post-mortems-retrospectives（事后回顾）

Post-mortems & Retrospectives

Scope

Covers

• Running blameless incident post-mortems and project/OKR retrospectives
• Turning "what happened?" into system learnings + decisions (not blame)
• Creating follow-through: owners, due dates, success signals, and review cadence
• Adding kill criteria / triggers so future pre-mortems lead to real action
• Institutionalizing learning via a lightweight "Impact & Learnings" review

When to use

• "Run a postmortem / retrospective for <incident/project> and write the doc."
• "We missed OKRs—lead a retro focused on learning and systemic blockers."
• "Create an after-action review with action items and owners."
• "Set up a weekly impact & learnings review so insights don't die in docs."
• "Do a pre-mortem and define kill criteria / pivot triggers."

When NOT to use

• The incident is still active (do incident response first; schedule the review after stabilization)
• The goal is to assign blame or evaluate an individual's performance (use HR/management processes)
• You need deep technical debugging without the right experts (this skill facilitates; it doesn't replace engineering investigation)
• You need to decide *what problem to solve* (use a problem-definition / discovery process first)
• You need to facilitate a meeting that is not a post-mortem or retrospective (use running-effective-meetings)
• You need to improve the shipping process itself, not review a past launch (use shipping-products)
• You need to change engineering culture or practices based on systemic patterns across retros (use engineering-culture)
• You need to plan for future risks and uncertainties rather than review past events (use planning-under-uncertainty)

Inputs

Minimum required

• What are we reviewing? (incident / project / OKR period) + 1–2 sentence summary
• Time window and key dates (start/end; detection time; resolution time if incident)
• Desired outcome (learning, prevention, speed, quality, alignment)
• Participants/roles (facilitator, scribe, decision owner; key stakeholders)
• Evidence available (timeline notes, metrics, dashboards, tickets, docs)
• Constraints (privacy; what to anonymize; audience)

Missing-info strategy

• Ask up to 5 questions from references/INTAKE.md (3–5 at a time).
• If details are unavailable, proceed with explicit assumptions and label unknowns.
• Do not request secrets or personal data; use anonymized descriptions.

Outputs (deliverables)

Produce a Post-mortems & Retrospectives Pack in Markdown (in-chat; or as files if requested):

1. Retro brief + agenda (purpose, attendees, roles, pre-reads, ground rules)
2. Facts + timeline (what happened; impact; timestamps; links)
3. Contributing factors + root cause hypotheses (systems lens; "why it made sense")
4. Learnings + decisions (what changes; why; tradeoffs)
5. Action tracker (owner, due date, success signal, follow-up date)
6. Kill criteria / triggers (signals → committed action) for future work
7. Learning dissemination plan (how to socialize + a recurring "Impact & Learnings" review)
8. Risks / Open questions / Next steps (always)

Templates: references/TEMPLATES.md Expanded guidance: references/WORKFLOW.md

Workflow (7 steps)

1) Classify the review + set blameless ground rules

• Inputs: request context; references/INTAKE.md.
• Actions: Identify the review type (incident / project / OKR). Set a blameless norm ("fix systems, not people") and decide whether to reframe language as "retrospective" to signal learning. Confirm facilitator, scribe, and decision owner.
• Outputs: Retro brief (draft) + attendee list + meeting invite outline.
• Checks: Objective is explicit (learning + improvement). Roles are assigned.

2) Assemble facts and a shared timeline (separate facts from stories)

• Inputs: artifacts (tickets, dashboards, logs, notes).
• Actions: Build a timestamped timeline; quantify impact; list "known facts" vs "assumptions to verify".
• Outputs: Facts + timeline section using references/TEMPLATES.md.
• Checks: Timeline has timestamps and links/evidence where possible. Assumptions are labeled.

3) Diagnose contributing factors (systems lens)

• Inputs: timeline + impact.
• Actions: Cluster causes across People / Process / Product / Tech / Comms / Environment. Use a "make it reasonable" lens: what conditions made the outcome likely? Optionally run 5 Whys on the top 1–2 factors.
• Outputs: Contributing factors map + root cause hypotheses.
• Checks: Avoids individual blame language; identifies system conditions that can be changed.

