Welcome back. Just over a year ago, Open AI’s o1 Pro cost up to $600 per million output tokens. Today, GLM-5.3 Flash costs a tiny fraction of that and actually outperforms it. There is no better time to build and ship the side projects we always thought of. In this issue, we cover insights from a YC talk on harness and how top AI labs might benchmark for high scores on paper.
Also: How Spotify engineers cut Claude Code token usage by 90%, a guide to leveling up your prompts for Astra, and read why the creator of C++ hates AI-generated code.
Today’s Brief
Master the two things models cannot out-code
How a founder can screw over your equity
How Stripe built its company brain (tutorial)
OpenAI’s tricks to get more out of GPT-6 Astra (cookbook)
TODAY IN PROGRAMMING
Watch Y Combinator’s full talk here.
The best AI upgrade might be your harness: A recent Y Combinator talk shifted the focus from the model to the harness. A harness is basically the scaffolding that connects a model to tools, memory, and integrations. Researchers found that improving this layer can alone push reasoning scores higher than upgrading your AI model. Some harnesses can even rewrite their own code. Even senior engineer Uncle Bob Martin admits he may be over-constraining his agents. Better systems could unlock much more from today’s models.
Harder benchmark exposes models with inflated scores: Artificial Analysis replaced Terminal-Bench 2.1 with a much tougher 4.0 test, and some scores cratered. While GPT-6 Astra and Claude Fable 5.1 stayed on top, Gemini 3.8 Flash fell from 89% to 19% and Muse Spark 1.3 dropped to 33%. SemiAnalysis calls this “benchmaxxing” which is basically when a model excels at one benchmark without being nearly as strong at general work. Meta’s Chief AI Officer Alexandr Wang pushed back, saying a big drop could simply mean the new benchmark is harder. Either way, it's a good reminder not to trust one leaderboard too much.
Devs can’t agree on how much of the codebase you should know: Senior dev Theo argued that real software eventually gets too big for one person to fully understand. His post went viral, and devs had plenty to say. Some agreed while others called it lazy engineering. Former Google engineer Arpit Bhayani pushed back, warning that as devs move faster and rely more on abstractions, they may understand their systems less. That’s a huge liability when things break and nobody knows where to look.
PRESENTED BY WISPR
Your dev stack got an AI upgrade everywhere except the input layer. You're still typing every prompt, every ticket, every review comment by hand.
Wispr Flow closes that gap. Dictate into Cursor, VS Code, Slack, Linear, or anywhere else you work. It's syntax-aware: camelCase, snake_case, acronyms, and file names all come through clean. Mention a file in Cursor or Windsurf, and it auto-tags.
It's the voice layer for an AI-native workflow. Speak your intent. Your tools do the rest.
Available on Mac, Windows, iPhone, and Android. Used by millions of developers, including teams at OpenAI and Mercury.
INSIGHT
The two things models cannot out-code
Source: The Code, Superhuman
The target keeps moving. In a viral essay, Sean Goedecke, staff software engineer at GitHub, argues that trying to stay technically ahead of AI is a losing game. Work that felt safe at the start of the year is already being automated. Refactoring huge codebases is a prime example. Hard engineering might buy you some time, but it won't hold forever. If models can tackle the Riemann hypothesis (a 167-year-old mystery about how prime numbers are distributed), low-level code isn't much of a feat.
Get better. Goedecke advises focusing on what models still do poorly, i.e., understanding context and communicating clearly. He splits this into two irreplaceable skills:
Know the system. Agents hallucinate less now, but they still lack context. They rebuild existing things, edit the wrong code, or turn simple fixes into over-engineered messes. To catch these anomalies, you need to know the whole codebase, not just the ticket on the screen.
Explain it well. Model writing often feels stiff and generic, and researchers struggle to benchmark good writing. Meanwhile, people tune out anything that reads like AI faster than ever. The value of human explanation only keeps increasing.
The edge. The durable skill is knowing the system better than the model does, then explaining what matters in a way people actually understand. Goedecke's guide to avoiding the common mistakes most people make in large codebases is a great place to start.
PRESENTED BY TRIGGER. DEV
Trigger. dev runs your AI agents as durable tasks inside your codebase, in plain TypeScript with full control of the runtime. No serverless timeouts, so even long agents run to completion, surviving refreshes, redeploys, and crashes along the way.
Write your logic, test it, and deploy agents that scale on demand.
IN THE KNOW
What’s trending on socials and headlines
Meme of the day.
Token Tricks: Spotify's engineers cut Claude Code token usage by 90% with a two-model routing setup, and their breakdown explains exactly how (9.7K bookmarks).
Prompt Refresh: Your old agent instructions might actually be hurting GPT-6 Astra. This guide explains which skills and AGENTS.md rules you should delete first (3.5M views).
For Founding Engineers: A VC's viral thread argues that a founder's character is what really decides a startup's fate, and this secret buyout story proves it (3.1 likes).
Hot Take: The creator of C++ just went viral for slamming AI-generated code. His argument for why humans are still essential is a must-read (1M views).
Marathon Model: A dev let GPT-6 Astra Ultra run for over 22 hours straight on a single project. What happened after the reset is the surprising part (3K likes).
Bet on it: OpenAI leadership famously refused to build internal tooling teams. This insider post reveals the logic behind that bet and why it actually worked (3.8K likes).
Design Crash Course: This video covers 95% of system design concepts, from the basics to handling 1M requests per second (34K likes).
TOP & TRENDING RESOURCES
Click here to watch the tutorial.
Top Tutorial
How Stripe built its company brain: In this tutorial, Stripe’s engineering leaders break down how they built Kai, the company’s internal agent platform. You’ll get a look at the system design, key technical decisions, and lessons learned from building an agent that can work across company knowledge and tools.
Top Repo
Dify (155K ⭐): One place to build AI agents, RAG pipelines, and multi-step workflows. It supports a huge range of models and tools, with visual workflow building and cloud
Trending Cookbook
How to get more out of GPT-6 Astra (by OpenAI): This guide walks through the best practices, new features, and migration advice for Astra. You’ll learn how to update your agent instructions and workflows for the new model, including features like async tool calling that let Astra keep working while other tools are still running.
AI CODING HACK
How to make Claude Code write PR descriptions reviewers actually read
Claude Code writes pull request bodies like essays. Walls of prose, "I ran tests" filler, and commentary on intermediate commits nobody cares about. An OpenCode engineer shared a skill that fixes it.
Step 1: Create the skill file:
mkdir -p ~/.claude/skills/writing-prStep 2: Save this frontmatter at the top of ~/.claude/skills/writing-pr/SKILL.md
---
name: writing-pr
description: Use when writing or editing a pull request title or body.
---Step 3: Then add the rules below it:
Don't write essays. Don't mention that you ran tests. Write a concise body.
Use bullet points, code snippets, and Mermaid diagrams.
For visual changes, show a before/after table with images.
For benchmarks, show before/after tables (baseline from target branch).
Skip intermediate PR details. Only the final squash commit matters.
For large or high-risk changes, write it like a technical blog post.Claude loads it automatically whenever it writes a PR. Skills also work in project repos at “.claude/skills/”, so the whole team gets the same PR style.
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
Amp Code’s co-founder shared an internal Slack post about what "owning" a problem in the age of AI actually looks like. It’s a must-read for all developers.
Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ 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
