2. prompt for self-verifiable solutions the agent can run. if agent can see and reason about what it does, it can detect divergence from plan or what you prompted for in the first place. browser use is the clearest example of this. however, it can also create tests/scripts to

3. use github cli a lot. with this, agent can see old PR implementation, check action status and logs, get PR comments to address etc. just to verify what's been done previously, but also automating some of the things i would've done by going to github website

4. using web search and curl from agent. a lot more docs are exposing it as .md files, which makes it easy for agent to curl. if not, it can use the web search tool go get the relevant pieces

5. best of n in background agents. let it implement 4 variants with 4 different models, then select the one that gets it the best. chances are the implementation quality varies a lot

6. setup commands you use on a daily basis. for me, that's /create-pr /review-changes /deslop (remove generated slop from models)

7. visual context! drop screenshots of ui you're working on, or just a quick excalidraw sketch. models these days are really good to convert the ui to code, and combined with 2. above it becomes nearly pixel perfect

to see how the broader team use cursor, check this out

@ericzakariasson Such a goated strat, really lets you align with the model before it creates 16 files of random shit lol

@ericzakariasson increase limits for adding custom modes. 6 is not enough bro

@ericzakariasson Hope this “How I’ve been using Cursor” becomes a series and a habit. It’s incredibly valuable for us users to see how you use it internally (even if everyone’s workflow ends up a little different). 💡 Thanks for sharing!

@ericzakariasson I’ve discovered that I can get even more from the plan mode when I keep all my tasks/CRs in a single backlog .md file, and for each task I have 1-9 bullet points with requirements, context, suggestions or references. Then my typical input to the plan mode is “do #42 from @

@ericzakariasson @threadreaderapp unroll

@ericzakariasson I’d be really curious to see under the hood and see exactly how the plan mode works. Like you, I always had it “plan” everything using .md files, and then execute. I find that the plan mode can be quite rigid, especially when it comes to brainstorming, so I don’t always use it.

@ericzakariasson Same. Plan is epic actually. Used to ask it to write .md docs but somehow the step where it asks questions super charges the results

@ericzakariasson tbh a line like "let's discuss the approach before making any changes" has worked fine for me in agent mode

@ericzakariasson Agreed on web search. I even was successful recently with the prompt: “verify all the external links in this docs page work and fix any that are wrong”

@ericzakariasson To be honest, it's impossible to use this plan mode yet. I came back in half a year and realized that it's also not usable, there is also a constant pause, there is no agency in the local mode, and to run an agent code that works somehow, I can't really see progress; it doesn't

@ericzakariasson yeah, plan mode makes everything smoother it helps get clarity before jumping in, which saves a lot of back and forth later

@ericzakariasson Same. I’ve also noticed similar workflows with others over the past month, including doing what plan mode does manually before it was released. I wonder what changed recently for folks to converge on this approach. It’s insane how effective agents can be doing it this way 🤯

@ericzakariasson how i've been using cursor in the last month 1. Open claude code in the integrated terminal and bring it into the editor

@ericzakariasson It so good. I fix bugs faster with it, code quality absolutely mooning!

@ericzakariasson GPT-5-codex plan and the agent is pretty much like using codex cli, its works so well

@ericzakariasson I get better result by planning with agent mode and creating a markdown with it.

@ericzakariasson These monthly Ai workflow updates are awesome!

@ericzakariasson Plan mode is 100% broken for the last 2 weeks 😭

@ericzakariasson works really well to make a “specs” directory, and just fill it with lots of markdown specs. and an index markdown to describe each one. not used plan mode yet 😆

@ericzakariasson The question is; WITH or WITHOUT appending as much relevant context as possible

@ericzakariasson I also use Plan mode a lot

@ericzakariasson When I want to update the plan, the tool call always fails. Now unfortunately I have to go back to making my own .md files and not use Plan Mode. Some feedback for you guys.

@ericzakariasson could you elaborate on /deslop?

@ericzakariasson am I supposed to read the code or just send to prod? AWS is done idk why

@ericzakariasson Plan mode + Claude is a really refreshing upgrade. BUT reaching the subscription limit is waaay easier than ever before 😱 Never hit it before in months of usage, now hit it within 2 weeks (not even)

@ericzakariasson Does plan mode search/read files like the agent?

@ericzakariasson All of that without reaching limits in 2 days??! Awesome

@ericzakariasson what model for planning, or running "Auto" ?

@ericzakariasson hi eric! love your posts about how you maximize cursor! lots of alpha to be gained towards optimizing for usage