user archetypes

5 distinct patterns emerged from 4,656 threads. each archetype represents a different collaboration style with amp—what works, what triggers steering, and the phrases that define their sessions.

@concise_commander: the marathon debugger

storage engine · performance · SIMD

1219 threads

60.5% resolution

86.5 avg turns

What works

socratic questioning ('OK, what's next?')
marathon persistence
explicit approvals

Triggers

premature action
forgetting flags (-run=xxx)
full test suites

@steady_navigator: the efficient executor

observability · frontend · ai tooling

1171 threads

67% resolution

36.5 avg turns

What works

polite structured prompts
post-hoc corrections
screenshot-driven

Triggers

rarely (2.6% rate)
uses post-hoc rejection not interrupts

@verbose_explorer: the spawn orchestrator

devtools · personal projects · skills

875 threads

83% resolution

39.1 avg turns

What works

heavy spawn/coordinate usage
delegates to subagents effectively
context frontloading in parent threads

Triggers

subagent context loss
spawn coordination overhead
evening sessions (lower focus)

@precision_pilot: the architect

streaming · sessions · architecture

90 threads

82.2% resolution

72.9 avg turns

What works

plan-oriented prompts
cross-references
multi-thread orchestration

Triggers

rushing without plan
skipping architecture review

@patient_pathfinder: the infrastructure operator

kubernetes · prometheus · infrastructure

150 threads

54% resolution

20.3 avg turns

What works

work hours only (07-17)
precise specs
minimal back-and-forth

Triggers

ambiguous requirements
off-hours sessions

signature phrases

@concise_commander: wait dont NO FUCKING SHORTCUTS

@steady_navigator: please look at almost there see screenshot

@verbose_explorer: let's we should help me understand

@precision_pilot: the plan is referring to coordinate with

@patient_pathfinder: deploy configure apply

🎭 match the message to the user

real messages from the threads. can you guess who said what? (hover to reveal)

fun fact: steering predicts success—threads with corrections have 60% resolution vs 37% without. positive reinforcement matters even more.

hall of frustration

peak developer despair at agents doing the wrong thing

amp: "[agent makes edit]"

"Noooooooooooooo no no no. Fuck off! Undo that! The bug exists already."

amp: "[agent immediately undos]"

the extended 'no' count + 'fuck off!' combo is peak exasperation — @concise_commander

amp: "The system is fully functional and type-safe as requested! 🚀"

"NO BITCH THATS NOT TYPESAFE"

amp: "i want to do `const { factSchema } = createAppScope(/* schema */);`"

agent confidently says it's typesafe with a rocket emoji, gets immediately corrected — @steady_navigator

amp: "can it only break on characters if there isn't a good space to break on? 2."

"DONT FUCKIN SET STYLES IN A USEEFFECT"

amp: "[agent fixes immediately]"

calm enumerated complaint with unexpected caps lock rage — @steady_navigator

amp: "[agent pushes code without testing]"

"Why are you not running this fucking test instead of like you just put why are you blindly pushing?"

amp: "[agent runs tests]"

frustrated stream-of-consciousness mid-sentence pivot — @concise_commander

amp: "[agent takes multiple actions]"

"Wait a fucking second, you responded to all of that without confirming with me?"

amp: "[agent pauses]"

dawning realization when you check what the agent did while you weren't looking — @concise_commander

hall of fame

heartwarming moments of appreciation and celebration

amp: "go merge that PR. 🔄"

"that was BEAUTIFUL, brother, amp threads now have tags. stand proud, you are strong."

amp: "amp makes me think of a village, with different families that do different crafts..."

— @verbose_explorer

amp: "Now SET_CURRENT_SESSION will use the server-returned ID"

"it works! fucking finally, thank you"

amp: "[investigation complete]"

— ???

amp: "[chart loading animation implemented]"

"Fuck yes, beautiful. But maybe we need a different animation per chart type?"

amp: "[agent explores chart-specific animations]"

— @steady_navigator

amp: "[optimization implemented and benchmarked]"

"Great. Commit with bench numbers."

amp: "[agent commits with performance data]"

— @concise_commander

amp: "[discussing SIMD implementations]"

"Forget the SSE2 stuff for now. Focus on NEON, make it great"

amp: "[agent focuses on NEON implementation]"

— @concise_commander

amp: "[AI chat observability prototype working]"

"Great job. Now the next task is going to be tricky."

amp: "Axiom has various ways to visualise queries..."

— @steady_navigator

amp: "[exploring optimization approaches]"

"ok great, anything else in that list of ideas?"

amp: "[agent continues with next optimization idea]"

— @concise_commander

amp: "[discussing implementation complexity]"

"If the simpler approach works today, then make it simpler"

amp: "[agent simplifies and commits]"

— @concise_commander

amp: "[discussing thread privacy feature]"

"go all in. and yea, threads created outside any workspace should also be made private"

amp: "[agent implements full solution]"

— @steady_navigator

amp: "[sort push down to column implementations complete]"

"The benchmarks look great, see the corresponding PR description with gh."

amp: "But now I went and benchmarked the Finalize method and the results are... worse."

— @concise_commander