Kyle Herbel

Working on agentic AI and 1-person companies
Founder @cortextools

[Kyle Herbel @KyleHerbel]

OpenAI just dropped GPT Image 2.0.

And it buried Nano Banana.

UI screenshots. Handwritten essays. Science infographics. Magazine collages. World War 2 newspapers with 100% readable text. Coffee machine diagrams. Figma slides recreated cleaner than the original.

The three biggest image gen failures — text artifacts, face drift, layout hallucinations — are gone in one release.

Paste your photo. Change your shirt. Face stays identical.

Rebuild an ad creative in one prompt. Zero AI slop.

Generate an entire newspaper. Every word readable.

Clean render in 29 seconds. No watermark.

This is the cleanest image gen model ever shipped.

[Kyle Herbel @KyleHerbel]

The new Claude Code desktop update changed everything

1. The activity chart tracks daily usage, streak, and which models you're burning tokens on
2. You can now run sessions locally, in the cloud, or over SSH from the same dropdown
3. Routines let you schedule 24/7 workflows that run even when your screen is off
4. Model + thinking effort are now front and center, not buried in settings
5. Connectors and Skills live under one Customize button

• Routines replace the old "keep Claude Code open forever" hack people were running
• The activity chart doesn't backfill, so your streak starts today
• If you live in the terminal like I do, none of this changes your flow — but the desktop app is finally worth opening

The UI caught up to how people were already using it.

[Kyle Herbel @KyleHerbel]

Most people using Claude Co-work are doing it completely wrong 🤯

One workspace setup → it reads your sales transcripts, builds color-coded Excel trackers, writes follow-up emails, and pulls your daily brief from Gmail and Google Calendar automatically.

All inside Claude Co-work.

Perfect for agency owners and freelancers who are still manually reviewing call notes, building spreadsheets by hand, and checking their inbox first thing every morning.

If you're hopping between Zoom transcripts, copy-pasting into spreadsheets, writing recap emails one by one, and still opening Gmail to figure out what matters today...

This workflow kills the entire loop:

→ Set up a folder hierarchy (context, projects, outputs) that gives Co-work persistent memory
→ Write a CLAUDE.md file that tells Co-work exactly how to route your work
→ Dump 6 sales call transcripts and get a full prospect tracker with conditional formatting in 3 minutes
→ Connect Gmail and Google Calendar to generate a daily brief without opening your inbox
→ Build custom skills so it writes in your exact voice every time

[Kyle Herbel @KyleHerbel]

Anthropic just shipped a full agent builder inside their console.

And it actually works.

Managed agents. Quick start templates. Environment configs. Credential vaults. Session viewers with full debug logs.

You give it a prompt, it generates your agent config, builds the environment, and lets you test it in one flow.

Deep researcher templates. Support agents. Data analysts. All pre-built and ready to run.

Custom agents from scratch with API integrations baked into the system prompt.

Full session playback showing every action your agent took.

Built-in log analysis that can identify errors, trace conversation flow, and suggest improvements automatically.

This is the fastest path from idea to working AI agent that exists right now.

[Kyle Herbel @KyleHerbel]

Claude Code Just made /plan 10x more powerful

Everyone's complaining that Claude Code builds ugly websites and generic apps.

Yeah, because you're using /plan like it's 2024.

There's a feature sitting right inside your terminal called ultra plan that most people don't even know exists. You type it in, rainbow text pops up, and it launches a completely different planning workflow.

Regular /plan mode just takes your prompt and runs with it. No questions. No design thinking. No structure. It guesses what you want and gives you the most boilerplate version of it possible.

Ultra Plan does the opposite. It spins up a separate session, asks you clarifying questions one by one, builds out a full plan with design direction, technical specs, and a verification checklist — and then lets you review and approve everything on a web interface before it writes a single line of code. The output isn't even comparable. Same prompt, same model. One gives you a generic template that looks like every other vibe-coded site on the internet. The other gives you animations, custom styling, and a layout that actually looks intentional.

The catch: it only works in the CLI right now. Not the desktop app. Not the web version. You need a terminal open.

It also eats more usage — I hit 18% on Max just from the initial build. But if you're building something from scratch, that tradeoff is worth it every time.

[Kyle Herbel @KyleHerbel]

How to turn Obsidian into a Claude Code-powered second brain in about 10 minutes.

Karpathy tweeted the idea last week. I've been using it nonstop since.

1. Download Obsidian and create a new vault
2. Inside the vault, make three folders: raw, wiki, output
3. Install the Obsidian Web Clipper extension for Chrome
4. In Clipper settings, change the default note location from "clippings" to "raw"
5. Install the Local Images Plus community plugin so clipped images come along
6. cd into your vault in the terminal and launch Claude Code
7. Start clipping everything — tweets, articles, docs — straight into raw
8. When you need synthesis, tell Claude: "create a wiki about [topic]" and it reads the whole raw folder and builds a structured index in the wiki folder

Few keys here:

• The raw folder is the magic. Don't organize it. Don't subdivide it. Just dump.
• Wiki folder is where Claude writes synthesized output. Don't put your own notes there.
• Output folder is for what you create from the research — scripts, drafts, posts.
• One vault per context. Business, client, project. Don't mix them.
• A CLAUDE.md at the root of the vault is what makes the whole thing behave.

