Bun rewrites in Rust + Claude Code CLI hacks

Claude Fable 5 & model neutrality in review

Following Anthropic's redeployment of Claude Fable 5 earlier this month with new security classifiers, engineering discussions throughout July have focused on the broader implications of safety guardrails. Industry analyses point to growing debates over whether AI coding assistants could become non-neutral infrastructure when assisting on competing products.

We are actively testing Fable and the results seem to show a lot of promise. Once we will have gathered more substantial data, we will share our findings on our blog and our podcast! Stay tuned!

The case against skills: Are we over-instructing LLMs?

Our CTO, Xavi, shared an insightful piece mid-month examining whether explicitly instructing models on tasks they are already trained for does more harm than good. The core argument is that over-specifying rigid skills fills up context windows and can actually degrade output quality.

Bun rewrites its core engine from Zig to Rust

In a major technical shift published this month, the Bun team detailed their complete core architecture rewrite. Carlos, a Tech Lead at MarsBased, brought this to our attention. The post details why and how they transitioned from Zig to Rust to achieve long-term stability and ecosystem scalability.

Claude Code productivity hacks

• Reduce Permission Prompts: Daniel Cay, an engineer at MarsBased, discovered Anthropic's fewer-permission-prompts skill. It scans recent session transcripts to build a tool allowlist, cutting down prompt fatigue while maintaining security standards.

• Manage Context & Token Limits: In response to questions about heavy token usage from Pablo, an engineer on our team, our CEO, Àlex, reminded the team to run /checkup in Claude Code to purge unused skills and split local CLAUDE.md files.