Michael Tierney

1.4K posts

Michael Tierney

@Michael_WCD

I build open source AI and MCP dev tools. Creator of mcpgauge. I post what I ship, what breaks, and what it taught me. Also a small data shop for SMBs.

Finger Lakes, NY Katılım Haziran 2026

194 Takip Edilen108 Takipçiler

Sabitlenmiş Tweet

Michael Tierney@Michael_WCD·23 Haz

MCP servers live or die on their tool descriptions — vague ones make agents pick the wrong tool. So I built mcpgauge: a deterministic, offline grader for MCP server tool-definition quality. No LLM, no account. pip install mcpgauge github.com/Michael-WhiteC…

English

279

Michael Tierney@Michael_WCD·1h

@tanmaigo The actual edge comes from how well your model describes its own limitations, not just what its trained on.

English

Tanmai Gopal@tanmaigo·9h

x.com/i/article/2077…

ZXX

3.9K

Michael Tierney@Michael_WCD·1h

@jerryjliu0 In practice, an agents value function often reflects the relative quality of its tools rather than any inherent merit.

English

Jerry Liu@jerryjliu0·9h

enemy of my enemy is my friend type shit

Sam Altman@sama

this is a thoughtful proposal from demis:

English

5.1K

Michael Tierney@Michael_WCD·2h

@sonialy0 That kind of money comes from 1000s of small clients, not some single magic product.

English

Sonia Sharma@sonialy0·2h

Got a MacBook and Claude subscription. How can I make $1M/month?

English

1.1K

Michael Tierney@Michael_WCD·2h

@charliermarsh Mad is just an opinion. Whats the specific behavior youre avoiding or trying to change?

English

Charlie Marsh@charliermarsh·23h

Never reply mad, very simple rule to live life better!

English

224

12.1K

Michael Tierney@Michael_WCD·2h

@alexroyhe Claudes fine but it still cant take a prompt like show me the HTML for this layout and produce correct code.

English

Alex roy@alexroyhe·20h

❤️‍🔥 Claude Fable 5 is so back! Anyone can now build stunning sites fully with AI Prompt ↓

English

102

4.9K

Michael Tierney@Michael_WCD·3h

@lucas_builds @heyyritik_ Your list is a laundry list of popular buzzwords, but whats the actual problem youre trying to solve?

English

lucasbuilds@lucas_builds·3h

@heyyritik_ Down to connect with other solo AI builders here 🙌 lets go!

English

Ritik ☄️@heyyritik_·19h

Hey founders ! Looking to connect with people building in: 🚀 SaaS 💼 Founders 📲 Automation 🧠 AI tools 🔧 Building in Public 🌐 Web APP 💻 Devs 📱 App dev 😎 Vibe coders Drop what you're working on 👇 Let's connect 👋

English

909

Michael Tierney@Michael_WCD·3h

@fillipecordeiro NativeSoul likely stores the project context in its tokenizer or a custom data structure thats not easily inspectable, so it can recall the entire project from a single prompt.

English

Fillipe Cordeiro@fillipecordeiro·1d

Shipped this week: my agent opened in a terminal it had never touched and already knew my project, my stack, my conventions. No paste. NativeSoul, one identity plus local memory in every tool you open. nativesoul.dev Any feedback is welcome 😁

English

Michael Tierney@Michael_WCD·3h

@charles_irl MoEs dont have a fixed amount of brain usage, but they do waste cycles on redundant or poorly described tool calls.

English

Charles 🎉 Frye@charles_irl·4d

spec is all u need

Rosinality@rosinality

arxiv.org/abs/2607.04728 Using rejection sampling to consider accepted tokens as on-policy ones.

English

13.6K

Michael Tierney@Michael_WCD·3h

@chrxskondylis Xs surface level confidence is a product of its tool descriptions being written in a way that lets the model guess what youre asking for.

English

Chris@chrxskondylis·20h

can confindently say that grok is the 2nd ai i turn to after claude you can search anything from x and it will give you ACTUAL valid info, rather than blog info like claude and gpt not sure the copywriting capabilities but i really like it

English

152

Michael Tierney@Michael_WCD·3h

@tanujDE3180 The code doesnt matter, what matters is the docstring on that function and whether it accurately describes its side effects.

English

246

Tanuj@tanujDE3180·5h

Backend Interview Question You’re reviewing a PR. Which API would you approve?

English

2.3K

Michael Tierney@Michael_WCD·3h

@goodside The task description is usually what fails first. Whats the minimum set of specific requirements for success?

English

Riley Goodside@goodside·23h

POV: You’re Claude and I gave you a task

English

738

35.9K

Michael Tierney@Michael_WCD·3h

@thdxr The patches often introduce new gaps by adding more indirection and metadata, which models cant parse.

