Incident Response

Manage active production incidents from detection to resolution. Stack-agnostic. Tool-agnostic.

This skill is for active incidents and incident process. For after-the-fact analysis, use after-action-report. For planned launches, use launch-runbook.

When to use

• An active incident is happening
• Building incident response procedures
• Defining severity levels
• Setting up on-call rotations
• Training a team on incident response

When NOT to use

• Post-incident retrospective (use after-action-report)
• Planned launches (use launch-runbook)
• Pre-launch issue triage (use qa-testing)

Required inputs

• Awareness of the incident (alert, customer report, internal observation)
• Access to production systems and monitoring
• Roles and authorities clearly defined
• Communication channels operational

The framework: 5 phases

1. Detection

How the incident becomes known.

Detection sources:

• Automated alerts (monitoring, SLO violations, error rate spikes)
• Customer reports (support tickets, social media, status page subscribers)
• Internal observation (engineer notices something off)
• Third-party (security researchers, partners)

On detection:

• Acknowledge within target time (typically 5 to 15 minutes for critical)
• Assess severity (see severity rubric below)
• Page the on-call if not already paged
• Open the incident channel

2. Triage

Establish severity and impact.

Severity rubric:

Severity can change. Re-evaluate as more info emerges.

3. Mitigation

Stop the bleeding before fixing the cause.

Mitigation patterns (faster than full fix):

• Rollback (revert recent deploy)
• Feature flag off (disable the broken feature without deploy)
• Failover (route to healthy replica or region)
• Scale up (more capacity to absorb the load)
• Throttle (reject some traffic to protect the rest)
• Graceful degradation (turn off non-essential features to keep core functional)
• Maintenance mode (last resort, blocks all users)

Mitigation principle: Stop user impact first. Cause analysis second.

4. Communication

Three audiences during an incident:

Internal team:

• Real-time updates in incident channel
• Cadence: every 15 minutes minimum during active incident
• Format: timestamped status updates with what we know, what we're doing, ETA

Internal stakeholders:

• Higher-level updates to broader org
• Cadence: every 30 to 60 minutes
• Format: business-impact framing, not technical detail

External / customers:

• Status page updates
• Cadence: every 30 minutes minimum during active incident
• Format: plain language, no blame, what users are experiencing, what to expect

Communication principles:

• Acknowledge before you have answers ("We're aware and investigating")
• Update on schedule even if no progress ("Still investigating, no new information")
• Never speculate publicly about cause
• Confirm resolution explicitly when restored

5. Resolution

Verified fix, customers restored, incident closed.

Resolution criteria:

• Mitigation in place and verified
• Root cause identified (or explicitly deferred to AAR)
• All affected systems back to normal
• Customers can resume normal use
• Final status update posted (internal and external)
• Incident channel can be closed (or archived for AAR)

After closure:

• Schedule AAR within 1 to 2 weeks
• Capture initial timeline while memories are fresh
• Track follow-up action items

Roles during an incident

For small teams or low-severity incidents, one person can hold multiple roles. Each role's responsibilities should still be explicit.

Decision-making during an incident

The IC's authority:

• Call rollback or other mitigations
• Pull additional people in
• Escalate severity
• Make the call when unclear options exist

Non-decisions to avoid:

• "Let's wait and see" when mitigations are available and impact is occurring
• Discussing root cause while users are actively impacted (mitigate first)
• Premature resolution announcements before verification
• Death-by-committee (pull in lots of people, no one decides)

When in doubt: act. A wrong action that can be rolled back beats inaction while users suffer.

Status page communication patterns

Initial:

"We are investigating reports of [issue]. Updates to follow."

Identified:

"We have identified the issue affecting [scope]. Engineers are working on a fix. Next update by [time]."

Monitoring:

"A fix has been applied. We are monitoring to confirm resolution. Next update by [time]."

Resolved:

"This incident has been resolved. Service has been restored. A full incident report will be posted within [timeframe]."

Patterns to avoid:

• Vague language ("experiencing some issues" - what kind?)
• Missing affected scope ("login is down" - everywhere or just one region?)
• Missing time commitments
• "Should be resolved soon" without verification
• Using "back up" before verification

Workflow

1. Acknowledge. First responder acknowledges within target time.
2. Assess severity. Use the rubric. Open the appropriate response channel.
3. Assign roles. IC, comms, ops at minimum.
4. Communicate. Initial status update. Internal channel active.
5. Investigate. Logs, metrics, recent changes. The four most common causes: a recent deploy, a configuration change, a third-party dependency change, a load spike.
6. Mitigate. Stop the bleeding. Don't wait for full root cause.
7. Verify mitigation. Don't trust dashboards alone; test the user flow.
8. Communicate resolution. Internal and external.
9. Close incident. Final timeline noted. Action items tracked.
10. Schedule AAR. Within 1 to 2 weeks.

Failure patterns

• No clear IC. Multiple people debugging in parallel, no coordination. Slower to mitigate, easier to make conflicting changes.
• Skipping mitigation, going straight to root cause. Users keep suffering while engineers debug.
• Premature "all clear." Announcing resolution before verification.
• Communication silence. Users don't know if anyone is working on it.
• Status updates too vague. "We're working on it" with no detail.
• Speculating publicly about cause. Often wrong, always damaging trust.
• Pulling in too many people. Coordination overhead exceeds value.
• No scribe. The timeline gets lost. AAR has to reconstruct from chat logs.
• Skipping AAR for "minor" incidents. Patterns get missed. Lessons get re-learned.
• Blame culture. People hide mistakes, incidents take longer.

Output format

During an active incident: incident channel updates and status page updates as per the framework above.

After incident close: a brief incident summary feeding into the AAR.

# Incident: [Brief title]

**Date:** [YYYY-MM-DD]
**Severity:** [SEV-1 / 2 / 3 / 4]
**Duration:** [Detection to resolution]
**Customer impact:** [Who, how many, how, or state the gap per the data-availability rule]

## Summary
[1 to 2 paragraphs]

## Timeline
[Timestamped events]

## Mitigation
[What was done]

## Action items
[Follow-ups, with owners]

## AAR scheduled for
[Date]

If required data is unavailable

This skill's output depends on data, measurements, or tool results it cannot generate on its own. When a required input, tool, or data source is unavailable or unverifiable, the sanctioned output is the deliverable with the gap stated: what was needed, what was actually obtained or verified, and which parts of the output are affected. Fabricating, estimating, or interpolating a required number to complete the deliverable is never sanctioned. A stated gap is a complete answer.

Reference files

• references/incident-playbook.md - Severity definitions, roles, status page templates, decision rubrics.