A new stealth model, polaris-alpha, has appeared on openrouter. Time to give it a try.

My earlier take is deprecated.

I've updated today the tanstack starter project + agent rules. You can build now quickly web apps with one the best web dev stack available.

f you use composer-1 in cursor for coding, it's helpful to run a quick /code-review to identify any code smells. In my case I use a slash command with gpt-5. Usually you find some flaws in the final implementation.

Finally I found a way to 10x codex! Sound on! 🔊

Cursor 2.0 update and their new agentic model "composer-1-alpha" I've been testing it for a few days and have put together a full summary in the thread below. Click to see whats new 🧵

This script plays a sound when Codex finishes your task.

If you work in a larger company and your team uses Jira, I can recommend the "mcp-atlassian" MCP. It's probably the only MCP I'm currently using. I fetch all stories for a given epic in markdown, including the specification stored in confluence and export the figma screens.

I've put together a collection of slash commands I use with Codex and Cursor. ⚡ Link in the comments

The new Plan mode in Cursor is well thought out. It builds on the existing task lists, showing they iterate on what’s already there and bring features together instead of shipping separate ones. I like that approach.

Cheetah is such a beast model. refactor to a monorepo, including smoke tests & CLI updates, in just a few minutes.

Codex CLI Model selector looks clean now. It seems that the configs for model_reasoning_effort (low, medium, high) are still available.

Look, the final boss spawned in Cursor.

A bit harsh but this is how i treat Claude plans when I hand them over to codex for implementation.

Loving the speed of the cheetah model in Cursor! Still, it's not flawless. It took 5 iterations to get green light from Codex to implement the plan. But first impressions are positive. Worth testing further within my workflow as a fast Q&A, planning, and peer review model alongside gpt Codex.

The new "cheetah" stealth model is the first model that is not free. Which probably means it comes from one of the major AI labs. If we follow the usual release loop of models, it fits for Gemini 3. And since it’s pretty fast, it feels more like a Flash model than the Pro one.

New Cursor 1.7 update is out 1/ Agent Autocomplete for prompts 2/ Hooks to control & extend the Agent loop 3/ Team Rules set global rules in the dashboard for all projects. Bugbot now follows team rules too. 4/ Share reusable prompts via links for docs 5/ Checking Agent Status with Menubar Monitor

Hacking the Agent loop with your own hooks is a feature I wished we had months ago. It's really cool that it finally made it into the new Cursor Version.

When using icon libraries for AI Coding i do the following. 1. First, I'm using Iconify, which is a developer friendly icon set framework. 🧵👇

New look, clearer Cursor.

It takes one day with Codex CLI to build your own infra that deploys a full-stack app (TanStack Start, Postgres, S3) on a €3.29 Hetzner Server. Add Terraform, Ansible, Dokku and manage it with your own CLI.

There are many rumors about what "code-supernova" is. Some say it could be another Grok model or even Claude 4.5. But Anthropic has never released a stealth model, and xAI published grok-code-fast-1 (sonic) recently.

OpenAI released GPT‑5-Codex today. - Install: cargo install --git https://github .com/openai/codex.git --tag rust-v0.36.0 codex-cli - Adaptive thinking speed + depth: 93.7% fewer tokens on easy tasks (up to 10× faster) and 2× reasoning on the hardest

To enable web search for Codex, set [tools] web_search = true

Added a new special offer for Instructa AI students. 1 Month of Cursor for Free. Thanks Cursor Team!

These are the AGENT rules I'm using for my project - Tanstack / react / shadcn - Cloudflare / Alchemy - Drizzle - Linting

I've just released a new version of Codex CLI 1UP Link in the comment 👇

These are the rules I use with Codex CLI to make code search faster and more accurate.

First, use grok-code-fast-1 to find the files you need for the refactor. After that, use gpt-5 to plan and carry out the changes in Cursor. This method is faster than having gpt-5 search for context on its own. Bonus Tip: If you want to increase the quality.

Using GPT-5 and Claude Opus for peer reviews is a great option when you're not sure about a code change. It feels like having a skilled developer give you a second opinion.

Try to keep an eye out for the stealth models, either in Cursor/Cline (like "sonic" now) or on OpenRouter These models are preview models, but they’re often really powerful, fast, and FREE. sonic -> grok code (?) horizon-alpha/beta -> Early GPT-5 quasar -> GPT-4.1

Cursor added a new stealth model called "Sonic."

Here are my quick thoughts on GPT-5: (after a weekend coding) Chat GPT-5: At this point, it's like a calculator... ask, then get an answer. Pragmatic, but not very relevant for coding.

Launch ChatGPT 5 with a prompt directly from your desktop. (Raycast) Press ⌘ + SPACE, select a model, and start asking. Link in the comments.

Get Next.js console errors live from the browser to fix them Cursor.

Here is my August Tool Stack for complex tasks and large codebases. 1. Identify a problem or task you want to solve. 2. Research, plan, and define the tech stack (GPT-5 Pro + Deep Research + GitHub). 3. Implement the task step by step using batch prompting (GPT-5 Pro).

One thing I've observed with Agentic Coding and new features is backward compatibility and fallbacks. GPT-5 and Claude behave the same way.

GPT-5: Like a waiter who delivers exactly the antipasti and wine you ordered. Nothing more.

Cursor CLI Updates shipped - Reference any files/folders with @ for better context - New Review Mode: lets you clearly see what's added and removed in a simple diff view. - You can type any follow-up instructions directly.

This is a prompt I built some time ago to split PRDs into smaller tasks. Feel free to use it.

The new Cursor update 1.4 is rolling out. - Interactive Agent Control: Queue (⌥ + Enter) or interrupt (⌘ + Enter) commands during execution. - Smarter Agent tools for huge codebases: full-file reads, tree lists, sharper grep & search. (Less tokens)

Why is o3 69.1 === 4o 30.8 ?

So, GPT-5 hasn’t solved the "purple problem"

Just tried out the new GLM-4.5 Model in Cursor. First impressions are good, the agent support isn't that good, though it's quite slow and dies multiple times. Maybe using the official API would be better. - 128k context window, - $0.60/$2.20 per mio tokens - Open Source

If you're working on a feature in Cursor and keep needing the same files, Notepads can be a big help. They've been around since the beginning of Cursor.... maybe forgotten by now, but still incredibly useful.

Just uploaded new lessons on Instructa .AI 🎓 - How to build with Cursor Background Agents - How to use Cursor and Slack together - Use Bugbot to find Bugs early - CMD+K (Update) - How Cursor's New Pricing Works

Current Aug Solo AI Tool Stack for Building Projects in Cursor - Starter for a solid base - Architecture doc with sequence diagram - Logger for all terminal output - Relevant .md docs in project - Minimal Cursor rules - Include SDK/Demos for reference - Use Git - Start new Chat ASAP

This video isn't sped up. The Horizon Beta Stealth model is just insanely fast.

Wow! Claude made my app 66% faster. Until I realized it added DEV_MODE=true flag which skips 50% of the slow parts...

I didn't try grok 4 heavy yet, but o3-pro continues to be the most reliable model for writing accurate code and tackling specific tasks in complex codebases. Gemini is still helpful for planning, though it often makes odd or unexpected assumptions.