The streaming renderer for Claude on web and desktop was overhauled to update only the parts of a response still changing. The result: long replies stall 9x less on slower laptops, worst-case freezes are 4.5x shorter, and 120Hz MacBooks now sustain a full 120fps from start to finish.
Developer Lydia Hallie calls out the Claude web and desktop streaming renderer improvements as a welcome quality-of-life upgrade, quoting the ClaudeDevs announcement of 4x smoother long-answer streaming.
Responding to feedback about the difficulty of using AGENTS.md or customizing the system prompt, the Claude Code team confirmed work is underway to make the tool more hackable, including easier support for AGENTS.md and other system prompt modifications. In the meantime, you can reference @Agents.MD from your Claude.MD. The team also noted they maintain different system prompts per model, since model families aren't interchangeable and prompt formatting has a big impact on performance.
Memory is now shared across Claude Chat and Cowork, so something you tell Claude to remember in one surface carries over to the other. The change came directly from user feedback, and the team is soliciting requests for what to unify next.
After detailing the reasoning behind model-specific system prompts, the Claude Code lead teased that more information on making Claude Code more hackable is coming soon.
Responding to the argument that cross-model prompt differences will converge and aren't worth optimizing for, the Claude Code lead counters with a concrete example: less capable models need hard constraints like 'DO NOT WRITE COMMENTS' to avoid bad behavior, but smarter models use judgment well, and applying those same hard constraints to smarter models forces unnecessary rigidity. The gap shifts but persists.
Asked whether the plan for making Claude Code more hackable involves open-sourcing the harness, the team confirmed they already do that on an annual basis.
