A batch that ranges from practical agent tooling to deep systems programming.
Arpan Patel’s Beyond the Prompt: Claude Code Mastery is the most thorough Claude Code guide I’ve read; it hit #7 on Hacker News for good reason. It covers the full .claude directory as a layered config system, how Boris Cherny actually writes CLAUDE.md (keep it short, let Claude write its own rules from mistakes, the “would removing this cause a mistake?” filter), using CLAUDE.local.md to capture PR feedback, skills with real examples, custom subagents, underused commands like /rewind and /goal, and MCP workflows including a three-tier Obsidian memory setup. The throughline (“setup is the work, execution is verification”) matches how I’ve come to use it. If you use Claude Code daily, block out time for this one.
Tejas Chopra’s Headroom post (with the open-source repo) diagnoses a real problem: MCP tools and database queries flood your context window with redundant JSON. His Claude bill hit $287, and 90% of it was tokens he didn’t need: 500-row query results when he needed 3, 50KB API responses where the signal was 200 bytes. Headroom is a proxy that sits between your app and the LLM and does statistical compression (keep anomalies, keep query-relevant items, compress the rest reversibly), cache alignment (moving dynamic content like dates to the end so prompt caching actually hits), and rolling-window management that never orphans a tool call from its response. He reports dropping from ~$280 to ~$110/month with no accuracy loss. Apache-licensed. A smart application of data-infra thinking to LLM plumbing.
Google Research’s Agentic RAG with Gemini Enterprise covers the shift from naive retrieval-augmented generation (retrieve once, generate) to agentic RAG, where the agent iteratively reasons about what to retrieve, evaluates the results, and refines its search before answering. The focus is on dependability for enterprise use: reducing hallucination by grounding responses in verified retrieval loops. Useful context if you’re building RAG systems and hitting the limits of the single-shot retrieve-then-generate pattern.
Justine Tunney’s Restartable Sequences calls rseq “the best kept secret at the frontier of system programming.” Restartable sequences (Linux 4.18+) let you build thread-safe data structures without mutexes or atomics that scale across many-core CPUs: the kernel restarts your critical section if the thread is preempted or migrated mid-sequence. Currently only tcmalloc, jemalloc, glibc, and cosmopolitan use it, and it requires handwritten assembly. Justine believes all system programming languages will eventually be redesigned to express restartable sequences. Dense, low-level, and fascinating if you care about high-performance concurrency.
The Lattice-Based Cryptography PDF from Cryptography101.ca is a course-grade primer on the math behind post-quantum cryptography. Lattice-based schemes are the foundation of most NIST post-quantum standards (like Kyber and Dilithium), and understanding them matters as the industry migrates away from RSA and elliptic-curve crypto ahead of practical quantum computers. Dense and mathematical, but the clearest structured treatment I’ve come across for developers who want more than a hand-wave.