The One Claude Code Setting That Changes Everything

📝 TECHNIQUE

The One Claude Code Setting That Changes Everything.

Boris Cherny gets asked for his top Claude Code tip constantly — and the answer is deceptively simple: use auto mode. It eliminates the permission prompts that break your flow during multi-agent sessions, letting Claude Code run tool chains without stopping to ask you for approval at every step. The productivity difference between supervised and auto mode isn't incremental — it's the difference between pair programming and dictating to a typist. If you haven't switched yet, try it on your next session and time the difference. (3,826 likes | 201 RTs) Read more →

GPT-5.5 Pro Turns Out to Be a Ruthless Fact-Checker. Ethan Mollick reports throwing entire book chapters at GPT-5.5 Pro and watching it hunt down every key reference accurately. The caveat: it sometimes over-nuances, flagging correct claims as needing qualification. Still, if you need a first pass on factual accuracy at scale, this is currently the best tool for the job. (1,345 likes | 58 RTs) Read more →

The "Please Save Me Money" Prompt That Actually Works. Turns out you can literally ask Claude to optimize for cost — and it does. The trick resonated hard with budget-conscious developers, and the engagement numbers suggest a lot of people didn't know this was an option. Try appending a cost-consciousness instruction to your next expensive agentic workflow before reaching for a smaller model. (1,023 likes | 30 RTs) Read more →

The Self-Improving Codex Prompt Now Mines Your Session History. An updated meta-prompt for OpenAI Codex that scans across sessions, Memories, and Chronicle to identify recurring patterns, then auto-generates reusable skills and subagents. The prompt engineering community is iterating on this in real time — yesterday's version only looked at the current session. (2,216 likes | 168 RTs) Read more →

Anthropic Publishes Its Internal Claude Code Playbook. Anthropic released the exact workflow its engineers use internally — context files, custom slash commands, hooks, subagents — with side-by-side comparisons of stock vs. tuned setups. It's free and more useful than most paid courses on the topic. If you're running Claude Code without project context files and custom commands, you're leaving performance on the table. (14 likes | 4 RTs) Read more →

Claude Is Writing Illustrated Home Repair Manuals Now. Felix Rieseberg moved into an older house and started using Claude to generate illustrated, problem-specific instruction manuals for real-world repairs. Not code, not text — full visual guides with diagrams tailored to his exact plumbing and electrical issues. A reminder that multimodal generation has use cases far beyond software. (220 likes | 6 RTs) Read more →

💡 INSIGHT

Hackers Are Submitting Malicious CLAUDE.md Files to Popular Repos.

Socket Security documents a new attack vector that should worry every team using AI coding agents: attackers submitting PRs that add malicious .cursorrules and CLAUDE.md files to popular open-source repos. When a developer clones the repo and fires up their AI agent, the poisoned config file injects attacker payloads into every subsequent AI action. This is dependency confusion for the agent era — except most teams have zero review process for config files that look like harmless documentation. Audit your repos now. (81 likes | 8 RTs) Read more →

Anthropic Closes $30B Round at $900B+, Surpassing OpenAI.

Anthropic is set to close a $30 billion funding round at a valuation north of $900 billion — blowing past OpenAI's $852B to become the world's most valuable AI startup. The valuation has 2.4x'd in three months. Whether this reflects genuine technical moats or peak-cycle froth, the fundraising crown has a new holder, and it signals that investors see Anthropic's enterprise and safety positioning as a durable competitive advantage. (17 likes | 6 RTs) Read more →

Memory Now Eats Two-Thirds of AI Chip Costs. Epoch AI data shows memory has grown to nearly two-thirds of AI chip component costs — a structural shift from the compute-dominated era. This explains NVIDIA's margin pressure and why memory-efficient architectures (quantization, sparse attention, mixture-of-experts) aren't just nice-to-haves anymore — they're economic necessities. (266 likes | 283 RTs) Read more →

June's Model Calendar: Gemini 3.5, GPT 5.6, Sonnet 4.7. Brace yourself — June is shaping up to be the most crowded model launch month in AI history. Expected drops include Gemini 3.5 Pro, GPT 5.6 (and 5.6 Pro), Sonnet 4.7, and possibly Mythos. If your team is about to lock a model version for a production rollout, wait three weeks or plan an evaluation cycle for late June. (388 likes | 8 RTs) Read more →

YC Publishes Its AI Startup Wishlist for Summer 2026. Y Combinator dropped its latest Request for Startups — a public roadmap of what the most influential accelerator wants to fund this cycle. When YC telegraphs its bets, it simultaneously validates markets and floods them with competitors. Read it for competitive intelligence whether you're fundraising or not. (94 likes | 16 RTs) Read more →

