Narrative Goldmine

Coding support

9 min read

Properties

Type
  • Note
Status
  • stable
Generated
  • by: process:vault-migrate/1.0 · at: 2026-09-22T12:41:58.55607396Z

Believably wrong answers

  • Study Details by Purdue University. Presented at the Computer-Human Interaction Conference in Hawaii. (CHI)
  • 517 programming questions from Stack Overflow.
    • 52% contained incorrect information.
    • 77% were verbose.
    • 78% showed inconsistency compared to human answers.
  • User Perception
    • Participants preferred ChatGPT answers 35% of the time despite inaccuracies.
    • Misleading AI responses were not detected by programmers 39% of the time.
    • ChatGPT’s answers were more formal, analytical, and positive in tone.
    • Politeness and comprehensiveness made ChatGPT answers appear more convincing.

Specialised Models

image.png

Products

Devin

Cody

  • The AI Coding Assistant
    • Introduction to Cody
      • Developed by Sourcegraph, co-founded by Beang Liu, CTO.
      • Aims to revolutionize software development with AI.
      • Integrates into various editors, enhancing developer productivity.
    • Foundation and Purpose
      • Rooted in Beang’s early interest in AI and machine learning at Stanford AI lab.
      • Addresses the gap between the potential of programming and the drudgery of day-to-day software engineering tasks.
      • Focuses on reducing time spent on reading and understanding existing code.
    • Defining Spatial Computing
      • Initially focused on advanced search capabilities in real coding environments.
      • Aimed at achieving ‘flow’ in programming through efficient information retrieval.
    • Integration of AI in Sourcegraph and Cody
      • Shift towards AI-enhanced coding tools around 2017-2018.
      • Early experiments with applying large language models (LLMs) to code search.
      • Development driven by advancements in AI, especially in neural networks and LLMs.
    • Capabilities of Cody
      • Provides AI-driven coding assistance in various IDEs.
      • Features include inline completions, high-level Q&A, and specific coding commands.
      • Unique in augmenting large language models with contextual information from Sourcegraph.
    • Future Aspirations for Cody
      • Aims to automate more complex software development tasks.
      • Foresees the potential for AI to generate pull requests and change sets from issue descriptions.
      • Emphasizes the importance of context quality in improving code generation.
    • Technical Challenges and Innovations
      • Balances traditional information retrieval with AI-driven approaches.
      • Focuses on optimizing search architecture and context retrieval for better code generation.
      • Explores the use of small models for faster and more cost-effective solutions.
    • The Evolution of Software Development with AI
      • Envisions a future where individual developers are more productive and cohesive.
      • Anticipates changes in the software development lifecycle due to AI integration.
      • Stresses the growing importance of CS fundamentals and domain expertise in an AI-augmented future.