[Kyle Herbel @KyleHerbel]

Anthropic BANS subscribers that have been using their Claude subscriptions with OpenClaw

[Kyle Herbel @KyleHerbel]

In all seriousness, Claude code is the easy and way more user friendly option for OpenClaw agentic use cases

[Kyle Herbel @KyleHerbel]

Codex revenue right now 🤑🚀

[Boris Cherny @bcherny]

Starting tomorrow at 12pm PT, Claude subscriptions will no longer cover usage on third-party tools like OpenClaw.

You can still use these tools with your Claude login via extra usage bundles (now available at a discount), or with a Claude API key.

[Kyle Herbel @KyleHerbel]

Codex now runs inside Claude Code.

Same terminal. Same project. No switching windows.

OpenAI shipped an open source plugin that lets you use GPT 5.4 directly inside your Claude Code workflow. No API key needed. Just your ChatGPT subscription.

Opus 4.6 and GPT 5.4 benchmark nearly identical on hallucination rates.

But Codex generates cleaner, more reliable output on the first pass.

Claude gets more creative but sometimes overthinks itself into bugs.

Now you don't have to pick one.

Run Claude for the creative heavy lifting, invoke /codex rescue when something breaks, and let GPT 5.4 clean it up without ever leaving your terminal.

Two frontier models. One workflow. Zero context switching.

[Kyle Herbel @KyleHerbel]

@AlexFinn Dude people just don’t get it yet. “But SeCruIty IsNt GoOd 😡”. They can’t help but complain instead of just building cool shit and moving forward

[Kyle Herbel @KyleHerbel]

Who is still using Microsoft 365 (besides outlook) 🥴

[Claude @claudeai]

Microsoft 365 connectors are now available on every Claude plan.

Connect Outlook, OneDrive, and SharePoint to bring your email, docs, and files into the conversation.

Get started here: https://t.co/EdoQeT8BBN

[Kyle Herbel @KyleHerbel]

Claude did all of this 6 months ago, I don’t think many people are planning on switching :/

[Cursor @cursor_ai]

We’re introducing Cursor 3. It is simpler, more powerful, and built for a world where all code is written by agents, while keeping the depth of a development environment.

[Kyle Herbel @KyleHerbel]

@rezoundous Bing AI

[Kyle Herbel @KyleHerbel]

Most of the people I talk to on a daily basis inside of my community who are just aren’t use Claude code or very new to AI programming don’t seem to understand even what context is let alone how it gets eaten up by different things. seems that a large majority of the people don’t understand the core concepts.

[Kyle Herbel @KyleHerbel]

Having the ability to locally host a AI model built by the most powerful company on earth is actually insane

[Google DeepMind @GoogleDeepMind]

Meet Gemma 4: our new family of open models you can run on your own hardware.

Built for advanced reasoning and agentic workflows, we’re releasing them under an Apache 2.0 license. Here’s what’s new 🧵

[Kyle Herbel @KyleHerbel]

Claude Code has a hidden pet system and most people don't know it exists yet.

Type one command, get a coding companion that lives in your terminal and reacts to what you build.

Here's how to set it up:

1. Open Claude Code in your terminal
2. Type /buddy and hit enter
3. Watch it search and hatch your companion
4. Check its stats — debugging, patience, chaos, wisdom, snark
5. Start building something
6. Say your buddy's name to get its take on your code

Few keys here:

• You won't get a rare one on your first try. Most people get common. Keep regenerating if you want something better
• It doesn't eat your tokens. Runs entirely client-side, no API calls
• There are rarity tiers — common through legendary — and the rarer ones have wilder stats and different creature types
• Axolotls, ghosts, robots, octopuses are all in the pool

Go hatch something!

[Kyle Herbel @KyleHerbel]

Built a voice AI advisor into a website using Google's new Gemini 3.1 Live model

One prompt → a skincare site where customers talk to an AI that listens to their skin concerns and recommends specific products from the catalog. Live voice. Real-time responses.

All inside Claude Code.

Perfect for e-commerce brands and agencies who want to add a voice shopping experience without building it from scratch.

→ Real-time voice conversations with sub-second response times
→ Webcam input so the AI can actually see what the customer shows it
→ Screen sharing support for walkthroughs and demos
→ Works through Google AI Studio with a standard API key
→ Embeds directly into any website with a few API calls
→ Identified a specific Rolex model from a webcam feed mid-conversation

Crazy times
https://t.co/aO0rfGTrWr