amp usage retrospective questions

structured questions for teams to discuss in retrospectives, organized by theme. each question is grounded in analysis of 4,656 threads.

thread health metrics

approval:steering ratio

• what is our team’s average approval:steering ratio? (target: >2:1)
• how many of our threads this sprint fell below 1:1 (doom spiral territory)?
• when we hit consecutive steering messages, what patterns emerge?

thread length

• are we abandoning threads too early? (<10 turns = 14% success rate)
• are threads dragging past 50 turns without resolution? what causes the elongation?
• what’s our “sweet spot” thread length distribution?

resolution rates

• what % of our threads hit RESOLVED vs HANDOFF vs UNKNOWN?
• are HANDOFFs intentional or abandonment in disguise?

prompt quality

context anchoring

• are we including @file references in opening prompts? (+25pp success rate when present)
• are openers between 300-1500 chars? (optimal steering rate 0.20-0.21)
• do we describe intent before action, or just give raw directives?

question density

• are we asking the agent clarifying questions >15% of messages? (excessive clarification signal)
• vs: are we asking enough? (<5% correlates with 76% resolution)

anti-patterns detection

agent behavior

• did we observe TEST_WEAKENING (agent “fixing” tests by removing assertions)?
• did the agent suggest “simpler approaches” when implementation got hard? (SIMPLIFICATION_ESCAPE)
• were workarounds applied instead of root-cause fixes? (71.6% workaround rate baseline)

user contribution to failures

• did we give polite requests (“please X”) that got ignored? (12.7% compliance rate)
• did we use prohibitions (“don’t”, “never”) that weren’t followed? (20% compliance rate)
• did we front-load context or drip-feed requirements?

process & tooling

task delegation

• are we using 2-6 task spawns? (77-79% resolution optimal range)
• are we over-delegating (>11 tasks → 58% resolution)?
• are spawn chains going past depth 10? (handoff risk increases)

verification gates

• do threads include verification steps before completion? (78.2% vs 61.3% success)
• are we running full test suites, or just the “changed” tests?

oracle usage

• are we using oracle EARLY (planning) or LATE (rescue)?
• 46% of FRUSTRATED threads show oracle usage—is ours proactive or reactive?

temporal patterns

time of day

• are complex debugging tasks scheduled during 6-9pm? (27.5% resolution—worst window)
• are we leveraging 2-5am or 6-9am windows for hard problems? (~60% resolution)

collaboration intensity

• are we rushing threads (>500 msgs/hr → 55% success)?
• can we adopt a more deliberate pace (<50 msgs/hr → 84% success)?

user archetypes & personal patterns

individual reflection

• what’s my personal approval:steering ratio?
• am i a “frontloader” (verbose openers) or “drip feeder” (context over time)?
• do i use socratic questioning style? (concise_commander pattern: 60.5% resolution)
• do my evening sessions perform worse than morning? (verbose_explorer pattern: 21% evening vs 59% morning)

team comparison

• whose prompting style consistently produces high-resolution threads?
• can we document and share those patterns?

early warning signals

doom spiral detection

• did we catch STEERING→STEERING transitions in real-time?
• how many turns before we recognized misalignment?
• did we pause and realign, or push through?

frustration detection

• did anyone hit level 4+ on the escalation ladder (profanity, caps)?
• what anti-pattern preceded the frustration? (usually: shortcuts, test weakening)

actionable improvements

next sprint experiments

• which anti-pattern will we explicitly watch for?
• what prompt template will we try standardizing?
• which time windows will we protect for deep work?

metrics to track

• can we instrument approval:steering ratio per thread?
• can we flag threads approaching >50 turns for review?
• can we surface “verification gate missing” warnings?

meta questions

• are these retrospective questions themselves improving our amp usage?
• what new patterns emerged this sprint that aren’t in the catalog?
• should we update the anti-patterns list based on recent experiences?

derived from analysis of 4,656 amp threads across multiple users and projects