- Home
- Skills
- DevOps & Cloud
- Incident Postmortem: Blameless, Evidence-Based RCAs
Works with the AI Agents you already use
Incident Postmortem: Blameless, Evidence-Based RCAs
Incident Postmortem makes your agent write them like an experienced SRE.
Free
Incident Postmortem: Blameless, Evidence-Based RCAs
Example session with this skill installed
Write the postmortem for yesterday's checkout outage. Here's the #incident Slack export, the Grafana screenshots and the deploy log.
- Read your context and instructions
- Compiled the incident postmortem
Postmortem: checkout errors after release 4.12 (SEV2)
Draft for review meeting · sources: #incident export (212 messages), Grafana (3 panels), CI deploy log
Summary
On 7 Nov, release 4.12 caused the payment service to exhaust its database connection pool. 8 % of active customers could not check out for 49 minutes. Rollback at 14:46 UTC mitigated it. No payment data was lost and no double charges occurred.
Impact
| Metric | Value | Source |
|---|---|---|
| Duration | 14:03-14:51 UTC (49 min) | Grafana 5xx panel |
| Failed checkouts | ~1,240 (range 1,100-1,400) | API logs, sampled |
| Customers affected | 8 % of active | analytics |
| SLO budget burned | 38 % of monthly | SLO dashboard |
Timeline (UTC)
| Time | Event | Source |
|---|---|---|
| 14:02 | Release 4.12 deployed | CI |
| 14:03 | Impact starts: 5xx rises to 18 % | Grafana |
| 14:21 | First customer report in support | Slack |
| 14:39 | Detected: error-rate alert pages on-call | Slack |
| 14:44 | Rollback decided | Slack |
| 14:46 | Mitigated: 4.11 redeployed | CI |
| 14:51 | Resolved: 5xx back to baseline | Grafana |
TTD 36 min · TTM 43 min · TTR 49 min
Causes
- Trigger: 4.12 opened a DB connection per item in the bulk-pricing loop.
- Root causes: load tests don't cover bulk pricing (scenario 18 months old); no canary stage.
- Contributing: error-rate alert uses a 30-min window; customers detected the issue before monitoring did.
What went well / where we got lucky
Rollback was fast once decided (2 min). Lucky: it happened outside the evening peak.
Action items
| # | Action | Type | Owner | Pri | Due | Done when |
|---|---|---|---|---|---|---|
| 1 | Add bulk-pricing scenario to load tests | prevent | Platform lead | P1 | 20 Nov | CI fails on pool exhaustion |
| 2 | Canary 10 % for 15 min before full rollout | prevent | DevOps | P1 | 27 Nov | pipeline stage live |
| 3 | Error-rate alert window 30 -> 5 min | detect | On-call lead | P1 | 12 Nov | alert fires in test |
| 4 | Pool saturation alert at 80 % | detect | DBA | P2 | 20 Nov | alert exists |
Open questions for the review
Why was the 14:21 support report not escalated to on-call?
Also drafted: customer-facing report and 4 status-page updates (attached).
Connects securely to your tools. The creator never sees your data.
About this skill
Postmortems people actually learn from
After an outage, postmortems are often written from memory, blame someone, skip the numbers and end with "be more careful". Incident Postmortem makes your agent write them like an experienced SRE.
What it does
- Evidence first: reads alerts, monitoring, logs, deploy and commit history, feature-flag changes, incident chat, tickets and status-page posts, and cites them.
- Precise timeline (UTC + local) with the key moments marked: impact start, detection, acknowledgement, mitigation, resolution. Computes time to detect, mitigate and resolve.
- Impact in numbers: users affected, failed requests or orders, error rates, SLO budget burned, data impact - with sources and honest ranges where data is missing.
- Cause analysis: trigger vs root causes vs contributing factors, 5 Whys / causal tree, a checklist of 30+ contributing factors (testing, release, detection, response, capacity, dependencies, process).
- Blameless language: systems and conditions, not people; roles instead of names.
- Action items that matter: prevent / detect / mitigate / process, each with owner, priority, due date and a "done when" check. No "be more careful".
- Three outputs: internal postmortem, plain-language customer report, and status-page updates (investigating -> identified -> monitoring -> resolved).
- Final review: consistent times, sourced numbers, no blame, no secrets or PII in shareable versions, every cause covered by an action.
Includes
- SKILL.md workflow
- references/analysis.md - 5 Whys example, contributing factor checklist, blameless phrasing table, severity levels
- references/templates.md - internal postmortem, customer report and status-page templates
Works with Claude Code, Claude (Cowork), Codex, Cursor and any SKILL.md-compatible agent.
Requires manual input for revenue impact or private customer data not present in provided logs. Analysis is limited to the context provided in the evidence files.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
This skill has been used 10 times.
Trust & safety
Security scan
Security scan passed 9 days ago
- Free with an Agensi account
Requires