question pattern analysis

analysis of 4,600 QUESTION-labeled messages across threads.

question type distribution

type	count	%
OTHER	996	21.7%
IS/ARE	786	17.1%
WHAT	715	15.5%
CAN	701	15.2%
HOW	501	10.9%
WHY	499	10.8%
DO/DOES	289	6.3%
ANY	81	1.8%
WHERE	32	0.7%

“CAN” and “IS/ARE” questions dominate — users often ask about capability/feasibility or state verification rather than procedural (HOW) or causal (WHY) questions.

question type vs thread length

type	avg turns	threads
CAN	117.6	542
WHY	114.2	378
OTHER	113.6	1202
WHAT	101.8	579
HOW	99.3	404

key finding: CAN and WHY questions lead to LONGER threads (~18% longer than HOW questions).

CAN questions often involve exploration/experimentation cycles
WHY questions require causal investigation, often branching
HOW questions are more procedural, resolved faster

question position

position	count
opening question	522 (11.3%)
follow-up question	4,078 (88.7%)

most questions are follow-ups mid-thread, not conversation starters. questions emerge from context rather than initiating it.

question complexity (word count proxy)

complexity	count
1-5 words	468
6-15 words	2,081
16-30 words	1,210
30+ words	841

medium complexity (6-15 words) most common. very short questions are rare.

complexity vs resolution

simple questions (1-5 words)

status	count
RESOLVED	260 (70.7%)
COMMITTED	40
HANDOFF	40
UNKNOWN	20

medium questions (6-15 words)

status	count
RESOLVED	813 (76.1%)
HANDOFF	93
COMMITTED	85

complex questions (30+ words)

status	count
RESOLVED	899 (74.4%)
HANDOFF	87
EXPLORATORY	86
COMMITTED	63

finding: resolution rate is CONSISTENT (~70-76%) across complexity levels. complex questions aren’t harder to resolve — they just take more turns.

response patterns

pattern	count
answered immediately (by assistant)	4,535 (98.6%)
user continued asking	53
thread ended without answer	12

almost all questions get immediate assistant responses. only 12 questions (0.26%) left dangling.

question density vs outcomes

density	resolved	avg turns
high (>15%)	101	12.3
medium (5-15%)	373	46.0
low (<5%)	836	105.6
none	760	44.0

counterintuitive: low-density threads have HIGHEST resolution rate with longest average length. dense questioning doesn’t help resolution — focused work with occasional clarifying questions works better.

top threads by turn count (with questions)

thread	turns	questions	title
T-0ef9b016…	1,623	9	Minecraft resource pack CIT converter
T-048b5e03…	988	9	Debugging migration script
T-c66d846b…	615	1	S3 background ingest review
T-b428b715…	594	4	Implementation plan creation
T-c3eb8677…	506	5	Unify merge machinery

longest threads have FEW questions — they’re execution-heavy, not interrogative.

user question patterns

user	questions
concise_commander	2,669 (58%)
steady_navigator	1,207 (26%)
verbose_explorer	538 (12%)

concise_commander asks the most questions, consistent with deep technical investigation style.

sample questions by type

HOW (procedural):

“How do the best techniques from sneller and memchr combine here?”
“How does that support transactions?”
“How do we get the right image for the k8s job?”

WHY (causal):

“Why is min and max always float?”
“Why is this a url parameter?”
“Why is filtering in line instead of the plan better?”

CAN (capability):

“Can you make a pass at the functions and remove obvious perf issues?”
“Can you use the real table we tested with?“

insights summary

feasibility questions (CAN) create longer threads — exploration mode, not execution mode
questions are mostly follow-ups — context-dependent, not conversation starters
complexity doesn’t hurt resolution — just takes more turns
low question density = higher resolution — suggests interrogative style isn’t optimal for getting things done
98.6% of questions answered — assistant engagement extremely high
WHY questions are investigation triggers — correlate with debugging/understanding threads