[feat] “Let Claude Think First” Prompt for Sonnet & Haiku

[dsnehitha]

Describe the feature

While trying out Onlook last week, I noticed responses from Claude Sonnet and Haiku are a little sluggish. Digging into Anthropic’s docs, I realised why: by default, Claude Opus gives a pause to think before it uses tool use query to determine necessary tools, but Sonnect and haiku don't - these two models are prompted to fire off tool calls as much as possible (~unnecessary tool calls). That guess-first mindset adds latency, burns tokens, and sometimes derails the flow.

Anthropic recommends a light “chain-of-thought” guardrail that makes the model think before tool calling.

I’d like to add a ~150-200-word version of this prompt to Sonnet/Haiku’s system messages.

Outcome: fewer tool calls, faster answers, happier users.

Let me know if you’re on board, and I’ll spin up a PR!

[Kitenite]

I like it, let's give it a shot! @dsnehitha

[Kitenite]

There's a thinking config you can use for Sonnet 4 as well instead of doing custom COT
https://ai-sdk.dev/providers/ai-sdk-providers/anthropic#reasoning

[dsnehitha]

ok, let me check that

[dsnehitha]

Here's my observation after trying both ways-

1. Thinking configuration is robust and gives reasoning output for debugging etc. But as a Tier 1 user, I have very little ITPM, OTPM, and had to set budget of thinking tokens to 2000 only(thinking tokens count towards output token). Total tokens consumed are 69k.
2. Whereas CoT prompting takes 200 input tokens, and the overall tokens consumed for the same prompt are 48k.

I believe CoT can be implemented to concentrate on optimising tool calls, but if we need an overall better reasoning then thinking config would suit.

Let me know your thoughts on this