Advice on AI coding

  • Choose Tools Strategically: Not all AI coding tools are created equal. Select the right tool for the job, considering the project’s scope and complexity:
    • Complex Applications: Cursor, Windsurf, or more established IDE integrations (see below) are often better suited for larger, more intricate projects.
    • Micro-SaaS: Bolt/Lovable are optimised for smaller, Software-as-a-Service applications.
    • Mobile Applications: Replit remains a good choice, alongside framework-specific tools.
    • UI Design: Consider using ‘vo’ or similar specialised tools for user interface design.
    • General Coding Assistance & IDE Integration:
      • GitHub Copilot: A widely used and powerful AI pair programmer that integrates directly into your IDE (VS Code, JetBrains IDEs, etc.).
      • GitHub Copilot Agents: Extend Copilot’s capabilities with specialised agents for tasks like code review, debugging, and test generation.
      • Aider: A command-line tool that helps you write and edit code using GPT models. Good for making changes to existing codebases, particularly for refactoring and adding features.
      • Roo: Provides code generation and chat capabilities within your IDE.
      • Cline: Good for command line interfacing, and code assistance.
  • Context is Paramount: Always provide comprehensive context about your project. AI tools cannot “guess” your intentions. Use Markdown (.md) documents to detail:
    • Product Requirements Document (PRD): Clearly outlines the purpose, features, and functionality of the application.
    • Technical Stack Document: Specifies the programming languages, frameworks, libraries, and databases to be used.
    • File Structure: Defines the organisation of directories and files within the project.
    • Frontend Guidelines: Describes coding standards, styling conventions, and component structure for the user interface.
    • Backend Structure: Outlines the architecture, API endpoints, data models, and business logic for the server-side code.
    • Use CodeGuide (or Similar): Consider using CodeGuide or a similar tool to help generate and manage these AI-specific coding documents. This ensures compatibility across various AI tools and helps maintain a single source of truth.
  • Incremental Development: Avoid overly broad prompts like “build me an AirBNB clone.” Instead, break down the project into manageable steps:
    • Page by Page: Develop the application one page at a time.
    • Component by Component: Within each page, build individual components sequentially.
    • Limited Task Execution: AI models typically perform best with a maximum of 3 concurrent tasks per request. Be mindful of this limitation, and break down larger tasks accordingly. Tools like Aider and Copilot Agents can help manage this complexity.
  • Select AI-Friendly Technologies: Certain technology stacks are better understood by current AI models:
    • Web Applications:
      • React (with NextJS or ViteJS): Provides excellent performance and is well-supported by AI tools.
      • Python (with frameworks like Django or Flask): Widely used and well-understood by AI models.
    • Mobile Applications:
      • React Native: A good choice for cross-platform development.
      • SwiftUI (especially with Claude): Works well, particularly with Claude models.
    • Avoid Older Technologies: Unless absolutely necessary, as AI model support may be limited.
  • Utilise Starter Kits: Save time and reduce token usage by starting with pre-built templates or boilerplates:
    • Example: The “CodeGuide NextJS Starter Kit” can provide a solid foundation.
    • Benefit: Accelerates workflow and provides a structured starting point. Most frameworks have readily available starter kits.
  • Define Rules Within Your Tools: Many AI coding tools allow project-specific rules:
    • Examples: .cursorrules (often “project rules”), .windsurfrules, or similar configuration files within your IDE or tool. Copilot and other IDE-integrated tools often have settings for coding style and preferences.
    • Purpose: Constrain the AI, preventing deviations from your guidelines and coding standards.
    • Coding Standards: Enforce coding standards using linters (e.g., ESLint for JavaScript, Pylint for Python) and integrate their configuration with your AI tools where possible.
  • Employ a Multi-Tool Approach: No single tool handles the entire workflow seamlessly. Combine tools:
    • Research: Perplexity.
    • Brainstorming: ChatGPT (voice features can be helpful).
    • Documentation: CodeGuide, or tools integrated within your IDE.
    • Data Scraping: Firecrawl, or libraries within your chosen language (e.g., Beautiful Soup in Python).
    • Code Generation/Assembly/Refactoring: Your chosen AI coding tool (Cursor, Windsurf, GitHub Copilot, Aider, Roo, Cline, etc.). Choose based on your workflow and project needs.
  • Patience and Persistence: Working with AI requires a specific mindset.
    • Prompt Engineering: Crafting effective prompts is crucial. Experiment with different phrasing and levels of detail.
    • Expect Errors: AI models are not perfect. Be prepared for errors.
    • Iterative Refinement: Stay focused, learn from mistakes, and iteratively refine your prompts and approach.
    • Debugging: Provide the AI with the full code and error message for assistance. Leverage Copilot Agents for debugging tasks.
  • Version Control
    • Use Git for version control.
    • Commit frequently with clear messages.
    • AI can help generate commit messages (Copilot, Aider, and others offer this).
  • Testing
    • Write unit and integration tests.
    • AI can assist in generating test cases (Copilot Agents are particularly useful here). Tools like Aider can help refactor code to improve testability.
    • Agent Frameworks

other links

Random Links

https://twitter.com/tldraw/status/1782443204710674571

VSCode Agents Tips and Tricks Training Modules

  • Cursor
    • really big detailed settings structures in complex extended codebases need this
  • Cline
  • Roo Code
  • Google Gemini
    • subtle whole codebase needle in a haystack logic problems
  • make notes about what works and doesn’t in the commits
  • reversion and blend strategies

Graph View