4) Extract learnings and decide what to change

• Inputs: contributing factors.
• Actions: Write 3–7 crisp learnings ("we learned that…"). Convert learnings into decisions (fix, guardrail, instrumentation, runbook, training, scope change). Keep OKR/grade discussion secondary to "why" and "what changes next".
• Outputs: Learnings + decisions section.
• Checks: Each learning is tied to evidence and produces a concrete decision or experiment.

5) Build the action tracker (owners + dates + success signals)

• Inputs: decisions.
• Actions: Create action items with an owner, due date, and success signal. Add a follow-up review date (or a recurring review). Limit to what can realistically be executed; explicitly park "later ideas".
• Outputs: Action tracker table + follow-up plan.
• Checks: No orphan actions: every item has owner + date. Top actions address top factors.

6) Add kill criteria / triggers (pre-commit to future action)

• Inputs: learnings; "what would we do differently next time?"
• Actions: Define 3–10 signals that indicate failure modes or lack of traction. For each signal, pre-commit to an action (pause, pivot, kill, escalate, add investment).
• Outputs: Kill criteria / trigger list.
• Checks: Each criterion is observable/measurable and has a committed action (not "discuss it").

7) Disseminate learning + quality gate + finalize

• Inputs: full draft pack.
• Actions: Create a 1-page shareout (TL;DR, top actions, decisions). Propose a lightweight weekly/biweekly "Impact & Learnings" review to socialize learnings beyond the team. Run references/CHECKLISTS.md and score with references/RUBRIC.md. Add Risks / Open questions / Next steps.
• Outputs: Final Post-mortems & Retrospectives Pack.
• Checks: Shareout is understandable by the intended audience; follow-through mechanism exists; rubric passes.

Quality gate (required)

• Use references/CHECKLISTS.md and references/RUBRIC.md.
• Always include: Risks, Open questions, Next steps.

Examples

Example 1 (incident postmortem): "We had a 45-minute outage in our payments API yesterday. Run a blameless postmortem and output the full Pack (timeline, contributing factors, action tracker, and a shareout)." Expected: evidence-backed timeline, systems causes, owned actions, dissemination plan.

Example 2 (OKR retro): "We hit 0.8 on our Q4 activation OKR. Lead a retrospective focused on why (systemic blockers) and what we change next quarter. Output the full Pack and kill criteria for the next initiative." Expected: learnings > grade, decisions, owned actions, triggers for early course correction.

Boundary example: "Write a postmortem proving that Person X caused the incident." Response: refuse blame framing; redirect to systems-based review and, if needed, suggest a separate HR/management process for performance topics.

Boundary example (neighbor redirect): "Our last three retros all surfaced the same 'deploy process is broken' theme. Fix the deploy process." Response: recurring themes across retros indicate a systemic engineering culture or process issue. Use engineering-culture to design the process improvement. This skill is for running the review itself, not implementing the changes it surfaces.

Anti-patterns

1. Blame in disguise — Using blameless language ("the system failed") while structuring the timeline and contributing factors to point at a single person. Contributing factors must focus on system conditions, not individual actions.
2. Action items without owners — Producing a list of "we should" recommendations with no owner, due date, or success signal. Every action must be owned, dated, and measurable.
3. Shallow root cause — Stopping at the first "why" (e.g., "the deploy script failed") instead of investigating systemic conditions (e.g., "no integration test coverage for deploy scripts, no runbook, no alerting"). Use at least 3 levels of "why" on top contributing factors.
4. Retro amnesia — Running retrospectives that produce insights but never feeding them into durable process changes. Every retro must include a dissemination plan and a follow-up review date.
5. Grade over learning — Spending most of the retrospective debating the OKR score or incident severity instead of investigating systemic causes and deciding what changes. Keep grading secondary to "why" and "what changes next."