Technical Context
I appreciate small features like this more than major announcements. The /btw command in Claude Code falls into this category: it doesn't revolutionize the market, but it noticeably improves daily work.
The concept is simple. I can ask a short, clarifying question during a main task, and this query doesn't get added to the session's main context. The history doesn't swell, tokens aren't wasted on trivial things, and the main workflow isn't cluttered with auxiliary remarks.
According to documentation and community discussions, this feature is part of Claude Code 2.1.72+ from March 10, 2026. Internally, it functions as a temporary, read-only episodic agent: it uses the existing session context but doesn't write its response back into the main thread.
And here's a crucial detail: /btw must be written along with the question. The format is like /btw what was meant in this function?. Just typing the command alone won't trigger this scenario.
I liked that the answer opens in a separate side panel and leaves no trace in the main dialogue after being closed. For long coding sessions, this is a real gift. Especially when I'm deep into a project's context and don't want to corrupt it with a question like, "what was the syntax for os.walk again?"
There's a limitation: /btw doesn't fetch new external information or run full-fledged tools. It's not a mini-research tool or a substitute for sub-agents. If I needed to retrieve something from outside or perform a separate workflow, I'd use subagents, not this.
As far as I can see, Anthropic hasn't released any major official benchmarks. However, the community estimates that these side clarifications can significantly cut the cost of long sessions, sometimes by tens of percent. The logic is sound: less junk in the history means you're not repeatedly paying for that junk in subsequent steps.
How This Changes Real-World Work
I wouldn't sell /btw as a "revolution." But for the AI architecture of workflows, it's a very useful building block. When a team constantly works with an agent in the CLI, the main problem is often not the model's quality but the context bloating from minor clarifications, self-checks, and side thoughts.
This is where the new command hits a major pain point. The main thread stays clean for code, edits, and sequential task execution. All the minor cognitive dust is moved to a side channel.
The primary beneficiaries are developers who spend time in long sessions: refactoring, debugging, migrations, and maintaining old repositories. In these cases, one bad stream of clarifications can quickly turn good context into a mess. /btw simply prevents that mess from being created.
The only ones who lose out are those who expected a full-fledged parallel research tool. There's no magic here. If a question requires new data, access to tools, or an autonomous task run, a different pattern is needed.
I also see this as a good signal for those building AI automation around development. Not all improvements come in the form of a new model. Sometimes, real savings come from the mechanics of context management: where to store state, what to write to history, what to move to an ephemeral layer, and how to avoid paying repeatedly for random noise.
At Nahornyi AI Lab, we constantly face this when designing agent-based scenarios for teams and products. Proper AI implementation isn't just about choosing a cheaper or smarter model. It's also about correctly organizing memory, roles, side queries, and the boundaries between the main workflow and temporary clarifications.
In short, /btw is a small command with a very mature idea inside: context must be preserved. And the longer your sessions, the more noticeable the effect.
This analysis was done by me, Vadym Nahornyi of Nahornyi AI Lab. I personally build and debug AI solutions for businesses where the focus is on reasonable cost, predictability, and real-world artificial intelligence integration into processes, not just fancy demos.
If you want to apply this approach to your stack, contact me—we'll figure out where your tokens are leaking, how to make your AI automation cleaner, and what's actually worth implementing.