User Rules, with memory, errors tracking, rules generation

Showcase Built for Cursor
41

That’s exactly the point!

It makes perfect sense to talk to the AI agent about such things, sometimes amazing things come out of it. I can advise everyone to ask the AI if things don’t work.
For example, you need comments or DocStrings in rules with code examples or invalid examples of how not to do things, all superfluous, suitable for humans but not for AI. Instead of forcing the human language on the AI for such things, we should accommodate the AI and speak its language. This saves a lot of space in the rules. A clean codebase and workflows from development make life easier.

42

@Tof Thanks for this, been experimenting with what I call ‘Sigma+’ similar to yours, but slightly different, closer to what @johnpeterman72 has done with the Riper5 system.

Typing out loud here - how many other applications have you experimented with? Coding is but one discipline. I’ve been goofing around with this, and I love it - game changer for sure - especially with longer prompts and GPT actions.

I was working on another implementation - and for the life of me, I couldn’t get the precise action I was prompting for - solved it with Sigma+ in 1 shot.

Thank You :folded_hands: !!

2 Likes
43

Someone can confirm that’s working?

I copied the pasted but don’t create the memory and task folders

44

adiiokid - I’ ve played with the full Riper5-Sigma System, but have since modified a custom version that suits me better. My opinion - Riper5 shines and is a major value add due to these features:


  1. Consistency – The Modes & Sigma+ language (calling it Sigma+ for now, lacking a better term) for AI is akin to mathematics for humans: clearer, less ambiguous instructions. Just as “1 + 1 = 2” is clearer than saying “Complete the following…,” Sigma+ language significantly reduces ambiguity.

  2. Simple Process – Clearly defined workflow: Research → Innovate → Plan → Execute → Review Provides consistent, solid, unambiguous results, reduces scope creep, and gives better control over each exact step.

  3. Effectiveness – With Sigma Riper5, I’m achieving better results using Claude 3.7 (“dumb” mode) than I previously did using Claude 3.7 with :face_with_monocle: Thinking mode (“:eyes: smart?” :man_shrugging:).

  4. All Tools Enabled – I keep write mode enabled and ensure all tools are turned on. Claude executes write commands strictly in permitted modes.

  5. Customization - Prompt Claude to deconstruct Sigma+, modify, put it back together as your own version.


EXAMPLE

1. (/r) RESEARCH MODE

/r @filename.xyz (or foldername)

image
image772×484 29.6 KB



2. (/i) INNOVATE MODE (Can combine with (/p) PLAN MODE)


Next prompt could be something like:

/i Brainstorm ways to complete the next task, create a  grading metric based on what's important for project goals, grade each idea with the metric.


/p Devise a synergistic plan with chronological flow broken down into - Phase{A,B,C}/Steps{1,2,3} (we will be completing it 1 phase at a time).

I switched the mode from innovate(/i) to plan (/p) within a single prompt, you can play around with this.


image
image568×552 31.6 KB


image
image490×662 40.9 KB



3. (/e) EXECUTE MODE


/e ONLY PHASE A of above plan, then provide a /review

image
image810×1288 84.2 KB

45

Tof Amazing job! Really appreciated using it! A little complex to understand, trying to learn with the ones you provided!

secprobe Amazing rule organization/structure! May i ask, is the agent choosing correctly? And, do you updated the rules?

46

very interesting, how can I test this?

47

First I have to thank @Tof , as a global rule it didn’t work, but it did when I put it directly into the project.

Later, I have to thank @johnpeterman72, @robotlovehuman and @bmadcode , their methods are almost divine, and I see that they all complement each other, it’s incredible how this seems out of this world.

I did several tests and created some fundamental rules, and after few hours researching the topic
I created an AMAZING MODE called “RIPER-5 Mode” Fixes Claude 3.7 Drastically! , to research the RIPER-5, I saw that everything I had done was already ready in this project, in fact I had created a small fraction of the project, it’s really incredible that this is not widely publicized here, we need to do this.

1 Like
50

Unfortunately, I can’t make a real statement about this yet, as I’m still lacking tests for this and Cursor has been releasing quite a few updates for the rules recently, such as rule types and so on.

But if you would like to test this, just play around with the rule types or trigger the rules manually in the prompt @Cursor Rules and let the agent develop a small project, then look at the code to see if the agent has followed the rules or if there are any rule violations.
Make sure that the agent has a clean chat for the context size of the AI, so it is important to avoid that the agent either does not follow the rules at all or only partially reads them.

