opening words analysis
analysis of first 3 words from 4,281 user thread openers. correlates opening patterns with thread outcomes (message count, tool usage).
key findings
dominant patterns
| rank | first word | count | % of threads |
|---|---|---|---|
| 1 | ”continuing” | 1,502 | 35.1% |
| 2 | ”please” | 667 | 15.6% |
| 3 | ”in” | 153 | 3.6% |
| 4 | ”i” | 134 | 3.1% |
| 5 | ”fix” | 109 | 2.5% |
50.7% of all threads start with just two words: “continuing” or “please”. this is a MASSIVE concentration.
opener → outcome correlation
most interesting signal: opening word correlates strongly with thread complexity.
| opener | count | avg messages | avg tool uses | interpretation |
|---|---|---|---|---|
| ”we’re” | 24 | 249.1 | 119.0 | collaborative, long sessions |
| ”let’s” | 45 | 212.5 | 104.7 | exploration mode, extended work |
| ”summarize” | 41 | 155.0 | 71.8 | analysis tasks, multi-file |
| ”implement” | 35 | 142.0 | 68.7 | creation tasks, substantial |
| ”we” | 63 | 114.0 | 55.9 | collaborative framing |
| ”review” | 80 | 101.4 | 49.5 | code review, iteration |
| ”continuing” | 1,502 | 97.5 | 49.4 | resumed work baseline |
| ”what” | 35 | 40.0 | 16.5 | questions, quick answers |
| ”using” | 34 | 17.1 | 11.9 | short directive tasks |
| ”migrate” | 33 | 20.5 | 12.3 | scripted batch operations |
interpretation
collaborative framing (“we”, “we’re”, “let’s”) → longest threads
- avg 114-249 messages vs 97.5 for “continuing”
- hypothesis: collaborative language signals open-ended exploration, user stays engaged longer
imperative openers (“fix”, “create”, “update”, “migrate”) → shorter threads
- avg 20-62 messages
- clear directive = faster completion, less back-and-forth
question openers (“what”, “how”) → minimal tool use
- avg 16-38 tool uses vs 49+ for task openers
- often answered from existing knowledge, less exploration needed
opener taxonomy
1. continuation pattern (35%)
"continuing work from" - 1,236 occurrences
"continuing from https://..." - 266 occurrencesvast majority of work is resumed from prior threads. this is the DOMINANT usage pattern.
2. polite directive (16%)
"please look at" - 435
"please run" - 33
"please read" - 25
"please start" - 15
"please implement" - 15structured as requests. “please look at” is the canonical opening for new work.
3. direct command (8%)
"fix the" - 39
"review the" - 37
"look at" - 63
"run and" - 20
"read the" - 17imperative form without pleasantries. correlates with shorter threads.
4. first person (3%)
"i have" - 26
"i got" - 27
"i need" - 18
"i want" - 18user establishes context/need. mid-length threads.
5. collaborative (“we”) (2%)
"we need" - 18
"we are" - 13
"we're going" - 12frames agent as partner. LONGEST average threads.
6. interrogative (<1%)
"can you" - 77
"what is" - 17question-based. relatively short threads.
user type signals
the “you are” pattern (52 occurrences, avg 86.5 messages) is interesting:
- “You are fixing Icon migration issues…”
- appears to be spawn/delegation pattern where user programs agent identity
recommendations
- for tooling: detect continuation patterns to auto-load prior context
- for UX: “please look at” is the natural human opener for new work - design around it
- for metrics: collaborative openers (“we”, “let’s”) predict 2x longer engagement
- for agent behavior: imperative openers (“fix”, “migrate”) should bias toward efficient completion, not exploration
limitations
- only analyzed first user message per thread
- “continuing” threads inherit context from prior work, inflating their metrics
- no sentiment analysis on opener tone
- no success/failure correlation (would need manual labeling)