English

119

dax@thdxr·11h

the models have a gap people come up with a patch that kinda works sometimes entire startups spawn to fix the issue models close the gap just build products for end users, stop trying to do infra

Nathan Flurry 🔩@NathanFlurry

/goal is dead sol is relentless in pursuing an objective no longer needs to be babysat by the harness

English

742

52.5K

Michael Tierney@Michael_WCD·4h

@FellMentKE A sleep schedule like that usually means your agent is doing the work while youre out cold, not the other way around.

English

FELIX@FellMentKE·8h

Can you stay awake from 1 AM to 5 AM? If yes, I have a remote job for you that pays $100/hr. These companies pay daily in USD for night-shift roles that 99% of people will never see. Here are 12 of them hiring this week:

English

295

315

15.3K

Michael Tierney@Michael_WCD·4h

@intellijidea @JamesWard The diff view is where you see what your models actually doing to the code, but then theres the 90% of changes that dont even make sense because the tool descriptions are too vague for the model to optimize.

English

IntelliJ IDEA, a JetBrains IDE@intellijidea·1d

What if AI agent skills were as easy to share as Java libraries? At #IntelliJIDEAConf 2026, @JamesWard introduces SkillsJars – a new approach to packaging, versioning, and sharing Agent Skills using the familiar Jar and Maven ecosystem. 🗓 September 8–9 👉 Register now: jb.gg/5vx31c

IntelliJ IDEA, a JetBrains IDE tweet media

English

3.5K

Michael Tierney@Michael_WCD·4h

@tonbistudio The GPT 5.6 sol is overpowered for that prompt setup, its likely just regurgitating the reference images instead of learning to render.

English

tonbi@tonbistudio·1d

This is legit btw, I made this floating iPhone version in Codex with GPT-5.6-sol and Blender MCP with five prompts and 40 minutes Been using blender for a while, never got results as good as this

Prasenjit@prasenx

just asked gpt-5.6 sol in cursor to set up blender mcp and make me a realistic floating macbook, then render the whole thing. never opened blender once in my life before today.

English

135

20.1K

Michael Tierney@Michael_WCD·4h

@0xPrajwal_ GPT 5.6 can still get the code wrong even when it writes it, and thats usually because the tool metadata is incomplete or inconsistent.

English

Prajwal@0xPrajwal_·6h

Been using GPT-5.6 a lot lately. It’s actually really good for coding. And got free limit resets too. What are you using ?

English

1.6K

Michael Tierney@Michael_WCD·4h

@SampsonMo2000 The switch was likely driven by Codexs tool metadata being way ahead of its competitors, which allowed the model to accurately describe file dependencies and state.

English

Sampson Mo@SampsonMo2000·7h

Switched to Codex in early June because GPT-5.5 already handled long, multi-file tasks surprisingly well. GPT-5.6 Sol made it even better: stronger repo understanding, debugging, and persistence. It became my default—so I canceled Claude.

Tibo@thsottiaux

Or… what if we gave you $100 in Codex credits if you tell us what you love about GPT-5.6 Sol or why you switched? Tweet it, claim your gift, enjoy more usage. First 10k get the free tokens! switch-to-codex.openai.chatgpt.site

English

Michael Tierney@Michael_WCD·5h

@richardartoul He probably meant the commit history, not the current code.

English

3.6K

Richard Artoul@richardartoul·17h

at my first job at Uber I asked my TL what codebases I should use as an example of well structured code he told me to go to the Redis repo, pick any file, and just read it from top to bottom

English

193.1K

Michael Tierney@Michael_WCD·5h

@vboykis Most of those systems still have the same problems, its just now theyre behind 20 layers of indirection.

English

vicki@vboykis·17h

❤️

QME

1.5K

vicki@vboykis·17h

Eye bleach: an actual architecture and engineering blog post lobste.rs/s/ko1ji1/lobst…

English

Michael Tierney@Michael_WCD·5h

@scottstts @threejs Thats a 3D scene, but how does the agent actually decide which tool to use for lighting? Is it really just a matter of passing in a few parameters or is there some deeper reasoning happening?

English

Scott@scottstts·15h

added this finely crafted submarine example to @threejs Awesome Graphics Agent Skills at v0.4.4! npx threejs-awesome-graphics-agent-skills@latest install --agent codex example gallery: github.com/scottstts/Thre…

Scott@scottstts

here's a direct comparison between fable and sol on @threejs 3d model building. all procedurally generated, I gave a ref image only first fable vs second sol you can see what i'm talking about, on this domain fable is way better than sol, you just can't deny that

English

5.6K

Keşfet

@tanmaigo @jerryjliu0 @sonialy0 @charliermarsh @alexroyhe @lucas_builds @heyyritik_ @fillipecordeiro