user behavior profiles
comprehensive analysis of 6 users with >50 threads from amp corpus.
summary table
| user | threads | msgs | avg msg len | question ratio | peak hours | primary domain |
|---|---|---|---|---|---|---|
| @concise_commander | 1219 | 11461 | 263 chars | 37% | 9-16, 22-00 | data viz, performance |
| @steady_navigator | 1171 | 4452 | 547 chars | 43% | 04-11 | frontend, ai tooling |
| @verbose_explorer | 875 | 4511 | 932 chars | 26% | 16-21 | devtools, personal finance |
| @patient_pathfinder | 150 | 585 | 293 chars | 7% | 07-14 | infrastructure, metrics |
| @feature_lead | 146 | 243 | 780 chars | 11% | 13-17 | observability, analytics_service |
| @precision_pilot | 90 | 602 | 2037 chars | 34% | 19-22 | streaming, sessions |
@concise_commander
thread count: 1219 (highest volume user)
avg message length: 263 chars (shortest, most terse)
question ratio: 37%
avg turns/thread: 86.5 (longest conversations)
communication style
- extremely terse, command-like prompts
- uses imperative mood heavily: “fix”, “run”, “push”, “commit”
- technical references inline (sql snippets, error logs)
- rarely asks “how” — states desired outcome directly
- pattern: pastes error → expects immediate fix
characteristic phrases:
- “DO NOT change it. Debug it methodically”
- “fix the tests”, “run the tests”
- “push”, “commit”, “ok”
topic clusters
- canvas/charts (28 mentions): data visualization components
- sortmatches/radix (17 mentions): algorithm optimization
- query optimization: postgres, vacuum, performance tuning
- storage_optimizer/data_reorg: storage system internals
temporal patterns
- bimodal: active 09-16 (work hours) AND 22-00 (late night)
- light activity 03-06
workflow indicators
- high steering (0.85 avg for resolved threads)
- high approval rate (1.7 avg for resolved)
- 61% threads resolved, 13% committed
- max 615 turns in single thread — marathon debugging sessions
@steady_navigator
thread count: 1171
avg message length: 547 chars (moderate verbosity)
question ratio: 43% (most inquisitive)
avg turns/thread: 36.5
communication style
- polite, structured prompts (“please look at”, “can you”)
- frequently references file paths explicitly
- visual/spatial descriptions (“flip the timeAxis”, “ticks point down”)
- iterative refinement pattern — multiple follow-ups on visual precision
characteristic phrases:
- “please look at the
Component” - “almost there”, “something is still off”
- “please fix”, “see screenshot”
topic clusters
- query builder (19 mentions): apl query construction
- canvas/component (21 mentions): UI components
- traces/spans: observability UI
- ai integration: tooling, ai-assisted features
temporal patterns
- early bird: peak 04-11 (unusual 4-7am activity)
- minimal evening activity after 18:00
workflow indicators
- low steering (0.10 avg for resolved)
- moderate approval (0.41 avg)
- 65% resolved, 23% unknown status
- heavily screenshot-driven workflow
@verbose_explorer
thread count: 875
avg message length: 932 chars (verbose, context-rich)
question ratio: 26%
avg turns/thread: 39.1
communication style
- first-person heavy (“i want”, “my account”, “i am”)
- provides substantial context upfront
- references past threads frequently
- personal finance mixed with dev work
- occasional frustration markers (“there are WAY MORE threads”)
characteristic phrases:
- “search my amp threads”
- “i want to”, “can we”
- “instead”, “why does”
topic clusters
- amp/skills: tool customization, skill development
- icon/UI components: visual components
- git worktree/handoff: workflow tools
- personal finance: account analysis, money tracking
- nixos/config: system configuration
temporal patterns
- night owl: peak 18-21
- secondary peak 10-12
- gap 03-08
workflow indicators
- moderate steering (0.55 for resolved)
- high approvals (0.98)
- 83% resolved, 4% handoff (power user who spawns effectively)
- max 1623 turns — extensive exploratory threads
note: prior analysis incorrectly classified spawned subagent threads as handoffs. @verbose_explorer runs 231 spawned agents with 97.8% success rate.
