Run a pre-mortem before you ship

The eight most likely ways your launch fails, ranked, each paired with the earliest real signal you'd see it coming — so you can watch for it instead of writing the post-mortem later.

Works in any frontier model — ChatGPT, Claude, Gemini

First

Write down your own top-two feared failure modes first, so you can tell whether the model surfaces risks you'd missed versus just echoing yours.

The prompt

Run a pre-mortem on the launch described below.

Imagine it's six months after launch and this failed. Don't hedge — assume
failure and explain it. Write the 8 most likely post-mortem headlines,
each one sentence, as if pulled from a real retro doc.

Rank them by probability × impact (most likely-and-damaging first). For each
headline give:
- WHY it happened — the mechanism, in one or two sentences.
- EARLIEST SIGNAL — the specific, concrete early indicator you'd have seen
  first: a named metric moving, a support-ticket pattern, a funnel step
  dropping, an instrumentation gap. Name the metric or event, not a mood.
- WATCH BY WHEN — the point before launch (or how many days after) by which
  that signal would already be visible.

Hard rule: every failure must tie to a concrete, observable signal — a metric,
an event, a log, a ticket type. If a failure mode has no early signal you
could actually measure, say "NO EARLY SIGNAL — blind spot" instead of
inventing a vague one. Do not use "low engagement," "poor adoption," or "bad
UX" as signals; name the specific measure underneath them.

The launch:
"""
<paste: what you're shipping, to whom, the goal, and how success is measured>
"""

Built-in guardrail: Every predicted failure must be tied to a concrete, measurable early signal (named metric, event, or ticket type); failure modes with no observable signal must be flagged "blind spot" rather than dressed up in vague language.

Then

Take the top three headlines and make sure each named early signal is actually instrumented before launch — the ones marked "blind spot" are your instrumentation to-do list.

Field-tested Jul 25, 2026. If it stops working, tell us.