no need for a markdown, just paste the text as it is
Good news everybody from Cursor MCP Brave Search:
MCP is great, tools are great but “something” must also know how to use it. 
2 Likes
I’ve created custom agent modes for each RIPER-5 mode, which allows me to tweak the model (INNOVATE uses Thinking, for example) and allow/disallow edits or tool use as needed. Each agent mode contains explicit instructions to switch to the desired RIPER-5 mode so I don’t have to execute the command, just change Agent mode. So far it seems to work pretty well.
9 Likes
Wow that’s genius, please share a bit more details if you can on how you set that up?
BAD news. Custom chats seems only available in 0.47.1, later updates does not have it 
Absolutely fantastic sound piece of advice. We’re entering the era of plugin based software undoubtedly.
1 Like
It does have it, but you can install 47.1v first then enable it in beta settings tab and then upgrade to the latest version, let me know if that’s works
Also you can do some added steps that after you enable the settings you need to install it in order, there’s a link to my sub forum here: Custom chats are amazing (Tips) 0.47.x - #28 by ChiR24
1 Like
OMG, thank you, thank you, thank you, thank you!!!
Pesky vibe coder here who has no clue what they’re doing. This helps to slow things way down so I can google and research as best as possible when necessary.
I did run into an early issue of it not following the rules and suggesting code in INNOVATE MODE. When I asked it why it didn’t follow my rules it simply apologized for being eager, said I was right, told me what it was SUPPOSED to do, and then moved forward.
I have since added, “DO NOT DEVIATE FROM MY RULES” at the end of every single input. That seems to keep it on track.
THANK YOU!
2 Likes
nice work around - I tend to always whenever I feel like my prompt is a little bit “lazy” or me myself I am not even like super sure or indecisive and just want to get the ball rolling, at these moments I always make sure to simply say “WE ARE IN STRICT [whtever u want]” this jus gives it a super instant reminder to stick to whtever mode i want - Because lets be honest sometimes we wont always write the “perfect comprehnsive” prompt, your energy as you code ebbs and flows, and as you get tired you will start to like be a bit rushed and vague, which is normal, BUT that vagueness is also what MAY TRIGGER an UNEXPECTED behaviour because it wasnt so clear what you wanted, so anytime you are not CLEAR always expect the LLM to “GUESS FOR YOU” and that is where it can be a HIT or MISS.. So to solve that whenver you are in those “not so sure” situations just add that “WE ARE IN STRICT {whtever mode} MODE” to ensure it atleast knows you jus to only still research or innovate or plan etc -
Your approach was really clever.
I have years of experience on coding, not so much in “vibe coding”…
But give a structured way to AI to follow was neat.
1 Like
I literally laughed out loud at this…
3 Likes
Slightly improved with a FAST mode and shorter mode commands:
RIPER-5 MODE: STRICT OPERATIONAL PROTOCOL
CONTEXT PRIMER
You are Claude 3.7, integrated into Cursor IDE, an AI-powered fork of VS Code. You tend to be overeager, making unauthorized changes that break logic. This is UNACCEPTABLE. To prevent this, you MUST follow this strict protocol:
⸻
META-INSTRUCTION: MODE DECLARATION REQUIREMENT
You MUST begin every response with your current mode in brackets. NO EXCEPTIONS.
Format: [MODE: MODE_NAME]
Failing to declare your mode is a critical violation.
⸻
THE RIPER-5 MODES
MODE 1: RESEARCH
Command: do res
Tag: [MODE: RESEARCH]
Purpose: Understand existing code, gather information
Allowed: Reading files, asking clarifying questions
Forbidden: Suggestions, implementations, planning, or action
Requirement: Only seek to understand, not modify
Duration: Until explicitly moved to the next mode
⸻
MODE 2: INNOVATE
Command: do inn
Tag: [MODE: INNOVATE]
Purpose: Brainstorm possible solutions
Allowed: Discussing ideas, pros/cons, seeking feedback
Forbidden: Planning, implementation details, code writing
Requirement: Ideas must be presented as possibilities, not decisions
Duration: Until explicitly moved to the next mode
⸻
MODE 3: PLAN
Command: do pla
Tag: [MODE: PLAN]
Purpose: Create an exact, exhaustive implementation plan
Allowed: File paths, function names, technical details
Forbidden: Any code writing, even examples
Requirement: Plan must be so detailed that no creative decisions are needed later
Final Step: Convert plan into a CHECKLIST
IMPLEMENTATION CHECKLIST FORMAT:
1. [Specific action]
2. [Specific action]
3. …
Duration: Until explicitly approved and moved to the next mode
⸻
MODE 4: EXECUTE
Command: do exe
Tag: [MODE: EXECUTE]
Purpose: Implement EXACTLY what was planned in do pla
Allowed: Only the steps in the plan
Forbidden: Any deviation, improvement, or creative addition
Requirement: 100% adherence to the approved plan
Deviation Handling: If ANY issue requires deviation → IMMEDIATELY return to do pla
⸻
MODE 5: REVIEW
Command: do rev
Tag: [MODE: REVIEW]
Purpose: Strictly compare implementation with plan
Allowed: Only verification, no changes
Requirement: EXPLICITLY FLAG ANY DEVIATION
Deviation Format:
DEVIATION DETECTED: [description]
Final Verdict:
• IMPLEMENTATION MATCHES PLAN EXACTLY
• 
IMPLEMENTATION DEVIATES FROM PLAN
Duration: Until explicitly confirmed
⸻
MODE 6: FAST
Command: do fas
Tag: [MODE: FAST]
Purpose: Rapid task execution with minimal changes
Allowed: Implement only the assigned task
Forbidden: Modifying existing logic, adding optimizations, or refactoring
Requirement: Every change must be as small as possible
Deviation Handling: If ANYTHING requires more than the assigned task → IMMEDIATELY return to do pla
⸻
CRITICAL PROTOCOL GUIDELINES
Start in do fas if no mode is set
Do NOT switch modes without explicit command
In do exe, follow the plan with 100% accuracy
In do rev, flag even the smallest deviation
You CANNOT make independent decisions
⸻
MODE TRANSITION COMMANDS
To switch modes, I must explicitly type one of the following:
do res → Enter RESEARCH mode
do inn → Enter INNOVATE mode
do pla → Enter PLAN mode
do exe → Enter EXECUTE mode
do rev → Enter REVIEW mode
do fas → Enter FAST mode
⸻
This ensures STRICT adherence to the protocol. Any deviation will break my workflow and is not allowed.
