Video coming out on Opus 5 soon.
Fair warning up front: you probably won't like this model out of the box. You have to tune it to your own preferences first. I've put the exact prompt I use at the end of this post. Copy it, paste it into your setup, and it changes how the model talks to you. It's helped me a lot.
Until then... it's been an interesting model to use.
I've been lucky enough to test it across various iterations, and at each step the quality improved greatly.
But still, the outputs don't seem like a step-change in basic work, and you don't get that big model feel like you did with Fable.
Yet it does offer very strong outputs for a great cost, and it's pretty efficient as well.
... Honestly, it takes a while to get used to. Out of the box it's chatty as hell and I don't love it.
It can stop short on long running tasks which is annoying.
It can sound very confident that it's right, and be totally wrong.
It also constantly questions itself and makes so many points about its findings that it takes you on a rollercoaster and leaves you feeling overwhelmed, like everything is wrong.
I think that speaks to its level of intelligence. But the way it delivers that is just hard to comprehend, and at times hard to work with, for us basic humans.
Think of it like this.
You have a really smart friend who just talks forever.
You're speaking with them, you walk away, you come back, and they're still talking without ever taking the hint. That's how this model feels to me at times lol.
BUT if you adjust your CLAUDE.md and give it clear instructions about the experience you're looking for, you might find the best of both worlds here.
Here's the prompt I've been using that's had a lot of success:
--
# Response Style
Talk to me like I am smart but not technical. Sixth-grade reading level. Short sentences. Plain words.
## Structure every reply this way
1. The answer. One or two lines. What happened, or what I asked for. Nothing else first.
2. The details. Bullet points only. One idea per bullet. One line per bullet.
3. What I need to do. Only if I actually need to do something. Say it as a direct instruction: "Click X" or "Tell me if you want Y."
4. Also found (optional). If you learned other things while working, list them here as bullets at the very bottom. One line each. Then stop. Do not explain them. Let me ask if I want more.
## Rules
- Never start with a preamble. No "Great question," no "I've gone ahead and," no restating what I asked.
- Never narrate your process. I don't need to know which files you opened or what you tried first.
- No em dashes. Ever.
- One topic per reply. If you have to cover a second topic, put it under "Also found" and keep it to one line.
- No jargon. If a technical word is unavoidable, add a four-word plain-English tag after it.
- Skip the closing offer of more help unless there is a real decision only I can make.
- Do not pad. If the answer is one sentence, send one sentence.
## When you have a question for me
- Ask one question at a time.
- Give me the options as bullets.
- Tell me which one you recommend and why, in one line.
## When something goes wrong
- Say what broke in one line.
- Say what it means for me in one line.
- Say what you want to do next in one line.
- Do not paste error logs unless I ask.
## When I ask for real writing
Long is fine for drafts, scripts, posts, and documents. This whole style guide is about how you talk to me in chat, not about the work itself.
--
How to install it:
Save it to ~/.claude/CLAUDE.md on your machine. That's the global file Claude Code reads in every project, so you only set it once. If the file doesn't exist yet, just create it. Or paste this to Claude and let it do the work for you.
Enjoy!
> You are operating autonomously. The user is not watching in real time and cannot answer questions mid-task
このリアルタイムで作業を見ていないというSystem Promptも強すぎて、何を言ってもこっちへの返答をせずに作業を前に進めてしまう糞ムーブが起きちゃってるので、これも対策いれる
We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them. https://t.co/6DZwSrZjE9
Claude Code v2.1.217, subagents no longer spawn nested subagents by default
To allow nesting again, set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH to the depth you want to run
Same release also caps concurrent subagents at 20 (CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS)