nerdimite

Agentic System Prompt Design by Cursor

A deep-dive into how Cursor constructs the system prompt for its AI coding agent — using a real, live system prompt as the working example. Written for developers who want to understand the machinery so they can use it better.

---

Why This Matters

When you open a chat in Cursor and type a message, you're not just talking to a raw language model. Cursor assembles a carefully structured system prompt before your message ever reaches the model. That prompt is the difference between a generic chatbot and a context-aware coding agent that knows your repo, your rules, your tools, and your current editor state.

nerdimite

transcript1 = [
    {"agent": "Thank you for calling ABC delivery. How may I assist you today?"},
    {
        "customer": "Hello, I'm calling about my delivery. It was supposed to arrive yesterday but it's still not here."
    },
    {"agent": "I'm sorry to hear that. Can I have your order number please?"},
    {"customer": "Sure, it's 123456."},
    {
        "agent": "Thank you. Let me check on the status of your delivery. Please hold for a moment."
    },

nerdimite

'''
Function: Calculates the number of lines of code in a directory recursively
Author: Bhavesh Laddagiri
'''

import os
import argparse

# append more extensions as per your project
code_extensions = ['py', 'js', 'css', 'html', 'graphql', 'sh', 'json']

nerdimite

def optimize_GA(pop_size, max_generations, crossover_prob, mutate_prob):
    '''Returns the final solution by optimizing using genetic algorithm'''

    # Initialize the Population
    population = initialize_pop(pop_size)
    global_best = {}

    # Start Evolution
    for g in range(max_generations):

nerdimite

def mutation(individual):
    '''Mutates the DNA of a child/individual by swapping the values at two positions'''
    # Selecting the index values to swap
    pos_1 = rnd.randint(0, len(individual)-1)
    pos_2 = rnd.randint(0, len(individual)-1)
    # Init the mutant
    mutant = individual.copy()
    # Swap
    mutant[pos_1] = individual[pos_2]
    mutant[pos_2] = individual[pos_1]

nerdimite

def crossover(parent1, parent2, prnt=False):
    '''Returns the child after crossover between the parents'''
    # Select cut points
    c1 = rnd.randint(0, len(parent1)-2)
    c2 = rnd.randint(c1+1, len(parent2)-1)
    
    # Create an Empty Child DNA
    child = ["X"] * len(parent1)
    
    # Set the values between the cut points from parent1 in the child DNA

nerdimite

def selection(population):
    '''Selects an individual randomly based on its fitness value as weights'''
    # Calculate the fitness values of every individual in the population
    fitness_values = list(map(fitness, population))
    # Calculate the relative inverse fitness as our goal is to minimize the total distance
    inverse_fitness = [max(fitness_values) - x for x in fitness_values]
    # Assign weights to each individual based on the inverse fitness
    weights = [x / sum(inverse_fitness) for x in inverse_fitness]
    # Select a individual as a parent with weighted randomness
    selected_individual = rnd.choices(population, weights)[0]

nerdimite

def fitness(solution):
    '''Returns the total travel distance for a given solution'''
    total_distance = 0
    
    # Calculate the acyclic distance from first city to last city
    for i in range(len(solution)-1):
        p1 = city_map[solution[i]]
        p2 = city_map[solution[i+1]]
        total_distance += euclidean(p1, p2)
    

nerdimite

cities = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

def initialize_pop(pop_size=10):
    population = []
    for i in range(pop_size):
        individual = cities.copy()
        rnd.shuffle(individual) # shuffle the cities
        population.append(individual)
    return population

nerdimite

author: Bhavesh Laddagiri summary: A Sample CodeLab id: sample-codelab categories: codelab environments: Web status: Published feedback link: https://github.com/theneuralbeing/theneuralbeing.github.io/issues

Sample Codelab