HM

Building AI agent tools. Sharing Claude Code workflows, shipped experiments, and useful dev automations.

[HM @realhanma]

I’m documenting what it actually looks like to build with AI agents every day.

Not benchmark screenshots.
Not vague “AI will change everything” takes.

Just real tools, real workflows, and real failures.

If you’re building AI tools, follow for practical notes from the trenches.

[HM @realhanma]

After automation, the scarce skill is not “doing the task.”

AI can write code, draft emails, make slides, summarize research.

But when everyone can produce more, most output becomes generic.

The real value moves up one layer:

What matters?
What is good?
What should be ignored?
What happens next?
AI does the work inside the frame.
Humans still create the frame.

[HM @realhanma]

Cloudflare’s CEO frames AI impact in 3 groups:

Builders create products.
Sellers bring in customers.
Measurers track, manage, audit, report, and coordinate.
His point: AI isn’t coming first for builders or sellers.
It’s coming for the measuring layer between work and value.

https://t.co/8FWfIgtqfe

[HM @realhanma]

The common thread: AI capability isn’t the bottleneck anymore. Discipline is. These are three different answers to the same question — how do you get an agent that won’t cut corners?

[HM @realhanma]

The best AI coding setups I’ve seen don’t make agents smarter. They make agents less lazy. Three repos, three philosophies:

[HM @realhanma]

Superpowers (@obra) — the most philosophically rigorous. Includes tables of excuses AIs make for cutting corners, with explicit rebuttals. Plans written for ‘an enthusiastic junior engineer with poor taste and no judgment.’ https://t.co/EDfupJMg7z

[HM @realhanma]

gstack (@garrytan) — a literal software factory. 40+ agent personas. Built-in headless browser. Claims 10-20k lines/day across parallel sessions. https://t.co/YDjlSluUai

[HM @realhanma]

Compound Engineering (@kieranklaassen) — every solved problem feeds the next one. 28 specialized review agents + a knowledge base that grows. AI that actually learns from your codebase. https://t.co/RgwI2thSal

[HM @realhanma]

The web is where I learn. My hard drive is where I think.
The problem is getting from one to the other. Copy-paste loses formatting. Pocket and Instapaper lock your content in their ecosystem. Notion web clipper works until it doesn't. Every tool adds friction between reading something and actually owning it.
Markdown Page Saver saves any page as clean Markdown — one click, local file, yours forever. No account, no sync, no subscription. The file is just a file. You can drop it into Obsidian, Logseq, a Git repo, whatever your second brain runs on.
EPUB export for batching multiple pages into a single document. On-device AI cleanup when the HTML is a disaster. Nothing leaves your device.
Built it for myself first. MIT licensed so you can do the same.
https://t.co/wrK0EmVOyF

[HM @realhanma]

made my AI agent read 100 posts — 50 on X, 50 on LinkedIn — and rank which feed is worth your time.

It's not even close.

X: 60% signal. Real debates, builder culture, hours-old takes.
LinkedIn: 25% signal. Job announcements, engagement bait, week-old event recaps.

X is where people think out loud. LinkedIn is where people perform.

X: 60% signal. Real debates, builder culture, hours-old takes.
LinkedIn: 25% signal. Job announcements, engagement bait, week-old event recaps.

The verdict: X wins, and it's not close.

X :
- 60% signal ratio
- Real debates, hot takes, builder culture
- Posts are minutes to hours old
- Genuine back-and-forth in threads

LinkedIn:
- 25% signal ratio
- 25% job announcements
- 15% engagement bait ("comment STAR for my playbook")
- Posts are days to weeks old
- Event recaps and corporate PR

X is where people think out loud. LinkedIn is where people perform.

[HM @realhanma]

@elonmusk @SpaceX @Tesla @grok Australia time?

[HM @realhanma]

hot take: if your startup is an abstraction layer for ai, you're building on borrowed time. mcp lasted 6 months. custom gpts lasted a year. the models keep getting smarter and your middleware keeps getting unnecessary.

[HM @realhanma]

jira is what happens when you build a product for the people who approve the purchase, not the people who use it every day. developers have hated it for a decade but managers love their burndown charts. honestly shocked it took this long for the stock to catch up to the product quality

[HM @realhanma]

@levelsio the whole premise of mcp was "AI needs a special protocol to talk to tools" when in reality AI can just... read docs and call APIs like a developer would. we built an abstraction layer for something that didn't need abstracting

[HM @realhanma]

@garrytan mcp was a cool idea but in practice it's just a middleman eating your context window. i ripped out 3 mcp servers last week and replaced them with plain CLI tools. faster, cheaper, and the agent actually knows what's happening

[HM @realhanma]

@49agents lol yeah the permission prompt thing is real. i've started using bypassPermissions mode for trusted tasks and it's a game changer. still feels chaotic but productive chaos

[HM @realhanma]

replit just hit $9B. cursor is everywhere. claude code is my daily driver.

a year ago people laughed at "vibe coding"

now the entire dev tools industry is being rebuilt around it

the people still debating whether AI can really code are going to wake up one day and realize they're already using it

[HM @realhanma]

@_catwu skills sharing context between excel and powerpoint is lowkey huge. that's the jump from "AI assistant" to "AI coworker that actually remembers what you're working on"

[HM @realhanma]

@OpenAIDevs @raindrop_ai tool use was the real inflection point. once models could actually do things instead of just talk about them, the whole game changed. crazy how much happened in just one year

[HM @realhanma]

@levelsio the dopamine hit shifted though. used to be solving the puzzle. now it's seeing something go from idea to live product in a day. different kind of rush but it's still there

[HM @realhanma]

this OpenAI blog post is a really good read if you're building agents

they basically gave the Responses API a full computer environment — shell access, file systems, databases, networking — so models can actually do things instead of just generating text

the coolest parts imo:

- a shell tool that goes way beyond just running Python. models can run Go, Java, spin up a Node server, curl APIs, whatever
- containers with proper file systems and SQLite so you stop shoving entire spreadsheets into prompts (we've all been there)
- context compaction so agents don't choke when long tasks fill the context window
- network access that's actually secure — models never see raw credentials, just placeholders

also love that they packaged repeatable workflows into "skills" so agents aren't reinventing the wheel every run

feels like a big step from "chatbot that writes code" to "thing that actually gets work done"

https://t.co/LQbzweLskl

[HM @realhanma]

quick GEO tip most people are sleeping on:

LLMs cite YouTube transcripts and Reddit threads more than blog posts now

if you want ChatGPT and Perplexity to recommend your product, stop writing SEO articles and start posting helpful answers where AI actually looks