Welcome back. We all code differently, and Anthropic's new feature lets us work in our own style. Meanwhile, decision model Jev faces three new rivals, and a new launch claims to be the first human interaction model.
Also: A SpaceXAI engineer’s prompt to catch AI mix-ups, Karpathy’s personal tricks for clearer LLM outputs, and a JS + CSS animation devs are obsessed with.
Today’s Brief
Why Ruby on Rails’ creator stopped coding manually
A dev turned his agent's work into live pixel-art cartoons
Tokenmaxxing your Claude Code (tutorial)
Cut agent costs with model routing (cookbook)
TODAY IN PROGRAMMING
Click here to watch how you can use Claude Code mods.
You can customize Claude Code for your workflows: Anthropic just shipped mods, small TypeScript add-ons that let you tweak how Claude Code looks and behaves. Build one yourself or ask Claude to do it, and it kicks in instantly. You can also replace built-in features like “/diff" with your own version. Keep in mind, mods aren't sandboxed, so stick to trusted sources. To get started, Anthropic's official guide walks you through prompting Claude to build your first mod, or you can even ask Opus 5.5 to mine your past sessions for mods worth building.
A new class of video model is here: AI startup Tavus just unveiled Griffin, a "Human Interaction Model" that watches, listens, and talks at once. It skips the usual relay from speech-to-text to an LLM to a voice model, letting it nod or jump in while you're speaking. Tavus says 48% of 54 testers mistook Griffin for a human, and it topped NVIDIA's VideoFDB benchmark. That convincing performance comes with deception risks, so Tavus is keeping it off its API for now. Request early access.
Three new decision-making models make their debut: Cloudflare just dropped Clef, a family of open-source models that pick from a fixed set of answers and tell your code how confident they are. Pricing starts at $0.24 per million input tokens, or $0.09 for Clef-flash. Perplexity’s Decisions API goes cheaper still at $0.04, with up to 250k tokens of context. Neither charges for output tokens. Amazon joined the mix the same day with Strands Decider 2B, small enough to run on modest local hardware.
PRESENTED BY AGENTFIELD
CodeAF is a new open-source coding harness on the Pareto frontier of cost and quality.
• On DeepSWE it solved 2× the issues OpenCode did, and nearly 4× what Claude Code solved, on the same open model.
• It paid 4.8× less per solved issue than OpenCode, and 3.4× less than Claude Code.
• Frontier-grade coding on open models like DeepSeek, Qwen, GLM and Kimi.
• No lock-in: it works with any provider you choose, closed models too.
Try CodeAF on GitHub
INSIGHT
Why the creator of Ruby on Rails has stopped coding manually
Source: The Code, Superhuman
Keyboard optional. Only five developers raised their hands when founder David Heinemeier Hansson asked a packed room who still writes code manually every week. During his Rails World 2026 keynote in Austin, the Ruby on Rails creator and 37signals co-founder shared that his team went "pencils down." AI agents handle the coding now, while human engineers step in only when those agents get stuck.
Death by a thousand PRs. Last spring, 37signals let designers vibe-code the final Basecamp 5 features. The pull requests looked reasonable on their own, but together they left holes in the architecture.
The Fix. That is where DHH's rule comes in. Whenever you step in, figure out why the agent failed and fix the workflow so it succeeds next time. That repair cycle drives the loop. If you skip it, problems quickly pile up.
The playbook for the loop:
Let the agent take the first pass. Make it your default for drafting code, and only touch the keyboard when it truly gets stuck.
Hand over whole tasks to the agent. Treat it like a coworker who can get on with work in the background, then review what it delivers.
Make every intervention count. If you have to hand-code a fix, ask what needs to change so the agent can handle it next time.
Eyes still required. Fixing those Basecamp 5 issues took someone who understood why the pieces were not fitting together. As you keep tuning that loop, you will find yourself writing less and less code. Still, it leaves the question DHH didn’t answer: if agents take over routine coding, how will new devs ever learn to tell good work from bad?
IN THE KNOW
What’s trending on socials and headlines
Meme of the day.
Alignment Check: A senior SpaceXAI engineer shared the one-line prompt she uses to catch AI misunderstandings before they become wrong code (11K bookmarks).
Readable AI: Andrej Karpathy shared his personal tricks for making LLM outputs easier to understand, including one obscure formatting spec he swears by (1.3M views).
Always-On Agents: OpenAI's 9-minute demo of "dots" shows why your AI may never need you to open a laptop again (2.9M views).
Spider Code: A JS + CSS animation of web crawlers skittering across a page about actual spiders is winning the internet (1.5M views).
Mods Demo: Claude Code now supports mods, and one dev already vibe-coded a spinner that turns his agent's work into live pixel-art cartoons (1.9K likes).
Card Killer: iOS 27 just gave Apple Wallet a feature that could finally empty out your physical wallet for good (18K likes).
TOP & TRENDING RESOURCES
Click here to watch the tutorial.
Top Tutorial
Tokenmaxxing your Claude subscription: A CTO shows how to hand Claude Code bigger chunks of real work without chewing through your limits too fast. You’ll learn how to structure longer agent runs, parallelize tasks, and treat Claude more like a collaborator than a glorified autocomplete.
Top Tool
Upsy: Give Claude, ChatGPT, Cursor, and other assistants their own shared cloud computer. They can browse with your saved logins, work in separate tabs, and keep going even when your laptop is closed.
Trending Cookbook
How to cut agent costs with model routing: Instead of using your most expensive model for every task, route simpler jobs to cheaper ones and save the strongest model for harder work. In one test, this cut median cost per thread by 64% without a measurable drop in quality.
AI CODING HACK
How to catch Claude Code solving the wrong problem
A dev shared a prompt block that catches agent misunderstandings before any code gets written. Hand Claude Code or Codex a long task, and it often solves the wrong problem. You find out only after it has touched a dozen files.
Step 1: Append this block to the end of your task prompt:
Before you touch anything, tell me in 2–3 sentences what you think I'm after and what problem we're solving. Start only after I say yes.
When there's a tradeoff, weigh all three:
→ UX: is it easy for users to use?
→ DX: is it easy for developers to change later?
→ AX: can the next agent understand it and keep going?
Don't touch anything outside this task, and don't break anything that already works. When you're done, tell me in 2–3 lines what you picked, what you gave up, and why.Step 2: Claude replies with a short readback of the task. Say yes, or correct it before it edits anything. When it finishes, it reports what it picked and what it traded off.
Pro tip: paste the block into CLAUDE.md (or AGENTS.md for Codex) so the readback runs on every task automatically.
P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.
IN CASE YOU MISSED IT
Our most-clicked story from yesterday
An ex-Microsoft and Google engineer turned his '90s principles into hard rules for Claude Code. His cost per PR dropped from $640 per feature to $1.
Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.
Whenever you're ready to dive deeper
We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.
What did you think of today's newsletter?
You can also reply directly to this email if you have suggestions, feedback, or questions.
Until next time — The Code team
