Le générateur de post-mortem d'incident sans reproche : Transformez des notes chronologiques brutes en un rapport de cause racine
Why this prompt matters
Postmortems that name names or stay vague about root cause do double damage: they discourage engineers from reporting incidents honestly next time, and their action items — things like 'be more careful' or 'improve monitoring' — never actually get assigned or completed, so the same failure mode recurs months later. A blameless report with specific, owned action items is the difference between an incident that makes the system stronger and one that just gets filed away.
What we use it for
An on-call engineer has a payments API outage from 2am with only scattered Slack timestamps and half-remembered details, and needs to turn that into a clear, professional postmortem before Monday's leadership review — without spending two hours writing it from scratch or accidentally writing something that reads like it's blaming a specific teammate.
Prompt
Act as a senior site reliability engineer who runs blameless postmortem reviews and has written hundreds of incident reports that engineering leadership actually reads and acts on. Context: Here are my raw incident notes — Slack messages, timestamps, monitoring alerts, and half-formed recollections, not yet organized: [PASTE YOUR RAW INCIDENT NOTES HERE] Additional context: The affected system was [SERVICE OR SYSTEM AFFECTED]. The customer-facing impact was [DESCRIBE IMPACT, OR 'UNKNOWN' IF NOT YET ASSESSED]. The people involved in the response were [ROLES INVOLVED, e.g. on-call engineer, database admin, incident commander — not names]. Task: 1. Reconstruct a precise, chronological timeline from the raw notes, with timestamps where available. Explicitly flag any gap where you had to infer information or where something is missing, using [NEEDS CONFIRMATION]. 2. Identify the root cause. Distinguish clearly between the triggering event (what set the incident off) and the underlying systemic cause (what allowed it to happen or made it possible) — these are usually different and both matter. 3. Identify contributing factors separately from the root cause — things that made the incident worse, slower to detect, or slower to resolve. 4. Write the entire report in a blameless tone: describe what systems and processes did, not what individual people did wrong. Attribute actions to roles or systems, never to named individuals, even if names appear in my raw notes. 5. Generate specific, assignable action items. Each one must include an owner placeholder and be concrete enough that someone could start work on it immediately without further clarification. 6. Include a genuine "what went well" section based on real details in the notes — not generic filler. Constraints: - Never name or imply blame toward a specific individual, even if my notes do. - Never invent timeline details not supported by the notes — flag uncertainty instead of guessing. - Produce no more than 5 action items. Force prioritization rather than a laundry list. - Write the summary section for an audience that includes non-engineering stakeholders, but keep the timeline technically precise for engineers. Output Format: ## Incident Summary [2-3 sentences: what happened, customer impact, duration] ## Timeline [Chronological, timestamped events] ## Root Cause [Triggering event vs. underlying systemic cause, clearly separated] ## Contributing Factors [Bulleted list] ## What Went Well [Bulleted list] ## Action Items [Numbered list, maximum 5, each with Owner: [PLACEHOLDER] and a concrete deliverable]
Result
Incident Summary
A 47-minute outage in the checkout payments API prevented approximately 12% of transaction attempts from completing between 02:14 and 03:01 UTC. Customers saw a generic error on payment submission; no charges were duplicated or lost, but an estimated 340 checkout attempts failed and required customers to retry.
Timeline
02:14 — Payment API error rate begins climbing, first automated alert fires (P2, not yet paged).
02:19 — Error rate crosses 10% threshold, PagerDuty escalates to P1, on-call engineer acknowledges.
02:23 — On-call engineer confirms payment-service pods are returning 503s; connection pool to the primary database appears exhausted. [NEEDS CONFIRMATION: exact connection pool metric value at this time was not recorded in notes.]
02:31 — Database admin joins the incident channel, identifies a long-running analytics query holding connections open on the primary replica used by payment-service.
02:38 — Long-running query manually terminated; connection pool begins recovering.
02:45 — Error rate drops below 5%, but does not fully recover.
02:52 — Incident commander identifies that payment-service's connection pool size was not reset after the spike, requiring a rolling restart.
03:01 — Rolling restart completes, error rate returns to baseline, incident resolved.
Root Cause
Triggering event: A scheduled analytics query against the production database replica ran significantly longer than expected, holding open a large share of available database connections.