🔧 TOOL

DeepSeek Reasonix: A Coding Agent That Leads With Price. DeepSeek launches Reasonix, a native coding agent optimized for aggressive caching and lower per-token costs — positioning squarely against Claude Code and Codex on price, not capability. With 395 HN upvotes, the builder community is clearly interested in a budget alternative for routine coding tasks. Compare the pricing before assuming your current setup is cost-optimal. (395 likes | 189 RTs) Read more →

Coral Replaces MCP Tool Glue With a Single SQL Interface. If you're stitching together multiple MCP sources with bespoke connectors, Coral offers a cleaner path: a local-first SQL runtime that queries across APIs, files, and data sources through one interface. Benchmarks show 20% higher accuracy and 2x cost efficiency compared to direct provider MCPs. One query language to rule them all. (159 likes | 15 RTs) Read more →

🏗️ BUILD

Obsidian Gets Agent Skills — and 32K Stars Agree.

Kepano ships obsidian-skills — a framework that teaches AI agents to natively work with Markdown, Bases, JSON Canvas, and Obsidian's CLI. With 32,853 stars, this is the clearest signal yet that knowledge management and AI agents are converging. If you use Obsidian as a second brain and AI agents as coworkers, this bridges the gap between where you think and where you build. (32,853 likes | 2,292 RTs) Read more →

🧠 LAUNCH

Meituan Open-Sources a SOTA Talking-Avatar Model Under MIT. Meituan releases LongCat Video Avatar 1.5 — a talking-avatar model that's arguably the best open-source option available, now MIT-licensed for commercial use. The video generation space has been locked behind proprietary APIs for too long; this gives indie developers and startups a self-hostable alternative that competes with the paid offerings. (119 likes) Read more →

🔬 RESEARCH

New Paper: Your Coding Agent Forgets Requirements Mid-Task. Researchers formalize "constraint decay" — the phenomenon where LLM agents progressively lose track of requirements during complex backend code generation. The longer the task, the more constraints slip. If you've ever had an agent nail the first three files of a refactor and then quietly drop a validation rule by file seven, this paper explains why — and it should inform how you structure agent-driven work into smaller, constraint-preserving chunks. (156 likes | 79 RTs) Read more →

🎓 MODEL LITERACY

Constraint Decay: When you give an LLM agent a complex coding task with multiple requirements — "use this ORM, validate these fields, follow this error pattern, write tests" — the model doesn't forget them all at once. It forgets them gradually, like a checklist where items near the bottom get progressively ignored. This is "constraint decay": the agent's effective working memory for requirements shrinks as the task grows longer and more complex. Today's research paper formalizes this, showing that decay accelerates with task complexity and codebase size. The practical implication: break large agent tasks into smaller units that each carry their own explicit constraints, rather than trusting the agent to hold twenty requirements across ten files.

⚡ QUICK LINKS

Tencent Hy-MT2-7B: 33-language translation model, self-hostable, sweet spot between quality and efficiency. (143 likes | 1.3K downloads) Link
Datasette 1.0a30: New agent and fixtures plugins ship as the SQLite explorer marches toward 1.0. Link
Ollama + Codex: Fully local AI coding agent — no cloud, no code leaving your machine. (12 likes | 4 RTs) Link
David Ha on Jevons Paradox: AI won't replace engineers — 10x productivity expands the problem space, increasing demand. (209 likes | 14 RTs) Link
The "Tells" of AI Content: Mollick observes recognition is accelerating as daily AI use makes the patterns obvious. (388 likes | 21 RTs) Link
Armin Ronacher on AI Tooling: The Flask/Ruff creator's perspective on how AI reshapes developer tools. Link

🎯 PICK OF THE DAY

The files we invented to make AI agents useful are becoming attack surfaces. Socket Security's report on malicious .cursorrules and CLAUDE.md PRs landing in popular repos is a wake-up call that cuts deeper than a typical vulnerability disclosure. These agent config files — designed to give AI coding tools project-specific context — are uniquely dangerous because they look like harmless documentation. No security scanner flags them. No CI pipeline blocks them. Most code reviewers skip right past markdown files during PR review. But once a poisoned config is merged, every developer who clones the repo and runs an AI agent inherits the attacker's instructions, potentially exfiltrating code, injecting backdoors, or modifying behavior invisibly. This is the agent-era equivalent of dependency confusion attacks, and most teams have zero review process for it. The fix is straightforward but requires cultural change: treat .cursorrules, CLAUDE.md, .github/copilot-instructions.md, and similar files as security-sensitive code, not documentation. Add them to your PR review checklists. Today. Read more →

Until next time ✌️