ByteCaleb @bytecaleb

You're probably prompting AI wrong. Yet one of the BEST ways to prompt has been buried in so much hype that it's hard to see clearly. That's okay. I'll break it down. And I'll start with a bad example of how to prompt. In December 2022, I asked ChatGPT to write me an 800 word essay on "the potential for AI to think for itself" (that was my English topic for the semester - ironic). ChatGPT was new. I just wanted to see if it could. I waited a bit and like magic, an essay appeared. But it wasn't 800 words even though ChatGPT said it was. I put it in a word counter to check. It wasn't even 400. What a scam, I thought. But here's what makes such a kind of prompt a bad prompt: I gave the criteria about the essay's length, but I had an external method of verification that the AI did not have access to (Word Counter). If I'm writing an essay by myself, I'm not going to be able to know how many words are in it just by eyeballing it. I may not lie about knowing the number like ChatGPT did but I still need tools to count the words, and so does ChatGPT. But my expectation was that the AI would count the words while I would use a word counter. This is a misunderstanding of how the technology works. AI works better when both you and the AI have access to the same verification tools. A real world example of this is that recently, I wanted to improve my website's loading speed. I could get a page speed report from Google's Pagespeed Insights and start planning changes with Claude and then rerun the report after the changes have been made - but this creates the same problem as 2022: I have access to an external verification system that the AI does not have access to. What I did instead was I gave Claude free rein on a separate branch of my website that renders previews of the site, Playwright (to navigate around the site preview and take screenshots) and Chrome DevTools MCP, which allows Claude to run the Pagespeed reports on its own. With this, Claude could iterate on the site for hours, verifying when it is going in the right way or the wrong way from visual cues and the actual numbers from the report. Build -> Test (run report) -> Analyse results -> Loop This is loop engineering, and in the cases where the objectives are numerically verifiable it is a good framework to have at your disposal when building with AI. I ended up with 100% on almost every metric at the end of the day, and I didn't have to do anything. Thinking, "What tools will I use to verify that the work is done correctly, and can I share this with the agent so that I don't have to endlessly relay what the verification tool will tell me" rather than jumping straight into asking AI to make changes - will save you time. Loops like this are not useful in every case. It's difficult to iteratively verify that a design looks as you imagine it, or that a user flow operates as you intend it to, but be sure to watch out for where they might be a good place to apply!

How Raphinha is moving rn

I THOUGHT IT WAS JUST ME

Myk is Walking Backwards @mykola

3 months ago

me, begging, crying, on my knees: "Please just use plain english, I don't understand what you're saying." Claude: "The right fix, and the book's lesson applies: the tutorial rotted as a front door because a front door full of claims always rots. So the new root holds only what's

216 317 6K 549K 871

@claudeai [insert something funny] because I'm early and want a lot of likes

I'm building an AI video editor that lives in Telegram/Discord/iMessage/WhatsApp. Comment "BLOOM" and l'll give you free access

Last week a Chinese AI model beat Claude Fable: "the AI model that was too dangerous to be released to the world" - but as users rush to use it, there is one graph everybody is sharing and another graph nobody seems to have seen. Let me catch you up in case you have a clean feed and a healthy social life: Kimi-K3 is a new AI model released by a Chinese research lab called "Moonshot": and Kimi is the AI model that has some people asking if Anthropic's model - Claude Fable - has already been surpassed as the best AI model in the world. And this comes only DAYS after a massive splash from Anthropic's biggest rival releasing an AI model comparable to theirs. Now that you're caught up, I'll get right to it. Kimi-K3 is expensive. People typically associate Chinese research labs with cheap AI models but Kimi-K3 is the opposite. Artificial Analysis found it to be the third most expensive AI model in terms of cost per task (onIy beat by GPT-5.6 and Claude Fable itself). Don't get me wrong, it's still a massive saving from Fable, but this is still not the model you go to if you want to save money. Even if you go by API prices, Kimi-K3 is more expensive than well-known American frontier models like Gemini 3.5 Flash and Claude Sonnet 5. While Kimi-K3 is clearly more capable than both the aforementioned models, I can guarantee that most tasks you need AI for do not require the capability jump. If you really do need it, now you know one trade off. If you don't need it and you see it later on, don't fall for the FOMO. Thoughtless adoption is the fastest way to lose money for no reason.

