G3 - AI

uses

• understanding code
• problem fixing
• code/idea generation
• context-aware code completion
• improving code quality (type hints, comments, docstrings)

image: prompt engineering workflow, Generative AI on Vertex AI, Google

• effective prompt: content + structure

• components:

  • objective
  • instruction

• additional optional components include system instructions, which the model processes before the prompts:

  • persona (role/vision: who or what the model is acting as)
  • constraints
  • contect
  • few-shot examples
  • reasoning steps
  • response format
  • recap
  • safeguards

<OBJECTIVE_AND_PERSONA>
**You are a [insert a persona, such as a "math teacher" or "automotive expert"]. Your task is to...**
</OBJECTIVE_AND_PERSONA>

<INSTRUCTIONS>
**To complete the task, you need to follow these steps:
1.
2.
...**
</INSTRUCTIONS>

------------- Optional Components ------------

<CONSTRAINTS>
**Dos and don'ts for the following aspects
1. Dos
2. Don'ts**
</CONSTRAINTS>

<CONTEXT>
**The provided context**
</CONTEXT>

<OUTPUT_FORMAT>
**The output format must be
1.
2.
...**
</OUTPUT_FORMAT>

<FEW_SHOT_EXAMPLES>
**Here we provide some examples:
1. Example #1
Input:
Thoughts:
Output:
...**
</FEW_SHOT_EXAMPLES>

<RECAP>
**Re-emphasize the key aspects of the prompt, especially the constraints, output format, etc.**
</RECAP>

source: Generative AI on Vertex AI, Google*

• prompt format includes style and structure
  • different models may respond better to specific formats
  • influences how AI interprets the request
  • natural language questions, direct commands, or structured inputs with specific fields
• context and examples
• fine-tuning the AI model on specific tasks using tailored prompts, and adapting prompts based on user feedback or model output
• multi-turn conversations

prompt types

• direct / zero-shot
  • direct instruction/question with no additional context or example
  • eg: idea generation, brainstorming, summarisation, translation
• one-, few- and multi-shot
  • provide model with additional examples before the actual prompt
• chain of thought (CoT)
  • break down complex reasoning into a series of intermediate steps
• zero-shot CoT
  • ask model to perform reasoning steps

issues with prompts

• overt manipulation attempting influence using emotional appeals, flattery or artificial pressure
• conflicting/redundant/irrelevant instructions and examples
• missing output format specification and role definition
• under-specification
• task outside of model capabilities or overloading
• non-standard data formant
• incorrect chain of thought order
• conflicting internal references
• prompt injection risk (check for explicit safeguards surrounding untrusted user )

prompt writing strategies

source: Google Cloud

• setting clear goals and objectives
  • using action verbs
  • setting the desired length and format
  • specifying the target audience
• providing context
  • including relevant facts and data
  • referencing specific sources
  • defining key terms and concepts
• using few-shot prompting
  • using XML-like markup and consistent formatting
• being specific
  • using precise language
  • quantifying requests
  • breaking down complex tasks
• iterate and experiment
  • trying different phrasings, keywords and prompt lengths
  • adjusting the level of detail and specificity
  • guiding model through the logic
• leveraging CoT
  • encouraging step-by-step reasoning
  • asking for explanations of reasoing process

WindSurf

• the first agentic IDE

  • agentic AI: autonomous artificial intelligence systems that can set goals, make decisions, and take actions with minimal human intervention to achieve objectives; unlike traditional AI that simply responds to commands, agentic AI proactively plans, reasons, and executes multi-step tasks, often using tools and adapting to changing conditions

• copilots are generally confined to short scoped tasks, and are great as they keep the human in loop

• agents act independently to complete much larger tasks, but the tradeoff is the loss of collaboration

• WindSurf combines the best of both worlds: collaboration + independence

• they call this system AI flows - collaborative agents - providing a new paradigm of working with AI

• forked from VS Code

• respect people's native editors

image: windsurf

Cascade

• WindSurf's AI code platgorm

image: windsurf

• flow awareness: action tracking to infer intent

• knowledge base: direct int4egration into repo and curated docs to build code understanding for accurate reflective solutions

• autocomplete lines as well as entire snippets predicting your intent - super complete

• predicts the location of your next edit and jumps there with a tab press, eg: variable name change, imports

• inline and terminal commands (cmd+i) using natural language

• 'codelenses' : explain, refactor, add docstring

image: windsurf

image: cursor