You can drop instruction files into ~/.claude/output-styles and select them under /config to control how Claude formats its responses. Useful for cutting down on verbose, jargon-heavy replies at the end of a long session.
Custom output styles stored in ~/.claude/output-styles differ from CLAUDE.md instructions in one key way: they live in the system prompt itself and can be switched on the fly via /config, rather than sitting in the project-level markdown file.
A follow-up to the output styles tip: one user noted they'd already been using a personalized version of the style, previously named after themselves, before the broader community discovered the feature.
Pushing back on the idea that agentic loops are a short-lived workaround for weaker models, trq212 draws a distinction between narrow execution loops (which may be fading) and event-driven or scheduled loops, which remain foundational. Claude Tag, for example, is described as mostly loops.
Responding to Simon Willison's question about how Claude Tag differs from a simple cron-scheduled prompt that calls tools in sequence, trq212 explains the key distinction: Claude Tag pulls in memory and organizational context from Slack, which a plain scheduled prompt doesn't do.
Bcherny's preferred approach to performance work: give an agent a goal (e.g. get p95 latency under 300ms), tell it to use a profiler, and let it run autonomously until it's done. The comment comes in response to a thread about AI finding a 15-30% memory efficiency gain in Turbopack nearly autonomously.
Android simulator support is on the way. The reply comes in response to a user asking whether Claude Code, which apparently already supports some mobile tooling, also works with Android simulators.
