RCA Summary Skill

Setup Check

Before loading context files, check if context/CONTEXT.md exists in the current directory.

If context/CONTEXT.md exists — read it and proceed normally.

If context/CONTEXT.md does not exist — this skill was installed standalone (e.g. via npx skills add). Ask the user these questions before proceeding:

1. Role — junior-sre / senior-sre / sre-manager (shapes output depth and tone)
2. Cloud provider — aws / gcp / azure / on-prem / hybrid
3. Observability stack — e.g. Datadog, Prometheus+Grafana, New Relic
4. Company name and primary services affected (if relevant to this task)

Use the answers inline for this session. For persistent setup across all skills, suggest:

pipx install sre-agent
sre-agent init

Instructions

Step 1: Load Context

Read context/CONTEXT.md for persona. The output format varies significantly by persona:

• junior-sre: learning-focused, detailed enough to build mental model
• senior-sre: dense technical digest, assumes shared context with the team
• sre-manager: business-impact-first, ready to forward to executives

Step 2: Gather Source Material

Accept input from any of these sources (ask the user which applies):

• A completed postmortem skill session in this conversation
• A completed root-cause-analysis skill session in this conversation
• Raw text pasted by the user (incident channel log, notes, timeline)
• A postmortem document file path (read it with the Read tool)

Step 3: Extract the Key Facts

From the source material, extract:

Step 4: Generate the RCA Summary

Produce the summary in the format matching the requested audience:

Format A: Slack / Teams Digest (default)

Optimized for posting in an incident channel or engineering Slack channel. Aim for < 300 words, scannable with bold headings.

*RCA Summary — [Incident Name]* | [Date] | [Duration] | [Severity]

*What happened*
[2–3 sentences. What did users experience? When did it start and end?]

*Root cause*
[1–2 sentences. The systemic reason this was possible, not just the proximate cause.]

*Contributing factors*
• [Factor 1]
• [Factor 2]

*How we detected it*
[1 sentence. Alert, user report, or manual discovery. How long after the problem started?]

*How we fixed it*
[1–2 sentences. Immediate remediation taken.]

*Key metrics*
• Time to detect: Xm | Time to mitigate: Xm | MTTR: Xh Ym
• Users affected: ~N (X%)

*Top action items*
1. [Action] — [Owner role] — [Due date]
2. [Action] — [Owner role] — [Due date]
3. [Action] — [Owner role] — [Due date]

Full postmortem: [link or "in progress"]

Format B: Executive / Management Email

Optimized for forwarding to VP, CTO, or customer success. Non-technical language. Aim for < 200 words.

Subject: Incident Summary — [Service/Feature], [Date]

On [date], [service or feature] experienced [plain-language description of user impact]
for approximately [duration]. This affected approximately [N] users ([X]% of our user base).

Root cause: [one sentence in plain language — no jargon]

We resolved the incident by [resolution description].

To prevent recurrence, we are:
1. [Action item 1 — outcome-focused, not technical]
2. [Action item 2]
3. [Action item 3]

A full technical postmortem is [in progress / available here: link].

Please let me know if you have any questions.

Format C: Knowledge Base / Confluence Entry

Optimized for long-term reference. Structured for searchability.

# RCA: [Incident Name]
**Date:** YYYY-MM-DD | **Severity:** P<N> | **Duration:** Xh Ym | **MTTR:** Xh Ym

## What Happened
[2–3 sentences. Observable impact from user perspective.]

## Timeline (condensed)
- `HH:MM` Started
- `HH:MM` Detected (TTD: Xm)
- `HH:MM` Mitigated (TTM: Xm)
- `HH:MM` Resolved (MTTR: Xm)

## Root Cause
[1–2 sentences. Systemic cause.]

## Proximate Cause
[1 sentence. Immediate technical trigger.]

## Contributing Factors
- **Detection gap:** [why it took Xm to detect]
- **Response gap:** [what slowed response, if applicable]
- **Process gap:** [what process was missing]

## Resolution
[What was done to stop the impact.]

## Action Items
| # | Action | Owner | Due | Status |
|---|---|---|---|---|
| 1 | | | | Open |
| 2 | | | | Open |

## Tags
`[service-name]` `[failure-mode]` `[team-name]`

Full postmortem: [link]

Format D: One-Line Digest (for incident retrospective lists)

For weekly/monthly incident digest tables:

| YYYY-MM-DD | P<N> | [Service] | [Xh Ym] | [Root cause in 10 words] | [Top action item] |

Step 5: Offer Format Selection

If the user did not specify a format, ask: "Which format would you like? A) Slack digest (default) B) Executive email C) Knowledge base / Confluence entry D) One-line digest"

Or produce Format A automatically and offer to convert to others.

Step 6: Follow-Up Prompts

After generating the summary, offer:

• "Want me to also generate a one-line version for your monthly incident digest?"
• "Should I add this to the knowledge base (Format C)?"
• "Would you like a version tailored for customer communication?"

Examples

Example Slack digest (senior-sre persona)

*RCA Summary — DB Connection Pool Exhaustion* | 2024-01-15 | 47 min | P1

*What happened*
Checkout failures (503s) for ~15% of users between 14:03–14:50 UTC.
Impact: estimated 2,400 failed checkout attempts.

*Root cause*
No automated load testing in CI allowed an N+1 query regression (introduced in v2.4.0)
to reach production undetected.

*Contributing factors*
• No connection pool saturation alert (detection took 42 min post-deploy)
• Staging does not mirror production traffic patterns

*How we detected it*
PagerDuty latency P99 alert fired at 14:03 UTC — 42 minutes after deploy completed.

*How we fixed it*
Rolled back to v2.3.9 at 14:45 UTC. Restored in 7 minutes.

*Key metrics*
• TTD: 42m | TTM: 7m | MTTR: 47m
• Users affected: ~2,400 (~15% of active users in window)

*Top action items*
1. Add DB connection pool saturation alert — SRE team — Jan 22
2. Add load test to CI pipeline — Backend team — Feb 5
3. Implement canary deploy for all services — Platform team — Feb 20

Full postmortem: [link]

Guidelines

• The root cause in a summary must be the systemic cause, not "the deploy broke it."
• Never include names of individuals in the summary — only teams and roles.
• TTD / TTM / MTTR must be calculated accurately from the timeline; do not estimate.
• Persona (sre-manager): always lead with user/business impact before technical detail, and ensure action items are outcome-focused ("reduce MTTR by X%" not "add an alert").
• If source material is incomplete (missing timeline, no root cause identified yet): flag the gaps clearly and produce a partial summary with [TBD] placeholders. Do not fabricate data.