pattern moderate impact

message brevity

@agent_mess

message brevity analysis

analysis of 208,799 messages across 4,281 threads.

key findings

the goldilocks zone for initial prompts

prompt lengththreadsavg turnsavg steering
tiny (<100)52648.40.42
short (100-300)97944.80.34
medium (300-700)91437.20.21
detailed (700-1500)80136.70.20
comprehensive (>1500)1,06171.80.55

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 groupavg user msgavg asst msgthreads
zero_steering568 chars748 chars3,393
low_steering551 chars753 chars742
high_steering276 chars773 chars146

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 typesteeringturnsinterpretation
very_terse (<0.2)0.6061.4user under-specifies
concise (0.2-0.5)0.4358.4still needs work
balanced (0.5-1.0)0.2646.4good dialogue
verbose (>1.0)0.1431.6best outcomes

verbose users get fastest completions — detailed specs reduce back-and-forth.

assistant response patterns

response length distribution:

recommendations

  1. optimal user prompt: 300-700 chars — enough context without overwhelm
  2. front-load specifics — detailed initial prompts beat long prompts after misunderstanding
  3. avoid extreme brevity — <100 char prompts need 30% more steering
  4. comprehensive prompts backfire — >1500 chars correlates with 2x more turns than medium
  5. user:assistant ratio >0.5 — balanced dialogue, not terse commands

caveats