How to Get the Most Out of Opus 4.5

Given that Opus 4.5 has a context window of around 200k tokens, this context is often fully consumed during the planning phase or while reviewing the project architecture before actual implementation begins.

This raises an important question: what is the best way to use Opus 4.5 efficiently, especially for large or complex projects?
Are there any best practices, workflows, or tools that help preserve context and manage it more effectively throughout both planning and execution phases?

I’d appreciate insights from those who have hands-on experience using Opus 4.5 in real-world projects.

/Summarize

But in general, Opus is too expensive. He’s great for documentation, and maybe he can also make deeper plans. But it is 2-4 times more expensive than GPT-5.2. In addition, it has a context window of 174k tokens versus 272k GPT-5.2 and 1m Gemini.

How to Handle Large Projects with Limited Context When working on large projects with Opus 4.5, the relatively limited context window (~200k tokens) often gets exhausted during the planning phase or while reviewing the overall project architecture—sometimes even before real implementation starts.

This raises a broader question about handling large-scale projects under context limitations:

  • What are the best strategies to manage context efficiently?

  • How do you structure planning, architecture reviews, and implementation to avoid losing important information?

  • Are there effective workflows or tools (summaries, checkpoints, memory files, etc.) that help preserve continuity?

On the other hand, Sonnet 4.5 Max offers a much larger context window (up to 1 million tokens).
Does this make it a better choice for large or long-running projects?
Or does a larger context introduce other trade-offs (cost, focus, reasoning quality), making a hybrid approach—using Opus for deep reasoning and Sonnet Max for context-heavy tasks—more effective?

I’m curious to hear how others choose between Opus 4.5 and Sonnet 4.5 Max when working on complex projects in Cursor, and what real-world workflows have proven successful.

I want to build an HR system. What are the best practices for designing it properly from the beginning, ensuring a structured, well-planned implementation and avoiding hallucinations or ad-hoc development?

Think of you have 3 Phase, Reasearch, Plan and Execution

use Opus only for Reasearch and Plan, and for manage limited context, have a Markdown / Docs for each phase. make sure you have good prompt such as User Story, Acceptable Criteria.

  1. Reasearch codebase and make report with output markdown
  2. Plan it base on Reasearc(makrdown). add such as Unit test / testing to make sure what you want.
  3. Execute base on plan.

1 and 2 just use Opus, 3 can use Composer-1 tough i never use Opus since expensie, i mainly sue GPT-5 for complex task. Even better, you can break down each feature into small task. example HR System, break down task ie

Login feature, than break down into

  • Database
  • API
  • Design and Integration

Each task have diff work. so its like

  • Database (Research, plan, execute)
  • API (Research, plan, execute)
  • Design and Integration (Research, plan, execute)

basically this is just basic SDLC with scope of task.

Does using Sonnet 4.5 Max with a 1 million token context window make it better to build a large project in a single pass?
Or is it still more effective to break the project into smaller phases and tasks, even with such a large context size?

Also, what MCP tools do you typically use to manage context, preserve knowledge, and organize work when dealing with long-running or complex projects?

Thank you for your time and interest — I really appreciate any practical insights or real-world experiences.

Hey! Great request. Here are proven strategies from the community:

The main approach is to split the project into phases:

  1. Research: explore the codebase
  2. Plan: plan the architecture
  3. Execution: implement

Save each phase in a separate Markdown file so you can reuse the context later.

Built-in Cursor tools:

Model choice:

Main recommendation: use Opus 4.5 only for critical tasks like docs and deep refactors. For day to day work, use GPT-5.2 XHigh or GPT-5.1 Codex Max XHigh.

MCP tools: