made a lil guide cursor canvas for @poteto's pstack 👀
one page of prompts. each points to the right skill for the job - low, medium or high budget
i used to reach for the same 5 skills. now i just pick the right one for the job
I stopped reaching for Zapier. The most powerful integration layer was already on my machine.
Managed platforms see what APIs expose. Your terminal sees the filesystem, socket events, process state. The connective tissue was already there. I just started using it. https://t.co/skp5V0FwOg
I built two memory systems for my agent. Researched Hindsight and Supermemory. What actually made it smarter: a promotion routine that updates a single markdown file. Runs between sessions, reads traces, writes rules. Retrieval is reactive. Promotion is proactive. https://t.co/mndu5T9mhR
I spent years learning how to build software that is reusable, maintainable, and scalable. That still matters. But with AI, I keep building software that is the opposite: narrow, temporary, and built for one job.
Ephemeral scripts. Ephemeral features. Even ephemeral software. https://t.co/W0psyW8Zeu
Prediction - companies will need an in house meditation teacher and offsites will probably look like meditation retreats - to deal with cognitive exhaustion
This task took 30 minutes. I was there for 7 of them. 3 decisions, the rest delegated. A principle I follow: once I know what needs to happen, AI should do the rest. The thinking is mine. The execution is delegated. https://t.co/p7fs2xu7iW
LLMs are trained on the world's generally available data. You get the wisdom of crowds - average reasoning on everything. Great when you're learning. But for anything you're above average in, the model is worse than you. Capture that gap. That's your alpha. https://t.co/S33c8z81jt
I never expected to be designing terminal UIs. Then Claude Code moved me into the terminal full-time. So that's where I built my notification system. The principle: build where you already are. Your best personal software lives in the context you're already in.
AI isn't replacing your role. It's replacing your tasks. Forget your title. List what you actually do, find the recurring themes, and automate them. The work that's left after that is the work worth protecting.
Building your own UI means skipping onboarding, billing, notifications, and feature discovery. No product fluff built for other people. Just the core thing, built for how you actually work.
For remote sessions, I used to spin up a fresh Claude remote env every session, which meant tons of time wasted on setting it up properly. Now I’ve got a persistent remote workspace that just connects and works. Didn’t feel like the best option before, but with Openclaw it does.
In the initial days of Google search going to page 2, 3 was normal, but not anymore, people only look top 3-5 links. I wonder same will happen with agent chats, people will probably never have a conversation with more than 2-3 turns.
Why are so many robot company demos showcasing how their robot can withstand kicks/pushes?
It might be technically complex to handle but that’s not what users want.
Almost all code next year will be written by AI, not because models will be better by then, but because developers will have become good enough at prompting them by then.
Every company should build their own memory layer that has all the information about processes, structure and product. An internal team should maintain this. Notion/confluence documentation doesn’t cut it, it was never meant for AI agents consumption.
Almost 100% of my code is written by either @AnthropicAI claude code or @cursor_ai agents - all running locally.
The review/edit cycle with cloud agents is too long, hopefully can start relying on them soon.
Have been planning to work on this. I would love a canvas like experience with pre built agentic flows as a DAG that I choose from. https://t.co/7q4zGRVteW
Claude 3.7 was released the right way, already available in bunch of apps the you use today, perfectly coordinated.
This is opposite if what other labs do, talk about it for months before releasing and then make it available weeks after to everyone.
Weirdly @grok 3’s insane response speed is making me “feel the AGI” more than the reasoning.
Since AGI might just be a loop, the speed would make a big difference.
New kind of prompt injection for thinking models - distract them!
Even for AI focus can be hard lol, one thing it shouldn’t learn from humans. https://t.co/R0mxlBTKrf
I spend more time editing and formatting the copy-pasted response from LLM into another app than editing the contents of the response itself.
Models are getting so good at One-shotting what I want.
Four thousand weeks by @oliverburkeman is my favourite self help book. Here’s my review - the best review that any self help book can get -
I didn’t need to read any more self help book after that. Thats was the last one.
Nowadays, it’s a bigger flex for a software engineer to show something they have built and say - didn’t write a single line of code, it was all AI.
Got this a few times this week!
The Times They Are A-Changin' 🎶
One other thing Deepseek got right is the branding / naming.
Deepseek app using the Deepseek-V3/R1 model built by company called Deepseek.
Vs
ChatGPT app using o1/gpt-4o built by company called OpenAI https://t.co/zrVeWzulqW
Every time I see a system prompt, it always look much simpler than what it can do. There is a lot of magic in orchestration and app/ux. Also the hack/approach to get the system prompt is much more impressive than the system prompt found. https://t.co/6vlKby7ufo
UX research will have a new meaning. Instead of researching whether a human can navigate the site/app, use ChatGPT operator to gather whether an AI can navigate effectively or not. Then make adjustments to make it easy for the AI. https://t.co/tSavQFcXkd
I love studying about consciousness, thats one thing that will keep giving me meaning/purpose in the AGI/ASI world, since it’s not clear to me that even ASI can solve the “hard problem of consciousness” - @davidchalmers42 .
Positive examples > negative ones when prompting AI. Why? Positive examples clearly show what to do, while negative ones only highlight what NOT to do. Sure, the model will not do that, but then what it should do is still open-ended and up to the model. https://t.co/iR0sog0N9Z
This is on point. You would feel not moving fast enough if you only look at the quantity of AI products, but based on quality you are not behind. No point trading off high quality for speed, the market/customers catches up quick. Survival of the highest quality product. https://t.co/nbCF2TvDTr
Everyone is equating AI driven age of abundance with utopia, but what about the big chasm that needs to be crossed - the Paradox of Choice.
Not sure thats an easy one with the social media native generation.
I often hear latency as a primary issue with agents, for me I'd rather wait than work with a half baked, average answer. It also seems that the expectation for lower latency by software is much higher than that of humans. This doesn't seem fair.
Accuracy > Latency
Gotta say GitHub copilot workspace has the best UX for viewing and editing the plan. That in itself makes it very useful because it’s all about promoting and being able to “prompt“ the proposed plan well is absolutely critical to see the magic of these agents. https://t.co/zbsak3J9wA
Been interviewing engineers - small note to everyone - gotta limit "we" in your responses, talk more about what you did specifically "I did...".
Common theme in last 3-4 interviews. No way for me to know what your actual contribution was, if "we" did everything.
Simple hack to 10x your AI use - swap the primary and secondary bars in VSCode/Cursor/Windsurf.
People naturally read left to right! https://t.co/I7a6dzcJgS
Working with AI needs a mindset and habit shift. For a lot of people, going to AI is not the first thought, making that happen is a big behaviour change.
Find it strange that models are close to 100% accurate in reading text from images, but the reverse is - close to 0% accurate on generating images with text.
The future of interfacing with AI is vision and voice. Today we use vision and touch to interact with our computers, but touch is too slow.
To get better at prompting move all instructions to voice input and screenshots. Reserve typing for rare scenarios.
Better to be a power user of an ok AI tool than a basic user of a good AI tool
Note to self: don’t keep looking for new tools rather keep going deeper in a few
Just had a creepy chatgpt voice moment, I asked a general question about baby development, it ended by asking how my child (BY NAME) is doing with his milestones. I know it collects memories, but experiencing that felt super weird. Flight/Fight triggered.
This makes a lot of sense, typically any chat gpt response requires a follow up clarification or something, takes 2-3 steps to get what you are looking for, even if the first prompt is detailed https://t.co/jVzsf1PwF5
