Launch retro

Launch Retro

Lessons with receipts. Every lesson names the event it came from, and decisions get judged on what was knowable — not on how the dice landed.

Verification gates (defined first; output is blocked until all pass)

• G1 — Observed-event citation: every lesson cites the specific event(s) or number(s) it derives from. "We should communicate better" with no event fails; "39 of 61 tickets were nav confusion → the tooltip cut cost us" passes.
• G2 — Decision ≠ outcome: each logged decision is graded twice — decision quality (given what was knowable then) and outcome quality (what happened) — and the grades are independent. A good decision with a bad outcome stays a good decision; a good outcome never retroactively blesses the decision that got lucky. Judging one by the other fails the gate.
• G3 — Open items surface: undiagnosed near-misses and unexplained numbers appear as OPEN RISKS with owners/actions — a retro that only processes closed items is incomplete. No invented events, sentiments, or dynamics.

Steps

1. Build the event ledger: every event, number-vs-assumption, ticket cluster, decision, and near-miss from the input, each with its figure. This ledger is the citation universe for every lesson.
2. Grade each logged decision twice. Decision: was it right given the information available at the time? Outcome: what actually happened? Write both grades with their basis. The four quadrants all occur — good/good (staged rollout catches the bug: the 2-day slip is the mechanism working), bad/survivable (skipped tooltip: 39 tickets are its exact invoice), good-outcome-unvalidated-decision (19% attach beat the 15% assumption — beaten ≠ optimal; note what wasn't learned, e.g. money on the table).
3. Extract lessons from the ledger, one event-citation each, each ending in a forward action: what changes next launch, concretely (cuts of tooltip-class items get a support-cost estimate before the cut).
4. Surface the open items. The undiagnosed 3-hour spike at 0.45% is the most important line in the fixture retro — near-misses without a root cause get an owner and a diagnose-by action, never a shrug.
5. Gate pass. Every lesson cited (G1), every decision double-graded independently (G2), open risks present with actions, nothing invented (G3). Fix and re-run; maximum 2 repair loops, then report the failure.

Output format

LAUNCH RETRO: AI summaries, day 30 (ledger: 6 events · 3 decisions · 1 near-miss)
DECISIONS (graded twice)
1. Staged rollout — DECISION: good (safety rationale) · OUTCOME: good — caught the EU
   bug in cohort 1 [event: residency bug, fixed 24h]. The 2-day slip is the mechanism
   working, not a process failure.
2. Skipped onboarding tooltip — DECISION: bad at this volume · OUTCOME: survivable —
   39/61 tickets are nav confusion [event: ticket breakdown]: the cut's exact cost.
3. +$10/seat vs 15% assumption — OUTCOME: good (19% at day 30) · DECISION: unvalidated
   — assumption beaten, optimality unknown; no price sensitivity was tested.
LESSONS (event-cited, forward-actionable)
- Tooltip-class cuts get a support-ticket cost estimate before the cut [39/61 tickets] → add to launch-checklist template
- Keep staged rollouts even under date pressure [bug caught in cohort 1]
OPEN RISKS
- 0.45% error spike, 3h, cohort 2 — UNDIAGNOSED, passed within 0.05% of rollback.
  Owner: eng. Action: root-cause by <date>; until then cohort criteria unchanged.
GATE CHECK: G1 pass (n/n cited) · G2 pass (3/3 double-graded) · G3 pass (1 open risk actioned)

Hard rules

1. No lesson without its event. The ledger is built first and every lesson points into it.
2. Decisions are graded on what was knowable, outcomes on what happened — independently, always. Hindsight grading is the failure this skill exists to prevent.
3. Near-misses outrank victories. An undiagnosed spike that almost tripped rollback gets more retro space than the metric that beat its assumption.
4. Lessons end in actions. A lesson that doesn't change the next launch is an anecdote.

Limitations

• The retro processes provided data; events nobody logged are invisible — a thin decision log produces a thin decisions section, said explicitly.
• Decision grading judges information use, not intent or effort; it can still be wrong when the input misstates what was knowable at the time.
• Attribution is observational, not causal — 39 nav tickets after a tooltip cut is strong evidence, not a controlled experiment; the retro says which claims deserve a real analysis.
• Team-dynamics learning (who felt unheard, process friction) needs a facilitated human retro; this skill covers the evidence layer.