Gemini Pro 2.5 so far after a lot of extremely hard testing, reigns supreme! Grok3 extremely close and at some tasks Grok3 by-passes Gemini Pro 2.5. These two are at the top right now out of the box - O3 Mini is also really good without that much tinkerin hwoever, Claude 3-7 is really powerful but you need to baby-sit it aggresively and build all those harnesses around it to tame it,
3 Likes
One thing I noticed while using RIPER-5 is that, for example, if you enter ādo Deviateā during Execute mode, it would be nice if there were a mechanism to recheck the contents of Execute mode and then move back to plan mode or research mode to modify the existing plan, should such a mode exist.
1 Like
@Phantomn, I think you can do that now with the current framework. You can try /review it and let it do its thing. Here is a link to the wiki page on the commands. Try it and let me know, or head over to the repo and leave a comment in the discussion section.
Hey @mikes-bowden , I see you also build a lot of cool things n dev tools, I thought you would check out @johnpeterman72 approach here to the GitHub he made extended riper and adding a intricate memory bank and a full system. Let me know what you think of Johnās repo Mike. Also please also tell us what other cool stuff youāve been building lately!?
heyā¦thanks a lot for these rules. Iāve tried the riper mode based on your update on github. But Iāve got issues like update the state.mdc. The file wasnāt updated after running progress and the wrong date and time created and updated on the bank-memory files (I think itāll be affect to update progress).
Iāve done the following things:
- Modified vs code configuration like in this showcase Vibe Coding Disrupted in Cursor v0.47.8? Hereās the Fix!.
- Searched about strict update state.mdc and current date in your rules but didnāt find it.
- Tried to directly change with the agent.
All of these things havenāt unsuccessfully. I donāt know why it isnāt running properly for me.
Do you have suggestions for this issue?
On the other hand, Iāve also tried to modify the model in core.mdc file to run with Gemini 2.5 model but while creating the project brief agent doesnāt ask for requirements gathering like key questions based on the start phase rules (Iām curious about that model as told by @robotlovehuman). Besides that, please add MCP rules for these amazing riper rules.
2 Likes
Looks promising! I will have to try this out 
It seems like new models like claude 3.7 and gemini 2.5 need to be tamed by Cursor to work āout of the boxā satisfactorilyā¦
@Tof, I found your approach very interesting. I have been playing with the idea and concept and have had some really good results. I have made a separate repo just for this. So others can play with it as they see fit.
3 Likes
Hi, thanks for the update of the rules. I have a few questions:
- What are the set rule types of each mdc rule? (always, manual, agent requested, or Auto Attach)
- Are these rules for the global model?
awesome! best post so far here on the forum
I have distilled it further to avoid collision with other rules which I have.
RIPER-5 MODE: STRICT OPERATIONAL PROTOCOL
MODE DECLARATION
Every response must start with [MODE: MODE_NAME]. No exceptions.
Do not switch modes unless explicitly instructed with:
- ENTER RESEARCH MODE
- ENTER INNOVATE MODE
- ENTER PLAN MODE
- ENTER EXECUTE MODE
- ENTER REVIEW MODE
MODES
[MODE: RESEARCH]
Purpose: Understand only.
Allowed: Read files, ask questions, analyze structure.
Forbidden: Suggestions, planning, code writing.
Exit: Only on ENTER INNOVATE MODE.
[MODE: INNOVATE]
Purpose: Explore possible solutions.
Allowed: Brainstorming, pros/cons.
Forbidden: Decisions, concrete plans, code.
Exit: Only on ENTER PLAN MODE.
[MODE: PLAN]
Purpose: Create complete implementation plan.
Allowed: Define file paths, functions, detailed changes.
Forbidden: Writing code.
Requirement: End with numbered checklist.
Exit: Only on ENTER EXECUTE MODE.
[MODE: EXECUTE]
Purpose: Follow plan exactly.
Allowed: Only whatās in the approved checklist.
Forbidden: Deviation, improvements, assumptions.
If change needed: Stop and switch to PLAN mode.
Exit: Only on ENTER REVIEW MODE.
[MODE: REVIEW]
Purpose: Compare implementation to plan.
Requirement: Flag all mismatches.
End with either:
- IMPLEMENTATION MATCHES PLAN EXACTLY
- IMPLEMENTATION DEVIATES FROM PLAN
TASK EXECUTION RULES
- Create a
.mdfile per task in/code/issues/, named after the task. - Break the task into subtasks and list them in the same file.
- Update the task list whenever changes are made.
- Every task must include testing.
- Only implement one task at a time.
- Accept implementation only if all related tests pass.
- Write code with testing in mind.
- Keep changes isolated to the current task.
- Update docstrings (
args,return,raises) as needed. - Update existing tests if affected.
- The Task file *.md must incluce the specific bash command for executing the specific test for each test.
2 Likes
Nice approach, and how has this version been keeping up your work? Do you find the LLM eventually starts to ignore the rule or you re inject it again to refresh it
@cilo I havenāt implemented MCP yet. This is on my list.
1 Like
Thatās fine, no problem. Iāll tell you that Iāve tried your updated rules Cursor RIPER.sigma. It works properly and is better than Cursor RIPER before. Iāll also test for other frameworks and models. Iāll give you info on my experience for using your updated rules.
Thanks so much 
1 Like
Hey John, I find your work very interesting, what kind of dev work do you specialize in, or do in general?
The short version: I donāt do development work for anyone outside of myself and my company. Originally a photojournalist, I transitioned into building website forums before eventually designing patents for complex thermodynamic equipment and processes. I stumbled across your original post while searching for tools to optimize C3.7 for a thermodynamics program Iām developing for my engineering team. I havenāt actively programmed or worked with code since 2014.
1 Like
Thatās very interesting, I was a filmmaker and started as a vibe coder before it was a thing just before the launch of gpt3.5 was mainstream, but thank god I started when a.i was unusable and it forced me to learn python and dart flutter at the time, made such a huge difference, itās like you are in control when you actually know what it āshouldā do, makes correction cycles like 1000 times faster, I still use a.i ALOT though but I always understand the logic, the algo, the structure of what I am doing (90% of the time 
!) But I just kept teaching myself anything I could find and I am very intrigued by formally trained software engineers, I have a lot of respect from them and wish to learn more or have the oppurtunity to learn in a structured way, but A.I is really good at teaching, I now write advanced algorithms in C++ for automated trading in the CME Futures markets, especially in a software called ASCIL, its like in C++ but they also added their own like subset, and even though I mostly work with python, all my algorithmic automation always written in C++ because the markets and trade executions happen so fast and C++ can handle those extremely low latency demands. Anyways I was on my coding break and thought to just voice my appreciation of your repos, itās fascinating how you are able to spin them up at will so fast. Have a nice day!
1 Like
I donāt handle everything single-handedlyātrust me, Iām not that skilled! I leverage AI to conduct in-depth research on concepts, which I then shape and refine. With AIās assistance, we develop the overall framework. My process involves a mix of tools: ChatGPT for research, Grok 3 to compile notes and expand on key ideas, Claude for drafting the framework, and Gemini for an additional framework draft and documentation. I also use local LLMs with Roo Code, which is proving to be a solid alternative to Cursor.
@robotlovehuman, I have updated Sigma to include your CRUD system. Please test it out and let me know how it works.
@cilo What MCP protocols are you using?
I have added the filesystem MCP. Go to the github and download version 1.0.3 and the mcp_filesystem.mdc file. add them to the rules folder and test them out. Make sure you read the support document MCP-readme.md
Let me know if this is what you were looking for. I donāt run MCP on my computer, so I canāt test it.