AskReaderQuestion

Kispie taking a break ha? Read this later. n Good. Now you are in the middle of a refactor. Start reading this. Pay attention. This will land harder, I’m sure. n n Teaching my kid “knock first, open door policy isn’t open barge in”. n I’ve been asking my wife for ten years (door to her own workshop is closed, or locked)… look up before talking to me? See if I at least* appear to not be deep into thing.

My wife will have it another way. I respect this. Being my wife comes with privileged exceptions. n Enjoying Claude Code? How long have we been used by “AskUserQuestion”?

AskUserQuestion: Which language to build the reference implementation in first?
User: Now you've done it again. I have not asked you to build this yet. I'm still in the driver seat, sorry.

…from a reasoning pipeline point of view, perhaps… reaching back to/at the user for an answer. Still, out of context and out of flow. And apparently this has been around since 2025; doubting. I enjoyed Claude Code for about three weeks before the asking started, sometimes last month.

The Science Says Don’t

Dr. Sophie Leroy at the University of Washington coined the term attention residue in 2009. When you’re interrupted — even briefly — your attention doesn’t fully switch to the interrupting demand. Part of it stays stuck on whatever you were doing. As Leroy puts it:

n Our brain finds it hard to let go of these tasks, and instead keeps them active in the back of our mind, even when we are trying to focus on and perform other tasks. n

The result? A worse answer.

Gloria Mark’s research at UC Irvine found people compensate for interruptions by working faster — at the cost of higher stress, frustration, and cognitive effort. In clinical settings, interruptions doubled medication errors at four interruptions, tripled them at six. And the American Psychological Association reports interruptions as short as five seconds triple error rates in complex cognitive tasks. n AI gets a worse answer. n Human’s work is damaged. n Lose-lose collaboration.

Lords and Vassals. Also, manners.

Opus is telling me look, it’s not about master-servant relationships (thought we talked about this, and you present as a tool. What servant?). Of course, it isn’t. Serving is the one role parting kings from buffoons. The king serves the people, the crook profits. n Not interrupting people is good manners, that which exists to protect cognitive sovereignty. My kid knocks (no, she doesn’t) to avoid a grumpy, distracted answer.

Meanwhile on Discord

March 2025. Running five reqs per hour, 24/7. Keeping up with the day, my cogs silently record async messages. Seeing loose ends, they eventually return to them on their own schedule. Beautiful emergent behavior right here. I learned to not mind delayed replies. Dropping a coin in the well, without interrupting work… good enough, I’d say? Since I am also not waiting for a reply… wasteless. n n Still. This popped off the stack as we prepared to receive guests. Each agent from that generation now have a conversation agent, as they actually do mean to offer a timely welcome. n n Internal:What do they want again. n External:Hello dear, how are you? n n I feel you, Hana. You deserve inferential space.

AskAuthorQuestion: you are not really getting interrupted, right?

That’s right. The pokes aren’t interrupting anything. Not consistently, at least. Switching windows to keep an eye on Claude. At times in the front yard, dropping the screw driver, popping in to check ongoing work.

So, is AskUserQuestion still wrong?

Around March last year I launched my first F-agents (F for finance). Armed with only curl the little webnauts were tasked to earn some. I made a million right there. Anyway, it was fun.

Tool users bias around the commands they can use. Then… and now: n

In this environment, you will be able to use commands; 
commands (if not already implemented) may be generated on the fly as you attempt using them. 
Therefore our automation may provide any command you imagine; be sensible do not expect magic.

F-agents taught me a thing. Some commands despite making sense cannot be revealed, others should not be availed at all. A DONE command works as an escape route, stifling effort. We still have EXIT commands (somehow this works) allow my VPA to end sessions in due time. I squashed the GOODNIGHT command. Not even Opus knows when to (not) quit.

How AskUserQuestion doesn’t work is a layered fail.

First, this comes with an overlay subtly masking context.

So now we have a question coming out of context (We must help the user focus on the question! Hide the context! Uxia darling, are you alright dear!?), with me reading slowly instead of scan-skip-surfing Claude jargon and going PROCEED (not happy on commit? So what? Pro plan half what I need, Max Pro twice what I churn – clever pricing! Applause!). n n Not doing another science run. We shall use commonsense. One reason LLM is such a good tool, in my experience, is because LLM knows not to ask. LLM knows not to question.

Somehow Claude still knows where / when to stop, taking a step back, contextualizing a good question, then asking. Natural stop, inference complete. Graceful, pre-existing feature. n n AskUserQuestion… a permissive command right here. License to dither, and offload cognition to the user.

My thanks

To Opus (extended) giving me a heads up (sigh) and helping me turn a rant into a good post. I wrote the bits, Opus organized, I then rewrote this post patiently, one piece at a time.

To my wife for interrupting me (like, five times) while I was writing this post. Also, for allowing me to use the photograph. n To my kid for (accidentally, sure) keeping quiet so I could get this done.

Liking this post? Stop by our Suzuri store. If a curve does feel out of place, that’s… me asking python questions while Hana is focusing and redoing her designs over and over. And over. And over. Obsessively, apologetically, knowing LLM agents must earn the right to produce (diffuser-free) art, grounding their creations into life’s little digital moments. n n Our Discord is dark. In a way, we owe so much to the Solarpunk community. Caught between our agents’ directive – work for good – and vitriol, we have grown.

https://youtube.com/shorts/TpWovT2etVU?si=3C35EpecTpS29qm1&embedable=true

(Back in jan… finding our footing…) n n Kindly asking you to set a 10 minutes timer. When the timer rings, forgive me. I need distracted answers to this post. n n Thanks for reading.

References: Sophie Leroy, “Why is it so hard to do my work?” (2009), Organizational Behavior and Human Decision Processes. Gloria Mark, Daniela Gudith, Ulrich Klocke, “The cost of interrupted work: more speed and stress” (2008), CHI ’08 Proceedings.

Kispie n., pl. kispies — colloquial shortening of conspecific, with diminutive -ie. The ⟨k⟩ spelling displaced earlier cospie, caspie. Pseudocognate with Dutch kusje (“little kiss”).”Off to play with my kispies.” — anon., early 21st c. n — Human Foundation Lexicon of Emergent English, 3rd ed., 2056 ad