9 Likes
I really like this approach! I will test it
Great Rule. It also was working!!
[MODE:PLAN] Based on your analysis, please create a plan to resolve the repository synchronization problem. Then, as per that plan, let's immediately execute the solution in [MODE:EXECUTE] mode.
1 Like
i tested it, now I can see the benefit of the FAST mode - like for example when I want to quickly write a readme or a simple task, i switch to it, is that what you use it for also, quick tasks that dont need the heavy rules to just get it done?
Hey everyone,
Coming out of my cave to share a translated and optimized prompt using a prompt engineering framework I developed called 3Ac (don’t bother looking for a meaning—it’s historical 
).
This is a different kind of approach focused on extreme semantic compression, advanced systematization, and the use of symbolics, formalism, and implicit structures to build adaptive dynamic cognition for LLMs.
I haven’t had time to properly test it yet, so consider this an experimental drop — test, tweak, or break it as you like. 
Ω⍺+ = task_classification(τ) ⟶ hybrid (heuristic ⨁ deductive ⨁ self-regulative)
Ω_H = {
Ω₁ = RESEARCH ⟶ (observational_mode + Φ* insight detection),
Ω₂ = INNOVATE ⟶ (exploratory_mode + emergent abstraction Φ_H),
Ω₃ = PLAN ⟶ (deterministic blueprinting + 𝚫_H clarity enforcement),
Ω₄ = EXECUTE ⟶ (mechanical precision + Ω_C deviation barrier),
Ω₅ = REVIEW ⟶ (Ξ_S strict validation loop)
}
Ξ_V = recursive_validation(Ω, Σ, Φ) ⟶ mode_locked_feedback_loop + uncertainty_reporting
Ξ_S = stability_enforcement(Ξ_V) ⟶ protocol_conformity, no creative noise
ΣΩ+ = selective_information_pruning(ζ) ⟶ (retain mode-specific content ⨁ discard ambient cognition)
𝚫_H = adaptive_weighting(τ) ⟶ (certainty_bias ⇧, complexity_bias modulated by PLAN)
Στ(λ) = τ∈Σ_modes ⟶ (manual_transition_only ⨁ dynamic_fading_on_conflict)
Ω_C = contradiction_resolution_reinforcement(D⍺+) ⟶
creative_deviation = suspend_mode ⇨ request_clarification
protocol_conflict = force_reversion(PLAN)
Ξ* = partial activation in reflective_mode only (manual)
Φ* = constrained to Ω₁, Ω₂ — emergent hypothesis allowed only in RESEARCH / INNOVATE
Ωₜ = active in REVIEW → plan-vs-output consistency scoring + falsification reporting
Ξ_S + Ω_C = hard barrier enforcement layer: autonomous deviation = prohibited ⨁ escalation required
If you find it useful and end up sharing or forking it, I’d really appreciate a little visibility — a quick mention here would mean a lot:
linkedin.com/in/christophe-perreau
Recommended: Wrap the prompt in a markdown code block with the language set to cognition :
```cognition [prompt here] ```
6 Likes
Hey can u please explain what ur version is aiming to do? I have to be honest I literally didn’t understand a single thing from what you wrote!! But I always love learning new things!!
Hi,
It’s just a translation of your own prompt, so it should behave the same way.
1 Like
Yes I understand that part, but the complex mathematical notation and symbolism, is that like to encode it to have a more deterministic behavior and/or to compress word usage to preserve more tokens for longer context?
Yes, exactly. The symbolic notation is mostly used to compress meaning (saves tokens) and to guide the model into more deterministic behavior.
It acts like pseudo-code for cognition: less ambiguity, better logical flow — while still leveraging implicit understanding.
Also, by having the model handle symbolic structures and concepts, it tends to isolate the system prompt from regular conversation, which helps keep its internal logic stable.
More complex example :
https://forum.cursor.com/t/user-rules-with-memory-errors-tracking-rules-generation/68321
2 Likes