale.bsky.social

20 octubre 2026 15:37

Claude Code can delegate to subagents in .claude/agents/. Plan mode even uses a research subagent so exploration doesn't clutter your main thread. Start simple: one subagent, one job. #ClaudeCode #AI

Sahil • gitsu, shareling, gitback

sahildave.com

20 octubre 2026 15:37

𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝘁𝗶𝗽: before you let agents loose on your GitHub issues, have Opus send out 𝟯 𝘀𝘂𝗯𝗮𝗴𝗲𝗻𝘁𝘀 to check 𝘁𝗵𝗲𝘀𝗲 𝟭𝟬 𝘁𝗵𝗶𝗻𝗴𝘀, then clean up in one pass. ⭐ Prompt to avoid agents burning tokens on tickets that already shipped #BuildInPublic

Dan Epps

epps.bsky.social

19 octubre 2026 15:53

So I worked with Claude Code and Codex to build skills with structured workflows that divide complicated tasks among subagents, look in the right place for primary sources, and include strong “deterministic” checks to minimize hallucination risk.