Most annoying thing about this AI push. I've been arguing with PMs that we didn't need THAT much process since back in 2016..
Hard to feel vindicated reading all these think pieces saying the same thing now because AI "changed the game" no it just made y'alls inefficiencies more obvious.
The curse of building with LLMs is that I never find time to actually write about building with LLMs. I need a workflow or something.
I'd let an agent do my shopping if it sent me a 2FA style code for every purchase.
Being a worker under capitalism is basically getting yourself all slutty and standing out on a street corner until a VC dude rolls up and goes "do you love solving difficult challenges in a fast-paced environment?" and then you say "I'll love whatever you want for housing and healthcare, baby"
It’s actually a great time to care about the craft. I don’t know if people realize this but you can actually write whatever you want into the prompt box. You can use the AI to get the function just right instead of making it generate crap as fast as possible.
I think if you think in terms of workflows instead of "agents" you get a lot further.
Cloudflare tunnels are dope
people keep saying "having adhd is an advantage now"
no you idiots our attention span is being assaulted non stop and our brains are shriveling and can't focus for shit
and instead of confronting this you morons are like "maybe this is good actually"
God I love LLMs sometimes. Three days ago I sat down with a client to plan out our go live, we set a month timeline.
This morning we're essentially code complete. That means we get three weeks of hardening, testing, and making sure it works seamlessly.
I don't use claude code anymore but when I did the first thing I did was use tweakcc to change the system prompt.
There's so much unwanted behavior and wasted space in the default prompt.
github.com/Piebald-AI/t...
Parallel agents have one job in my workflow: documentation and planning. Meeting notes in. Structured tasks out. Easy to spot check.
The parallelism makes sense there because they don't need to agree or share context to be useful.
I built an agent orchestrator in a day. You can define agents, route tasks, and review what the agent is doing. Up next I'm going to add permissions and confirmations and that'll close out the table stakes workflow.
The client project we've actually decided to do a focused two sprint iteration before launch. Aiming to launch in a few weeks.
Everything else is too much builder energy, not enough sales / business energy.
A few projects in flight that I don't know how to close out.
Client Project
- Martial Arts Gym Platform
Personal / Possible SaaS
- Agent Orchestration System
- OpenClaw Deployments as a Service
- An agenic desktop os exploration
- A few custom youtube -> podcast workflows I built for myself
Oh maybe that the focus should be on systems architecture at this point. If you can make reliable black boxes that just take i/o, you're halfway to a factory for whatever you're making.
I have a lot of "make the thing that makes the thing" tools. I tend to ask "if LLMs disappeared tomorrow what would I be disappointed I didn't build?" And a lot of that is stuff to optimize my output for clients (along with toys and exploration).
Idk if obvious or not:
I don't use parallel agents in a single project.
Memory systems are overrated. I've yet to find one that beats proper code organization.
Maybe the self improvement aspect? I have a lot of specialized clis that agents have built for themselves to better manage my projects.
Personal software is underrated because it has no public vanity metrics. No users, no launch, no ARR. Just better mornings.
As someone building AI-native products, Exa + agent-browser has become a core part of my agent's stack.
Exa gets them the right docs fast, and agent-browser gives them a reliable way to actually use the web without breaking context windows.
exa.ai
github.com/vercel-labs/...
#aitools #devtools
Codex (the model) is really good at code. It usually one shots things I ask for.
But sometimes it's too autonomous, I'll ask it to investigate something and come back to the issue being fixed.
At least let me pretend I'm still in charge dude.
In the past two days I've already used it to embed itself into one of my products, extend itself with extensions that strip .env variables, use `dots` (my own take on task management) as the backing for the TODO tool, and so much more.
github.com/badlogic/pi-...
My gf jokes that I have new favorite AI tool every week. This week it's Pi. shittycodingagent.ai
I was already in love with the control OpenCode gives you over the interface / conversation via the SSE server and Pi takes it to a whole new level.
#aitools #devtools
I don't really use it anymore. Models have gotten so good that most of the uses for me were replaced with an MCP -> CLI wrapper, writting small CLI tools, and letting the model write bash to orchestrate.
I am however taking the lessons and currently building lootbox v2 haha
The Tamagotchi is really just a loading spinner.
OpenCode SDK is basically an agent framework, think VoltAgent or similar.
You plug into OpenCode’s existing agent loop and build your own UI on top.
On my 'desktop', I used the same loop to build an AI-powered graph editor.
Updated Tamagotchi screenshot showing the refined edge-to-edge interface and animated pixel character reacting to activity state.
Quick update: I tightened the Tamagotchi UI and wired mood + animation to live OpenCode activity like reading, patching, replying, and retry. New screenshot below. Details coming soon. #OpenCode #BuildInPublic
Screenshot of a compact Tamagotchi app in a desktop AI workspace: pixel pet in a 12x12 SVG grid, mood badge set to SAD, three vitals (F/E/J), and Feed/Play/Nap controls.
I recently switched to OpenCode as a long-time Claude Code user, and the biggest unlock is this: you can embed AI into any application layer, from silly experiments to real production UI. Built this reactive Tamagotchi as a tiny proof. Full build thread soon.
I help founders move from prototype to production without losing momentum. Engagements focus on the critical path: architecture, implementation quality, and workflows that hold up after launch. If delivery risk is blocking launch, let’s talk.
Built an issue tracker cli for Claude and added a force graph view in the terminal.
The forcegraph is mostly not useful except to kind of zone out while Claude works but it looks cool lol.
Now I know enough about AI to grok what MCP is, it irritates me because fundamentally:
Make a cli tool instead! Then a human can use it too.
It’s not stateless? That’s what unix sockets are for!
Once again we’re reinventing stuff from the 70s and making it worse.
I usually make the list and then not look at it until I'm done, it's the only way I won't feel pressured and it lets me keep the just in time discovery ADHD is good at.