@patient_pathfinder
thread count: 150
avg message length: 293 chars
question ratio: 7% (lowest, most directive)
avg turns/thread: 20.3
communication style
- infrastructure-focused, operational prompts
- polite (“please”, “can you”)
- specific technical references (prometheus, eks, eu regions)
- concise, task-focused
characteristic phrases:
- “please”, “can you”
- specific infra terms: “liveness probe”, “readiness probe”
- references databases, metrics systems
topic clusters
- platform/metrics (32 mentions): observability platform
- eu/eks (14 mentions): kubernetes, regional infrastructure
- prometheus/gateway: monitoring stack
- mcp/db: database operations
temporal patterns
- work hours only: 07-17
- peak 08-12
- minimal evening/night activity
workflow indicators
- low steering (0.22 for resolved)
- low approval (0.23)
- 53% resolved, 44% unknown
- clean operational patterns
@feature_lead
thread count: 146
avg message length: 780 chars
question ratio: 11%
avg turns/thread: 20.7
communication style
- feature-spec oriented prompts
- references otel/kubernetes fields directly
- detail-oriented: field names, dataset specs
- code review integration (coderabbit mentions)
characteristic phrases:
- detailed type references (kubernetes, otel fields)
- “analytics_service”, “search_modal”, “metrics”
- feature implementation context
topic clusters
- search_modal/analytics_service (54 mentions): primary feature area
- metrics/deletion service (59 mentions): data lifecycle
- kubernetes/otel (73 mentions): observability integration
- datasets/fields: data modeling
temporal patterns
- afternoon focus: peak 13-17
- minimal morning activity
- gap before 05:00
workflow indicators
- 45% handoff (highest delegation rate)
- low steering for handoffs
- 17% resolved, 9% committed
- external code review integration
@precision_pilot
thread count: 90 (smallest sample)
avg message length: 2037 chars (most verbose)
question ratio: 34%
avg turns/thread: 72.9
communication style
- architecture-focused, plan-oriented
- asks for generated plans to feed other threads
- cross-references extensively
- streaming/session state expertise
- thinks in terms of message flow and stitching
characteristic phrases:
- “generate a plan for me to feed into another thread”
- “update … with the new architecture”
- “shouldn’t we also update”, “why does”
topic clusters
- streams/durable (43 mentions): streaming architecture
- sessions/timeline (21 mentions): session state management
- migration/sse (22 mentions): data migration, server-sent events
- web_platform/frontend: application architecture
temporal patterns
- evening focus: peak 19-22
- bimodal: also active 00-04
- gap 05-14
workflow indicators
- high steering (0.43 for resolved)
- 82% resolved
- complex multi-thread orchestration patterns
cross-user patterns
message length vs question ratio
inverse correlation observed:
- terse users (@concise_commander: 263 chars) ask more questions (37%)
- verbose users (@precision_pilot: 2037 chars) provide more context upfront
temporal clustering
- morning crew: @steady_navigator (04-11), @patient_pathfinder (07-14)
- afternoon crew: @feature_lead (13-17)
- evening/night crew: @verbose_explorer (18-21), @precision_pilot (19-22), @concise_commander (22-00)
domain specialization
clear domain ownership:
- observability ui: @steady_navigator, @feature_lead
- data systems/perf: @concise_commander
- infrastructure: @patient_pathfinder
- streaming arch: @precision_pilot
- tooling/personal: @verbose_explorer
steering intensity
high steering correlates with:
- longer threads (@concise_commander)
- debugging sessions (@concise_commander, @verbose_explorer)
- frustrated status (all users with FRUSTRATED have steering >1.0)
generated by mary_glimmerflick | thread analysis pipeline