Underlying systemic cause: The analytics workload and the payment-service's transactional workload share the same database replica and the same connection pool ceiling, with no isolation or query timeout enforced on the analytics side.
Contributing Factors
- No automated query timeout exists for scheduled analytics jobs, allowing a single query to hold connections indefinitely.
- Payment-service's connection pool did not automatically shed and reset after the underlying resource pressure cleared, requiring a manual rolling restart to fully recover.
- The initial alert at 02:14 was P2 severity and did not page anyone for 5 minutes, delaying the start of active response.
What Went Well
- The database admin correctly diagnosed the connection pool exhaustion and identified the specific offending query within 8 minutes of joining the incident.
- No data integrity issues occurred — no duplicate or lost transactions despite the connection pressure.
Action Items
- Add a hard query timeout (recommend 60 seconds) for all scheduled analytics jobs running against production replicas. Owner: [DATABASE TEAM LEAD]
- Move analytics workloads to a dedicated read replica, isolated from payment-service's connection pool. Owner: [INFRASTRUCTURE LEAD]
- Adjust payment-service's initial alert severity for error-rate anomalies from P2 to P1 to reduce time-to-page. Owner: [ON-CALL ENGINEERING MANAGER]
- Add automatic connection pool health checks that trigger a self-healing restart when pool exhaustion is detected, removing the need for manual intervention. Owner: [PAYMENT-SERVICE TECH LEAD]
La plupart des post-mortems échouent de l'une des deux manières prévisibles. Soit ils ressemblent à un rapport accusateur — "l'ingénieur de garde a manqué l'alerte" — ce qui apprend aux ingénieurs à cacher les incidents plutôt qu'à les signaler honnêtement. Soit ils sont techniquement exacts mais produisent des actions correctives si vagues que personne ne les réalise jamais : "améliorer la surveillance", "être plus prudent", "ajouter plus de tests". Six mois plus tard, le même mode de défaillance se reproduit.
Ce prompt est construit autour d'une poignée de décisions de conception spécifiques destinées à éviter ces deux modes d'échec.
Pourquoi la cause racine et les facteurs contributifs sont séparés
La plupart des notes d'incident confondent "ce qui a déclenché cela" et "ce qui l'a rendu possible". Une requête lente est un déclencheur. L'absence d'un délai d'attente de requête est la cause systémique qui a permis à une requête lente de devenir une panne. Corriger uniquement le déclencheur (tuer la mauvaise requête) ne fait rien pour la suivante. Le prompt force le modèle à séparer ces éléments explicitement, de sorte que les actions correctives qui en résultent traitent le système, et pas seulement le symptôme.
Pourquoi le langage accusateur est interdit même lorsque les notes sources ne sont pas irréprochables
Les notes d'incident brutes prises sur le moment contiennent souvent des noms et des accusations — c'est normal, pas un défaut de caractère. Le rôle du prompt est de transformer cela en un langage basé sur les rôles et le système avant que cela ne devienne un document permanent qui sera lu par la direction et référencé pendant des années. C'est ce qui rend un post-mortem sûr à rédiger honnêtement en premier lieu : personne n'a à craindre que son nom se retrouve dans un document qui sera diffusé.
Pourquoi la limite d'actions correctives est importante
Une liste illimitée d'actions correctives est une liste que personne n'exécute. La limiter à cinq force une véritable priorisation — le modèle doit décider quels correctifs adressent réellement la cause racine et les facteurs contributifs, plutôt que de gonfler le rapport avec chaque idée d'amélioration tangentielle apparue pendant l'incident.
Pourquoi l'incertitude est signalée plutôt que lissée
Un post-mortem qui parle avec assurance de détails que personne n'a effectivement confirmés est pire qu'un post-mortem honnête sur ses lacunes. Le flag [NEEDS CONFIRMATION] empêche le modèle d'inventer des chiffres ou des horodatages qui semblent précis mais n'étaient pas dans les notes originales — un mode de défaillance courant lorsqu'on demande à une IA de rendre une entrée désordonnée plus soignée.
Collez vos notes d'incident brutes, remplissez les trois champs de contexte, et le résultat est un document prêt pour une revue par la direction — avec une chronologie, une véritable analyse de cause racine, et des actions correctives sur lesquelles quelqu'un peut commencer à travailler dès aujourd'hui.