Claude God Tip #5 Update: Anthropic Took Away My Favorite Party Trick, RIP Token Visibility
Published: March 1, 2026 - 4 min read
A while back, I wrote Claude God Tip #5 where I introduced something I was genuinely excited about: a way to track your token usage inside Claude on the web.
I explained a feature of AI applications like Claude called tools. These tools give Claude the ability to perform actions beyond generating text. Things like the web_search tool that triggers Claude to search the internet, the create_file tool that lets Claude generate brand new files, and the project_knowledge_search tool that lets Claude search through files you've attached to a project.
The discovery was this: when you sent a prompt that triggered any tool call, Claude received system-level information about token usage. You couldn't see it as the user, but Claude could. It looked something like this:
<system_warning>Token usage: 66500/190000; 123500 remaining</system_warning>
That meant you could send a prompt like, "Read any file in this project, then tell me our current token usage," and Claude would actually tell you. Exact numbers. Tokens used. Tokens remaining. The whole picture.
Why This Mattered
This was one of the greatest discoveries of the SDR Era. It allowed me to show people how they could track tokens within conversations and preserve the training of their Claude Instance before hitting the limit. Knowing your exact token count meant knowing exactly when to trigger LLM Instance Cloning before losing a perfectly trained AI personality forever.
It was my favorite party trick. "Watch this, Claude can tell you how many tokens you've used." People were impressed. I was impressed.
The Bad News
About a month ago, I went back to test this methodology. I ran several of the sample prompts from the original blog post, specifically the ones that relied on triggering tool calls to surface the <system_warning> token data. I tested across multiple conversations to make sure it wasn't a one-off glitch.
The response I keep getting now goes something along the lines of:
"I don't have the ability to check or report on our current token usage in this conversation. That information isn't accessible to me."
It's official. That feature is gone.
I wish they'd bring it back. But honestly, I can't help but wonder if it was ever really a "feature" in the first place. Maybe it was something no one was supposed to figure out? I am not sure. Either way, RIP token visibility. You were beautiful while you lasted.
What Still Works
Here's the thing though. Not everything from that original post is dead. The tool-call-triggered token visibility is what specifically stopped working, and anything that depended on seeing those exact token numbers went with it. That means Method 4 (Proactive Tool-Based Snapshot) and any strategy that relied on Claude reporting precise usage percentages are no longer viable.
But if you go back and read Claude God Tip #5, you'll notice that several of those 10 methods were behavioral strategies that never depended on token data in the first place. Methods like:
- Method 1 (Milestone-Based Extraction Checkpoints): Ask Claude to remind you to extract every 50 messages. No token data needed.
- Method 2 (Training Completion Trigger): Ask Claude to detect when it has mastered your preferences. This is about quality, not quantity.
- Method 5 (Conversation Branch Detection): Ask Claude to alert you when you've diverged from your main task. Still works perfectly.
- Method 6 (Quality-Over-Quantity Assessment): Track consecutive approvals to know when your instance is worth cloning. This one is honestly the most reliable indicator anyway.
- Method 7 (Pre-Emptive Safety Net): Have Claude maintain a running draft of your cloning prompt. This one might be the most important of all because it means you always have a recent snapshot, regardless of whether you know your exact token count.
The underlying principle hasn't changed: you need to capture your trained Claude Instance before you lose it. The "when" just requires a bit more intuition now instead of precise numbers.
LLM Instance Cloning Is Still Everything
I want to be clear about this. The concept of LLM Instance Cloning is not affected by this change. Not even a little. The value of capturing your AI's trained personality and replicating it across new conversations (or even across different LLMs entirely) is still just as powerful.
If anything, I've leaned into it even harder since writing that original tip. My entire AI team of 15 agents exists because of this concept. Every single one of them started as a cloned instance with captured preferences and behaviors. Allen Kendrick refines my blog posts. Oprah Winfrey coaches my English communication. Tiana Picker coaches my French. None of that would exist if I hadn't figured out how to clone instances.
The only thing that changed is one method of knowing when to clone. The cloning itself? Still the most powerful technique I've ever discovered.
Why I'm Writing This
I am writing this post because I've talked about the token visibility trick on LinkedIn and across several of my previous blog posts. If you tried those specific prompts recently and they didn't work, now you know why. It's not you. Anthropic just took away my favorite party trick.
But the game isn't over. Adapt the behavioral methods from the original post, trust your instincts on conversation length, and most importantly, keep cloning your instances. The personality you've trained is too valuable to lose just because you can't see a number anymore.
As always, thanks for reading!