Once the tools all get figured out, the marketplace for skills, subagents, workflows, etc is going to be so enormous. Today you can create them and share them, but you can’t really measure or improve them - there’s no way to really evaluate other than vibes and your own direct trial/error.

Thoughts on this terminal/cli agent interaction? Personally ... I'm obsessed.

How no one hasn't come up with a better approach for feeding/providing an LLM access to frontend/backend logs is beyond me. Please tell me someone has a great solution for this?

It’s great Anthropic is publishing these articles - the best practices for agentic coding are spread all over the nerdiest corners of the internet.

After several full days working with Opus 4.5, i say this without hyperbole. We are in a completely different world now. Product development has changed. I truly believe that.

Just to pile on --- Opus 4.5 is insanely good. What a release from Anthropic. Crazy we got 5.1, 5.1 Codex Max, Gemini 3 Pro, AND Opus 4.5 in like 1 week.

Bringing my kids drawings to life with AI is so much fun and the kids absolutely love it. My daughter’s “land of Sinba” - she’s creating a whole story.

Fun facts We named the company Codename because we *love* the thrill of new projects with a great code America New June’s @newjunehq codename was Hogan. The iOS app will forever be called Hogan. Iykyk

My obsession this weekend is figuring out how to create an agent agnostic framework/system for skills, subagents, slash commands (human initiated + agent initiated), and I guess plugins. I love Claude Code. It’s my fave.

Spec driven development is about “alignment” not about creating a perfect blueprint. Llm knows what needs to be built from spec, you align with the LLM on the design to deliver that, tasks keep the LLM on track.

Finally getting around to messing around with Claude Skills (i know i know I've been a little busy) and wow. I've kind of "hacked" this with subagents -- we have a test-lead agent for example with a .md stuffed with golden test examples and best practices.

As someone actively working on AI features that even foundation models struggle to get consistently right, features that are critical to our “free” product @newjunehq… Kimi K2 Thinking is so wild. Overnight our cost of “free” cuts in half. Less depending on how you host/deliver the model/tokens.

How to tell someone lives all day in Claude Code - thinking in Context Windows as a measure of complexity

Been playing around with a few memory tools and plugins for Claude Code. I've never been more convinced that the UI for Agentic Coding needs to change dramatically. - is memory enabled? - what does it remember? - can i control when it remembers?

Its been said before, but more people need to realize that LLMs are mirrors. They take what you say and expand upon it. Very easy to silo yourself in a direction that you yourself are biased towards. You need to explicitly ask for counter arguments, options, different perspectives.

Building @newjunehq is such a roller coaster sometimes. Some days we're like "f*** yeah this works perfectly" then the next day we're like "f*** there is no way we can make this work" and then the next day we're like "f*** yeah this works 10x better than I thought it would" repeat.

I've been working out of coffee shops more frequently. Great people watching. A lot of people bring computer monitors! I was very surprised by this. But on days I spend more time here - I see like 3 or 4 people that do this!

Having an "only with AI experience" and its really cool I have to share. We're working on our @newjunehq shot detection/prediction algorithm. The concept of working with a highly capable engineer, that also deeply understands the nuances of golf and can reason about...

OMG i've never been so happy. Claude Code startup is basically instant now! Thank god... i was dealing with literal 15-20s startup times. I've been submitting bugs and seems like they are really paying attention to these. Incredible work from the Claude Code team.

Super smart of Cursor to be focused on speed and innovating on the editor UX. I think speed is really a top painpoint. Current CC/Codex models straddle between sync work and async work. I'd rather be either FULLY sync or FULLY async. cursor.com/blog/2-0

Have been chatting with some early stage investors -- been super helpful forcing function to work on the @newjunehq pitch. I had it all in my head but having to explain the whole value prop AND strategy AND risks to other humans live - different beast entirely.

A current fave prompt “You are a Sr Principal Engineer. The owner of this plan took suspiciously long to author it.

I wish Claude Code would give estimates in tokens instead of weeks or days.

man great icons make such a big difference. really loving huge icons. @huge_icons

3 "holy shit" moments w/AI yest. 1 @tomf009's first attempt at merging a bunch of context into @newjunehq's shot prediction "oracle" nearly perfectly predicted shots on a given hole from scattered context.

The next big dev workflow bottleneck for us I think is UX workflow. We haven't had much of a process, we basically 1. provide some initial direction (written spec) 2. pulling the AI UX slot machine 3. spend way too much time iterating

We’re experimenting a lot to go faster with AI while maintaining test coverage & clean strict lints. Goal is to ship insanely fast and have zero fear of breaking anything.

Legit, we're going to need Claude the Code Reviewer to get harsher and look deeper. This is getting automatic now with Sonnet 4.5/Codex and our established workflows! 🤣

Using Haiku in my /handoff and /resume_task slash commands is so amazing -- so much faster to transition across threads and maintain context.

I say this with complete seriousness. I’m starting an AI Native band.

I wish Slack had a "posts" format where you could turn a given channel more into a threads/x/LinkedIn style feed. We have "#{user}-ramblings" channel, but the chat/messages format makes it hard to visually distinguish between separate topics. Reverse chron feed makes such a big difference

Haiku 4.5 is not as good at calling subagents, which is a shame! Probably will stick to Sonnet 4.5 and then have dedicated subagents/prompts that use Haiku 4.5 vs having the primary model in the primary agent thread by Haiku.

I have never had more fun in my entire life than I am working on @newjunehq . Every single day I cannot wait to get to my desk. I truly feel like I was made for this. It took me a while to get here, and that's a long story, but I'm here. I do not intend to leave.

Haiku 4.5 comes out, started a single task and immediately hit my weekly limit 🫠 Resets tonight but I have so much fomo. Looks super fast.

This is how a simple way to think about building products when AI is getting so smart so fast: When intelligence becomes infinite, context is the constraint.

Codex and Factory Droid are great, but if you care about building systems to scale agentic coding - Claude Code is just second to none and keeps getting better.

Its been fun working with @tomf009 on @newjunehq. He started off more of a skeptic when it comes to agentic coding, and now he is a complete convert. Insane to watch an already incredible engineer figure out how to accelerate himself. Its wild.

Big News! I left Meta last month to start a company with my brother, Tom Fernandez (ex-Netflix / Amazon Principal Eng … and for those that know us personally — finally 😂) Our company is Codename, and the product we’re building is New June @newjunehq

The current pricing for AI subscriptions (for coding) looks exactly like smartphone era. • data plans with different tiers • “credits” to explain tokens like data usage What we haven’t seen yet: • overages, data packs, top ups • throttling heavy users • “free” use cases like summarization/chat

Tried Factory Droid CLI out a bunch this weekend. tldr; really promising, but not ready to replace Claude Code. It really could though -- possibly soon. pros/neutrals/cons 👇

This single principle is why spec driven development and context engineering are killing it right now. Scope ambiguity = death Simple solve (1) have the agent ask you questions to eliminate ambiguity, (2) read the specs/plans line by line until you are 100% aligned

Someone has to be building the model agnostic Claude Agent SDK right? Mastra AI should it would kill.

I hit my Claude Max weekly usage limit in ~36hrs. I happened to be churning through a ton of tokens this week getting our test coverage up + big refactors before a big push. But work doesn't stop when CC stops, so I'm off to spend 5 days minimum with OpenAI Codex.

Help - I need your best advice on getting LLMs to write and fix tests! I am just absolutely churning through tokens and it works but it seems beyond inefficient.

1. Bonkers 2. Right now, this is the worst this will ever be openai.com/index/sora-2/

2026 - Multi Agent UI is going to be a big deal. If you are living in Claude Code/Codex you learn quickly the UI is a limiting factor on productivity.

At this point one of the main reasons I’m reaching for Codex over Claude Code is that CC takes like 10s or more to start and Codex is instant. So painful.

Not you too new Alexa!!!!

If you've been dabbling with agentic coding and flirting with starting a real project, I cannot recommend the following guardrails enough. ⚠️ warning ⚠️ This will feel like overkill. BUT - you will never have to feel the pain of a AI causing cascading regressions.

A new skill when working with Agents is to identify when the agent *could* be doing helpful work but isn't, and then build prompts & automations to try and get idle/unproductive time as close to zero as possible. One example: