message brevity analysis
analysis of 208,799 messages across 4,281 threads.
key findings
the goldilocks zone for initial prompts
| prompt length | threads | avg turns | avg steering |
|---|---|---|---|
| tiny (<100) | 526 | 48.4 | 0.42 |
| short (100-300) | 979 | 44.8 | 0.34 |
| medium (300-700) | 914 | 37.2 | 0.21 |
| detailed (700-1500) | 801 | 36.7 | 0.20 |
| comprehensive (>1500) | 1,061 | 71.8 | 0.55 |
sweet spot: 300-1500 chars — lowest steering corrections, fewest turns needed.
very long prompts (>1500) paradoxically cause MORE steering and MORE turns. hypothesis: overwhelming context leads to misinterpretation or scope creep.
user message length correlates with success
| steering group | avg user msg | avg asst msg | threads |
|---|---|---|---|
| zero_steering | 568 chars | 748 chars | 3,393 |
| low_steering | 551 chars | 753 chars | 742 |
| high_steering | 276 chars | 773 chars | 146 |
users in high-steering threads write ~50% shorter messages (276 vs 568 chars). shorter messages = more ambiguity = more corrections needed.
user:assistant length ratios
| ratio type | steering | turns | interpretation |
|---|---|---|---|
| very_terse (<0.2) | 0.60 | 61.4 | user under-specifies |
| concise (0.2-0.5) | 0.43 | 58.4 | still needs work |
| balanced (0.5-1.0) | 0.26 | 46.4 | good dialogue |
| verbose (>1.0) | 0.14 | 31.6 | best outcomes |
verbose users get fastest completions — detailed specs reduce back-and-forth.
assistant response patterns
- avg assistant: 753 chars
- avg user: 519 chars
- ratio: 1.45:1 (assistant writes ~45% more)
response length distribution:
- brief (<500): 117,188 (63%)
- medium (500-2k): 54,634 (29%)
- long (2k-5k): 10,422 (6%)
- very long (>5k): 3,293 (2%)
recommendations
- optimal user prompt: 300-700 chars — enough context without overwhelm
- front-load specifics — detailed initial prompts beat long prompts after misunderstanding
- avoid extreme brevity — <100 char prompts need 30% more steering
- comprehensive prompts backfire — >1500 chars correlates with 2x more turns than medium
- user:assistant ratio >0.5 — balanced dialogue, not terse commands
caveats
- steering_count may undercount corrections (only explicit labels)
- completion_status appears unfilled (all 0.0%) — relying on steering as proxy
- no approval_score data populated for response quality assessment