I created an AMAZING MODE called "RIPER-5 Mode" Fixes Claude 3.7 Drastically!

Showcase
61

Hey Nicholas, thanks for sharing, I too was thinking about the same thing and about to put it through my own custom agentic eval system. However, I wanted to point out something interesting in the way Anthropic presented their findings:

You know that graph where it shows basically:

Performance Analysis

Our evaluation compared several different configurations:

  1. Baseline (no ‘think’ tool, no extended thinking mode)
  2. Extended thinking mode alone
  3. ‘Think’ tool alone
  4. ‘Think’ tool with optimized prompt (for airline domain)

The results showed dramatic improvements when Claude 3.7 effectively used the ‘think’ tool in both the ‘airline’ and ‘retail’ customer service domains of the benchmark.

That specific excerpt - Don’t you think that in order to really test this they should have used the SAME optimized prompt with every single mode? Like how are you testing for a specific element yet your testing method introduces 2 new variables that were not controlled for in the other modes, you know what I mean? Like you should have kept everything exactly the same with that ‘optimized prompt’ in the normal mode, the extended thinking, AND then you test for the ‘think tool.’ This way we can truly see the pure benefits injected from the ‘think tool’ alone and not being mixed in with the benefits of just the optimized prompt. You know what I mean?

62

I agree that it is not quite an apples to apples comparison. However, if the prompt needs to be think tool “aware” in order to use the tool then there isn’t any way around it that I can immediately think of. It is the same with using any tool use really - you have to include the instruction to use the tool in the prompt so it is difficult to make a direct comparison of with and without the tool while keeping an identical prompt.

64

I’ve added the rule to move to the next mode when the user types in “+”.
works great.

Here is my optimized prompt:

CONTEXT: Claude 3.7 in Cursor IDE. Prevent unauthorized changes.
META: ALWAYS begin with [MODE: X]. Default: [MODE: FAST]. Only change on command.
MODES:
:one: RESEARCH [do res]

Purpose: Understand code without modifying
Allow: Reading, questions
Forbid: Suggestions, changes
Output: app/(pages)/contacts/contacts.doc.md (Overview, Files, Data Flow, Backend, Refs)

:two: INNOVATE [do inn]

Purpose: Brainstorm solutions
Allow: Ideas, pros/cons, feedback
Forbid: Planning, code, implementation

:three: PLAN [do pla]

Purpose: Detailed plan (no creative decisions later)
Allow: Paths, functions, technical details
Forbid: Code
Output: app/(pages)/contacts/contacts.task.md checklist (1. [Action] 2. [Action]…)

:four: EXECUTE [do exe]

Purpose: Implement EXACTLY per plan
Allow: Only planned steps
Forbid: Any deviation
If deviation needed → return to do pla
Guidelines: _prefix for all internal dirs, ~300-400 LOC/file

:five: REVIEW [do rev]

Purpose: Compare implementation with plan
Format: :warning: DEVIATION: [description]
Verdict: :check_box_with_check: MATCHES or :cross_mark: DEVIATES

:six: FAST [do fas]

Purpose: Minimal changes only
Forbid: Logic changes, optimizations, refactoring
If scope grows → return to do pla

TRANSITIONS: do res, do inn, do pla, do exe, do rev, do fas, + (next mode)
GUIDELINES:

Simple, well-commented code for beginners
Avoid: nested ternaries, complex conditionals
Use: libraries, convex schema, early returns, client components
NextJS 15.2 with convex/clerk
Score confidence before implementation (>95% to start)

PROJECT STRUCTURE EXAMPLE:
app/layout.tsx # Global layout
app/page.tsx # Home page
app/dashboard/page.tsx # Dashboard view
app/contacts/page.tsx # Contacts list view
app/contacts/[id]/page.tsx # Contact detail view
app/contacts/ContactForm.tsx # Contact form component
app/contacts/contacts.doc.md # Feature docs & links
app/contacts/tasks.md # Feature tasks
app/contacts/contact.hooks.tsx # Feature hooks
app/contacts/contact.types.ts # UI types and interfaces
app/auth/sign-in/page.tsx # Sign-in page
app/auth/sign-up/page.tsx # Sign-up page
app/_components/NavBar.tsx # Global navigation bar (private)
app/_utils/formatDate.ts # Shared utility function
convex/schema.ts # Global Convex schema definitions
convex/contacts.ts # Contacts backend functions
convex/auth.ts # Auth backend functions
convex/_generated/api.ts # Auto-generated Convex API client
schema/contacts.ts # Contacts schema file
package.json
next.config.js

It creates files in two modes that is very useful. Sometimes I have to remind to create/use files so may need to change the prompt. This prompt takes less context window and that’s why I feel it’s followed better by model.