You can also use the cursor’s memory bank and build something like a rule check or rule logging into the templates.

Also very interesting, I might use that for the rule check, the “Custom Modes”
This allows you to create your own modes for the input for cursor, rules, templates and so on.

image
image843×527 20.4 KB

image
image474×335 9.85 KB

image
image446×583 16.9 KB

I will continue to update the rules, provided that it works as I have imagined.
The most important thing for me was not to make the rule files unnecessarily bloated in terms of size or lines and to make them AI and non-human compatible.

51

Why would you say performance benchmarking (aka testing) doesn’t bring real value? It’s the foundation of the scientific process. I’m asking because while I really appreciate the new approach, I’m inherently skeptical in science and engineering when someone just says 'Trust me bro!" ;).

At first glance, most of the training data for Claude/GPT4x/Gemini is in normal written sentences before tokenization, so this seems like saving tokens but at the expense of going further from the native language. In fact @bmadcode also had the same concern.

For kicks, I asked Claude 3.7 about your symbolic rules (RIPER.Tof.1.0.0.md) vs RIPER-5’s (RIPER.sigma.memory-bank.1.0.3.md) rules by @robotlovehuman / johnpeterman72 based on this thread. I got

While RIPER.Tof.1.0.0.md offers elegant abstract principles, its mathematical notation and lack of concrete examples make it less suitable for direct LLM implementation. LLMs typically benefit from explicit examples, clear boundaries, and specific procedures - all of which are abundant in the more comprehensive sigma.memory-bank document.

The ideal approach might be to use RIPER.Tof.1.0.0.md as the conceptual foundation, but implement its principles through the detailed mechanisms specified in RIPER.sigma.memory-bank.1.0.3.md for optimal LLM rule compliance.

Finally, all the leaked prompts to date, whether OpenAI, Anthropic, Cursor etc … they’re all in plain simple English. Simple, brief, to the point, but all in English.

Now, I understand the insanity of using Claude to reason like that. That’s why I was hoping to find some concrete data (e.g. 5 runs each) or some additional historical data/paper on symbolic prompt.

Even for the mermaid chart based ones, without data, it’s hard for the community to know concretely. There are several other factors like just-in-time prompt/context injection and memory which are great BTW.

Again, thanks for the work. Just trying to ground this in the scientific process.

Cheers!
Sid

52

Automated benchmarks can’t measure everything. I prefer to test through daily use, in real-world situations.

It’s an empirical approach, conducted outside academic frameworks, but with rigor, iteration, and confrontation with the real, like any form of applied research that lives and evolves.

Just because OpenAI or Anthropic do things a certain way doesn’t mean it’s the best way. Their own design choices deeply influence them, sometimes without their full awareness.

In fact, they acknowledge it themselves: what happens in the latent space of a LLM largely escapes their understanding. That’s precisely where my work takes place.

I’m not trying to prove anything to anyone. My concern isn’t validation, it’s exploration.

Just ask yourself this: where do the emergent capabilities of a LLM come from?

Here, the annotation-based syntax is intentionally extreme, not for style, but to isolate the prompt from its context. What matters is not the form, but what it represents and what it makes possible.

1 Like
53

Sid, this really is a game changer—and I’m 100% with @Tof on this. I’d even take it a step further: benchmarks for LLMs are incredibly linear. Prompting, on the other hand, is both the secret weapon and the Achilles’ heel. Small changes, and suddenly the AI goes from Einstein to eating glue.

To put the scale in perspective: GPT-4 was trained on the rough equivalent of a million books. That’s like reading 50 books a year… for 20,000 years—even if you had Audible on 3x speed, good luck catching up.

Pretty much all my primary prompts are heavily laced with Sigma+ which is inspired by @Tof 3ac and @johnpeterman72 CursorRiper Sigma.

1 Like
54

I have updated Sigma to include MCP and other functions. I have updated Sigma CursorRIPER.sigma/.cursor/rules/RIPERsigma1.0.5.mdc at main · johnpeterman72/CursorRIPER.sigma · GitHub to include MCP and some other functions.

1 Like
55

BTW, for anyone wanting to export their custom english rules to symbolic, I had created a custom GPT several months back at ChatGPT - Prompt Engineering: Markdown to Symbols . I still won’t claim performance one way or the other but it’s a tool for prompt transformation.