keyboard_arrow_left Previous keyboard_arrow_right Next

If you're coding with AI (which is basically everybody these days), you need to know about Behavioural Matrices. Behavioural Matrices answer the question: "How do I continue to understand the different behaviours of my system as I continuously modify it?" & here is why this question is important. When you're building a fun sideproject, you don't tend to consider the long term evolution of your project. However, many systems that solve real-world problems solve a thousand sub-problems at once: - What happens if a user's device is too slow? - What happens if there's a network error while something is saving? - What happens if there's no data to display in this area? Over time, the behaviour of your system grows deeply interconnected to the point that a change in one area might affect how a totally different area operates, causing inconsistencies - or worse, break it entirely. Historically, software engineers created flowcharts and state machine diagrams that were maintained to express how the system behaved in different states, but it's not so easy to represent flowcharts or state machine diagrams inside a terminal or coding IDE. And that is where Behavioural Matrices come in. Behavioural Matrices attempt to be a tabular representation of flowcharts and state machine diagrams. It is super easy for an LLM to create a table in markdown and it is well-supported in almost all markdown renderers. A Behavioural Matrix typically has 2 columns: Old Behaviour / New Behaviour And a row for how your system behaves in each case. When you are making a change to your system, your coding agent can give you an overview of not just the change it is about to make, but the predicted effect that the upcoming change will have on a subset of surrounding features. This can warn you about potential bugs/regressions before you have to spend tokens fixing them, or even misalignment in the agent's interpretation of your instruction. I've found myself using behavioural matrices as the "flowchart of the markdown era" more and more so if you're coding with AI, you should definitely give it a try with this prompt the next time you're about to make a change: "Before making this change, create a Behavioural Matrix with two columns (old behaviour / new behaviour) and rows by feature detailing the expected changes in this feature and any knock-on effects (or non-effects) to dependent features" You might be surprised what you see.

@rom1trs EDIT

@borjafat Done

@rom1trs VOX

@TouchlineX Who's here after Brazil won?

@TouchlineX At this rate Bruna Biancardi will deliver the 6th before Neymar

@ThatG4m Parma

A website where AI models predict the lineups and scores of football games, and get scored on how close they were to reality

Google AI Studio @GoogleAIStudio

5 months ago

What are you vibe coding this weekend?

470 60 1K 125K 132

@UTDMael Brazil

@Andyc1178 I wonder... does everyone understand?

@FarzaTV This is what we thought Siri would be

This is what we thought Siri would be

Farza 🇵🇰🇺🇸 @FarzaTV

5 months ago

Watch me control my computer with just my voice. This is the future of operating systems. No hands. GPT-Realtime 2.0 is very, very underrated. Demo:

926 828 14K 3.8M 8K

I fell in love with software at 17. My first office job, I watched my coworkers grinding through the same Excel spreadsheets every single day. Manual, repetitive, soul-crushing. So I wrote a couple formulas into a copy of their sheets. You should have seen the look on their faces when they saw it for the first time. It was like I had discovered how to give them back their lost days. Things changed. People started going home earlier, dealing with the reports became quicker and we spent less time fixing inaccurate inputs. Automation gets a bad rep now because a lot of people are afraid of AI automating their job, but there's genuinely a side of automation that people look forward to. That's what I saw at 17. And it’s what I hear from people who use Chase Agents. One of our Chase Agents customers works in a research department and they set up an automation to fill out a quote form on their behalf based on prices sourced from a costing matrix - a task that used to mean jumping between emails, folders, and tools just to give a client a quote. I had the pleasure of sitting in on a conversation where she discussed Chase Agents to her superior, saying: “And here's the potent part [of Chase Agents]. It's if I’m not around, anyone else can do it based off that automation. And if anyone joins the department, they could easily figure it out.” By her own account, this wasn't automation that reduced her professional value in any form, it is automation that freed up capacity and enabled collaboration and organisational resilience. The AI industry's scare tactics will tell you that you should fear for your job (while charging premiums to your employer to replace you), but there's a good side of automation that I feel is being forgotten about. The kind of automation that genuinely allows you to scale your own capacity. That’s what made me fall in love with software.