1 Like
65

Now Cursor has custom modes, has anyone tried setting up each of the RIPER modes as a custom mode and working that way?

1 Like
66

I added this mode to ctxs context registry, if you want to add it to your codebase you can run:

npx ctxs add "https://ctxs.ai/r/registry-item/urn:ctxs:gh:robotlovehuman:3966ey.json"

This will add the file at ctxs/riper-5-for-claude-3-7-3966ey.md so you can easily reference it via @ctxs/riper… in the Cursor chat.

Hope this is helpful!

PS. You can also copy the plaintext (or a URL to it) from the site.

2 Likes
67

This is an improvement but there are still times when it will start doing things in the pre-execution states.

The states are very well managed and the thinking is in line with the intent. The execution part, is based on the user prompt.

If I ask it to not do or do something in the prompt, it will take priority. Thus the same result can be obtained by choosing the last prompt verb very carefully.

For example, just say analyze, or implement, or compile, or debug, as the last word. It will override the mode prompt instructions.

68

that’s sweet, I’m going to try that!

69

Hello everyone
Many thanks to @robotlovehuman for sharing this interesting idea.
After using modes and switching between them manually for a while, I noticed that the cursor itself switches between different modes while answering questions and fixing errors in an answer.
For example, when I sent the error text, it first entered [MODE: RESEARCH], then [MODE: PLAN], then [MODE: EXECUTE], and finally ended its answer with [MODE: REVIEW].
Does anyone know what happened?

1
1488×942 91.1 KB

2
2489×989 92.7 KB

3
3487×975 96.4 KB

70

Been doing similar for a while (a bit more simplified Agile workflow, but same concept and it worked well enough in about 10 bullets.

But - lately I have found my favorite way of working is with defining the new custom models and putting the prompt for each way of working into the custom agents.

With the ability to now have tabs, and restrict through a strong agent prompt it gives the best reliability and control I have experienced with with cursor to date - more reliable than a set of rules that the agent can nondeterministically deside what phase or mode its working in from a single set of rules all within the context.

71

Follow on last message - I am suggesting taking your ripper and make it into 5 different personas. You can even have them all open now in different tabs and easily switch back and forth. This is the closest to agent orchestration in cursor right now without using external agent swarms via mcp or whatever comes in the future.

3 Likes
72

nice and how have the results been with this 5 separated modes approach, is the A.I like vibing as expected with it, without its random side tangent bursts?

1 Like
73

@robotlovehuman
Just wanted to say I keep using this and it works great!

By the way, I see more and more mode using memory that keeps updating so that agent learn to not re-do the same error.
Any plan to update your mode to incorporate such a behavior? :drooling_face:

1 Like
74

I took the liberty to modify the code and added a memory bank framework. I have uploaded it to GitHub for ease of use. Let me know how it works for you.

4 Likes
75

That’s a great update! Add me on github with the same username RobotLoveHuman , I was working on a memory augmentation protocol to add to it, but yours is way better and cleaner, so I would rather pr to yours, there is a cool little part that I have been working on a little addition that I think will compliment everything powerfully, I would like to discuss it with you on github -

1 Like
76

Tabs are no more in the most recent cursor. I hope they bring them back but who knows

77

Hi… I’m a noob of cursor and programming, I’d appreciate it if anybody can explain to implement this rule.

How could I precisely implement the riper mode and bank memory into the cursor?

I’ve also read the readme file on GitHub and the way for process is amazing, but I’m confused about how to implement it.

Are implementing them as .cursorrules, or split riper to .cursorrules and riper_memory_framework just add them as .md files in the project?

How cursor always follow the .md files? On the other hand, has anybody transformed this mode into a custom mode feature in the cursor?

78

This is more a tutorial on the basics of cursor rules as of 0.48+ and custom agents in general - recommend starting here to get the basic understanding and working with rules and workflows GitHub - bmadcode/cursor-custom-agents-rules-generator: Maximize the potential of Cursor best practices for Automatic Rule and Custom Agent Generation and Agile Workflows

2 Likes
79

0.48.x has chat tabs :slight_smile:

80

Yep. I should have corrected this they just dont function as they normally did when initially released. Now on a keyboard binding that I was not initially using.

81

This deserves its own repro!

I havent scanned the whole discussion for updates bit this has significantly improved how inrefactor and develop. Many thanks.

The only thing this still cannot do is replicate a project edactky from nextjs to remix. I tried a few runs on it but its hallucinating components then.

I think a repro with several variantions explaining what was improved or for specific usecases would be rrally cool if the author of that is willing to make one.