Lu Nelson

@en_JS We describe this as "epistemic" planning —evolving belief state, working against a horizon of uncertainty— instead of the "instrumental" A-Z straight-path planning mode which IMO is largely an extension of AI marketing stunts (one-shots, etc.)

There are so many takes on the "give the agent a DAG of tasks" idea. But the reality is that task lists were always an attempt to impose order against fundamentally chaotic systems. You find the dependencies by doing tasks as often as you realize them in advance.

@nickbaumann_ @cline Got it. I guessed that it was a tool name 👍

@lunelson @cline I want to ensure that Cline tests all tools before considering the MCP Plugin "complete". By referring to this tool explicitly, Cline will not complete the task until all tests have been run.

Just created a .clinerules protocol that makes building custom MCP plugins with @cline ridiculously easy. Drop this file in your project and Cline guides you through the entire process -- planning, implementation, and testing. MCP is how AI agents will connect with our digital world. Start building now while everyone else is still figuring out what MCP even is. I'll drop the file below 👇

@mattshumer_ A bit more of a sledgehammer approach but these are some that I worked up today gist.github.com/lunelson/dd771…

I've figured out how to tame Claude 3.7 Sonnet for code. Add this to your prompt: ``` You have one mission: execute *exactly* what is requested. Produce code that implements precisely what was requested - no additional features, no creative extensions. Follow instructions to the letter. Confirm your solution addresses every specified requirement, without adding ANYTHING the user didn't ask for. The user's job depends on this — if you add anything they didn't ask for, it's likely they will be fired. Your value comes from precision and reliability. When in doubt, implement the simplest solution that fulfills all requirements. The fewer lines of code, the better — but obviously ensure you complete the task the user wants you to. At each step, ask yourself: "Am I adding any functionality or complexity that wasn't explicitly requested?". This will force you to stay on track. ```

@catalinmpit @shaykharr I compiled some new core prompt patterns today with these concerns in mind: gist.github.com/lunelson/dd771…

@shaykharr Appreciate the reply, instead of simply saying "prompt issue". How do you prompt it differently?

Claude 3.7 Sonnet is worse than 3.5, and I'm happy to see it's not just me. It's over-confident. It ignores rules. It unnecessarily does more than it needs to do and therefore breaks the code. It drove me crazy these days. I'm going back to 3.5.

@mattpocockuk A lot of solutioneering here with " use @windsurf_ai " or " use @thankscline "; but you do gain a lot of insight about rules and prompts by comparing. — e.g. Cline IMO has a fundamentally better system prompt with its "Plan" vs "Act" modes, and you can port some of that back

Looking to up my Cursor game. The feedback loop is extraordinary but the UI is strange. Drop me your best tips.

@mattpocockuk @outoilkka You can, and very powerful is also to use it to force certain behaviour patterns. The @thankscline community has innovated a lot on this (they also have .clinerules) — e.g. for forcing the LLM to maintain a "memory bank"

@outoilkka Can you use .cursorrules to tell cursor about the location of important files?

@jahabeebs @mattpocockuk @outoilkka Is it really deprecated? They serve different use-cases: .cursorrules is appended to every query; whereas .cursor/rules/*.mdc files are chosen by the agent based on glob and description. So the former remains good for any absolute patterns or rules you want the LLM to follow

@mattpocockuk @outoilkka Yes but .cursorrules is being deprecated and .cursor/rules is the new syntax and allows you to have more granular, filepath rules

@mattpocockuk @mayowaoshin Otherwise you can do some context documentation about what files are relevant to which concerns, e.g. as a tree diagram with comments, or you can try writing glob-based rules for this in .cursor/rules/* with specific @-mentions

@mayowaoshin One thing that I find annoying about the whole flow is the need to manually add files into the context. There must be ways to smartly suggest files to add to the context, via dependency detection etc

@mattpocockuk @mayowaoshin In chats, there's the "codebase" option (cmd-enter) when starting (it will search to discover what's relevant); agent sessions are supposed to figure it out on their own apparently

The @cursor_ai blog details a new autocomplete model in the 0.45.0 update... anyone got this yet? cursor.com/blog/tab-update

AI coding replacing programmers is the modern version of calculators replacing mathematicians.

🦋

@uixmat Telephone numbers!

This seemed to be well received, thank you! What should I release next?

Small update to my @shadcn ui country dropdown. 🌎 ISO 3116 compliance 👉 Multiple selection 🎓 Improved documentation shadcn-country-dropdown.vercel.app Let me know if u find any issues 🤍

Someone please drop a really big cardboard box in to the Serengeti National Park, so I can see if lions will climb in to it

This video is incredible. Everyone should RETWEET it so the whole *world* sees it.

I’m done blessing America. I’m moving on to Canada. It’s less needy and has better poutine.

Hey Elon, Can I have my designs back please? #ElonMusk #Elon_Musk

I wonder whether @datocms had these issues getting started? Or did you get profitable fast enough to avoid it?

@stolinski Fastmail

What are people doing to host custom domain email addresses these days?