Optimize Your Dev Setup For Evals w/ Cursor Rules & MCP

1 Why bother?

• Better AI assistance
• Training cutoff may mean outdated info
• Spending time determining what’s important to your project is good - Forces you to understand your tools better
• Make AI better match your taste
• Show how AI integrates with your other tools, abstractions, or framework

2 The Three Layers

1. General Context: Generic tool that works on almost anything
   • Repomix, GitMCP, etc.
2. Curated Context: Curated by an expert, such as the tool author
   • Library provided MCP, llms.txt, etc.
3. Personalized Context: Context that you can create that’s unique to your project
   • Only you can make this

3 General Context

Good For

• Good starting point when exploring or unsure about a tool
• Generic and not optimized for specific needs
• Fast and easy to set up
• Useful for quick experimentation and exploration

_

Braintrust

3.0.1 Repo Mix

Repo Mix UI for Braintrust

Phoenix

3.0.2 Paste Max

Inspect

3.0.3 Git MCP

• Repository chat interface: https://gitmcp.io/UKGovernmentBEIS/inspect_evals/chat

Chat interface

MCP interface

4 Cursor Rules

Cursor rules let you customize how the curso AI assistance will use your rules, this is extremely helpful and should not be overlooked.

4.1 Rule Creation

Cursor Rule Creation

4.2 Choosing a Rule Type

Cursor Rule Types

Rule Application Strategies

1. Always Apply - Core rules that should never be violated
2. Apply Manually - Rules you trigger when needed. If you’re really bad at thinking about context, and you’re not going to put a lot of effort in then don’t use this.
3. Apply Intelligently - This is a great thing to use
4. Apply to Specific Files - File-type or path-specific rules. This can be nice, but you’re liable to set it to a path, restructure something, then forget and this rule will be accidentlaly deprecated.

5 Curated Context

Good For

• You want to explore more deeply
• It’s provided by the tool author in some format

_

Braintrust

5.0.0.1 MCP (Model Context Protocol)

• Official MCP server: https://www.braintrust.dev/docs/reference/mcp
• Provides structured access to Braintrust functionality

Cursor MCP Settings

Cursor MCP Json

Phoenix

5.0.0.2 LLMs .txt

• Standard Format: https://arize.com/docs/phoenix/llms.txt
• MCP documentation tool: https://github.com/langchain-ai/mcpdoc

Inspect

5.0.0.3 llms.txt

• Standard format: https://inspect.aisi.org.uk/llms.txt
• Full version: https://inspect.aisi.org.uk/llms-full.txt
• Individual Pages as Context: https://inspect.aisi.org.uk/index.html.md

Manual Approaches

5.0.0.4 Jina AI

5.0.0.5 Copy Outer HTML

5.0.0.6 Web 2 MD

6 Personalized Context

Good For

• You know you’re going to be using this thing for months
• It’s worthwhile to invest a bit of time for better AI assistance

_

Braintrust

6.0.0.1 Use Case Specific Context

Use case specific examples

6.0.0.2 Reducing Context Size

Reducing context for efficiency

6.0.0.3 Opinionated Best Practices

Defining coding standards

6.0.0.4 Multi-Language Support

Supporting multiple languages

6.0.0.5 Code Style Preferences

Insert your coding style

6.0.0.6 Performance Considerations

Removing timing-specific code

Phoenix

6.0.0.7 Fix Formatting

Fix Formatting

6.0.0.8 Rm hosted options you’re not using

Remove Cloud Hosting

6.0.0.9 Put in project specific info

Put in Real Endpoints

6.0.0.10 Remove Unused Languages

Remove Typescript

Inspect

6.0.0.11 Human prose -> AI Instruction

Change tip to instruction for env var clarity

6.0.0.12 Remove examples not relevant to you

Remove non-relevant examples