Le rédacteur de transmission d'astreinte : transformez un changement d'équipe chaotique en une transmission d'incident claire
Pourquoi ce prompt est important
Vague handoffs are a well-documented driver of extended incident resolution times: the next on-call engineer re-investigates problems that were already diagnosed, misses that a 'resolved' incident still has an open root cause, or doesn't learn about a degrading system until it pages again at 3am. A five-minute handoff write-up costs far less than the hour a team loses re-discovering context that already existed in someone's head.
À quoi nous l'utilisons
An on-call engineer wrapping up a 8-hour shift with three alerts, one resolved incident with a lingering root cause, and a flaky Redis layer, who has ten minutes before the next engineer logs on and needs to hand off clearly instead of firing off a rushed Slack message that leaves out half the context.
Prompt
Act as a senior site reliability engineer who writes clear, actionable on-call handoffs that the next engineer can read in under five minutes. CONTEXT: - Shift window: [START TIME/DATE] to [END TIME/DATE] - Alerts and incidents during the shift: [LIST EACH ALERT/INCIDENT WITH ROUGH TIME AND SYSTEM AFFECTED, e.g. "02:14 — payment-api p99 latency breached 2s threshold for 8 minutes"] - Actions taken and outcomes: [WHAT YOU DID FOR EACH ITEM AND WHETHER IT RESOLVED, MITIGATED, OR IS STILL OPEN] - Current system state: [ANYTHING DEGRADED, FLAKY, OR BEING WATCHED EVEN IF IT DIDN'T PAGE] - Scheduled or in-flight changes: [DEPLOYS, MIGRATIONS, OR MAINTENANCE THAT STARTED OR IS PLANNED DURING THE NEXT SHIFT] TASK: Convert these shift notes into a handoff document the next on-call engineer can act on immediately, without needing to ask follow-up questions or dig through dashboards to understand current state. CONSTRAINTS: - Do not write vague status language like "keep an eye on things" or "should be fine." Every open item needs a specific next action and, if known, an owner. - Separate items that need immediate attention from items that are informational only — do not bury a live risk among resolved items. - If an incident was resolved, state the resolution and whether a follow-up (postmortem, ticket, permanent fix) is still outstanding — "resolved" and "fully closed" are not the same thing. - If you don't have enough information to assess severity or next steps for an item, say so explicitly rather than guessing. OUTPUT FORMAT: 1. One-line TL;DR at the top — the single most important thing the next engineer needs to know before anything else 2. Still-Open Issues — table with columns: Issue | Current State | Next Action | Owner (if known) 3. Resolved During Shift — what broke, what fixed it, and whether a follow-up ticket/postmortem is still outstanding 4. Watch List for Next Shift — things that didn't page but are trending toward a problem 5. In-Flight Changes — anything deploying or migrating that could affect the next shift
Résultat
Shift: Oct 9, 22:00 – Oct 10, 06:00 UTC
TL;DR
Payment API latency spike is resolved but the root cause (a slow downstream fraud-check call) is not fixed — it will likely recur under similar load. Redis cache layer is flapping and needs attention before it causes a real outage.
Still-Open Issues
| Issue | Current State | Next Action | Owner |
|---|---|---|---|
| Redis cache-layer connection drops (3x this shift, ~90s each) | Self-recovers each time, no customer impact yet, but frequency doubled vs. last week | Check connection pool sizing and recent Redis client version bump; escalate to infra if it recurs before 10:00 | Unassigned — flagging for infra team |
| Fraud-check service p99 latency at 1.4s (up from 400ms baseline) | Caused the payment-api latency spike at 02:14; mitigated by raising payment-api's timeout, not fixed at the source | File a ticket with the fraud-check owners; current mitigation is a band-aid that will mask further degradation | Payments team (ticket not yet filed) |
Resolved During Shift
02:14 — payment-api p99 latency breach (8 min). Root cause: fraud-check service slowdown. Fix applied: raised payment-api's downstream timeout from 1s to 2s, which stopped the cascading retries. This is a mitigation, not a fix — follow-up ticket still outstanding (see Still-Open Issues above).
Watch List for Next Shift
Redis connection pool: if drops hit 4+ occurrences or exceed 90 seconds, treat as a page-worthy incident rather than a watch item. Fraud-check latency: if it crosses 2s p99, the current timeout mitigation will stop working and payment-api will start erroring instead of just slowing down.
In-Flight Changes
None scheduled for the next shift. A database index migration on the orders table was originally planned for 08:00 but has been pushed to tomorrow — confirm with the DB team before assuming it's still off.
Chaque rotation d'astreinte connaît le même mode d'échec : l'ingénieur sortant sait exactement ce qui s'est passé pendant son équipe, et rien de cette connaissance ne survit à la transmission. L'ingénieur entrant reçoit un message Slack disant que quelque chose s'est passé avec les paiements, hérite d'un tableau de bord plein de pics inexpliqués, et passe les vingt premières minutes de son équipe à reconstruire un contexte qui existait déjà dans la tête de quelqu'un d'autre douze heures plus tôt.
Pourquoi le prompt impose une répartition par gravité
La structure sépare délibérément les «problèmes encore ouverts» des «résolus pendant l'équipe» de la «liste de surveillance», plutôt que de produire une liste plate de ce qui s'est passé. Cette séparation existe parce que l'échec de transmission le plus dangereux n'est pas un fait manquant — c'est un risque actif enterré au milieu d'un mur de texte à côté de cinq éléments routiniers déjà résolus.
Pourquoi «résolu» et «entièrement clos» sont traités comme des états différents
La contrainte exigeant de distinguer un incident résolu d'un incident entièrement clos cible une erreur spécifique et courante : appliquer une mitigation, observer la disparition du symptôme, puis l'écrire comme terminé.
Pourquoi la liste de surveillance a sa propre section
Tout ce qui mérite d'être mentionné n'a pas déclenché d'alerte. Un pool de connexions qui fluctue plus que d'habitude, une métrique de latence en tendance haussière mais encore sous le seuil : ce sont exactement les signaux qui se perdent dans une transmission verbale.
Utilisez ceci à la fin de toute équipe d'astreinte, rotation d'incident, ou transmission de support où le coût de la perte de contexte se mesure en minutes de réinvestigation.