I keep watching people throw every single task at the same Claude model and then wonder why their outputs feel off, their tokens vanish, or their files come out messy. There’s a better way, and I just stumbled onto a brilliant decision tree that finally makes model selection feel obvious.
This savvy professional on LinkedIn laid out a 5-step flowchart for picking between Sonnet 4.6, Haiku 4.5, and Opus 4.7 based on what you’re actually trying to do. I was blown away by how clean the logic is. No guessing, no wasted tokens, just a clear path from your prompt to the right model.
Here’s how the original poster breaks it down, step by step.
Step 1: Is it more than a quick answer?
This is the gatekeeper question. Before you even think about which model, ask yourself if the task needs real thinking or just a fast response.
- No: Go straight to Sonnet 4.6.
- Yes: Move to Step 2 for deeper routing.
Rationale from the author: most tasks are lighter than we think. Defaulting to the heaviest model for every question burns tokens and slows you down.
Step 2: If you pick Sonnet 4.6
Sonnet is the workhorse. The expert lays out three common scenarios where it shines:
Simple task. Use a tight, structured prompt like:
You are a [role]. [Task] this [input]. Keep it under [length]. Tone: [casual/formal]. No preamble. Just the output.
Create files (Excel, docs, slides). The original poster suggests this prompt template:
Create a professional Excel spreadsheet (.xlsx) for: [PURPOSE]. Context: [WHO / HOW USED]. It should cover: [LIST]. Rules: Use Excel formulas, never hardcoded calculations. Put editable assumptions in their own labeled cells.
Connect with more connectors. Slack, Google Drive, Notion, Figma, and 50+ more integrations fit naturally with Sonnet for day-to-day work.
Step 3: Does it require working with files?
If your task is complex, this question splits the path.
- No: Use Haiku 4.5. It’s lightweight and saves tokens.
- Yes: Use Opus 4.7. It’s the reasoning model for heavy lifting.
Rationale: Haiku handles chat-heavy work beautifully. Opus earns its keep when you’re reading long files, doing multi-step reasoning, or producing something that needs real depth.
Step 4: If you pick Haiku 4.5
This is where the author’s token-saving philosophy really clicks.
- Chat only. No files needed.
- Turn on web search for fresh info.
- Plan in Chat, build in Cowork.
Prompt example from this industry pro:
I want to [desired results] with [constraints]. Ask me questions using AskUserQuestion before you start.
Token-saving habits the creator recommends:
- Edit your message, don’t send a follow-up.
- Batch tasks into one message.
- Say “No commentary. Just the output.”
- Convert PDFs to .md before uploading.
- Turn off Search and Extended Thinking when you don’t need them.
- New topic = new chat. Always.
- Keep files under 2,000 words.
I think these habits alone could cut most people’s token usage in half. They’re the kind of small moves that compound fast.
Step 5: If you pick Opus 4.7
Opus is for the big stuff. Here’s the author’s setup checklist:
- Turn on Extended Thinking. Always.
- Upload .md files, not PDFs.
- Open Cowork and select your folder.
Prompt example the expert shares:
I want to [task] so that [success looks like]. Read the uploaded files completely. DO NOT start yet. Ask me clarifying questions (use AskUserQuestion) so we can refine the approach step by step.
After Cowork, the creator suggests:
- Download your file.
- Save session-notes.md for continuity.
- Start a fresh session to save tokens.
Why this matters
Claude is not one tool. It’s three models. Most people use the wrong one for every task.
That line from the post hit me hard. I’ve been guilty of throwing Opus at one-liners and Sonnet at tasks that needed real reasoning. Matching the model to the job changes everything, from output quality to how long your context window lasts.
If you work with Claude daily, this decision tree is worth bookmarking. Run every prompt through it for a week and watch your workflow tighten up. Check out the full LinkedIn post for the complete flowchart and more prompt examples from the original creator.