alert-triage（警报分类）

Alert Triage

Use this skill to turn a stream of alerts into a clear policy and response model.

Core principle

Do not start with channels or tools. Start with the decision.

For each alert, decide:

1. Is it actionable?
2. Is it urgent?
3. Is it trustworthy?
4. Is it new information?
5. Who actually needs it?
6. Should it be immediate, batched, suppressed, ignored, or escalated?

Output model

Classify each alert into one of these outcomes:

• send-now
• batch-later
• ignore
• suppress-as-duplicate
• escalate

When helpful, also assign:

• severity: critical | high | medium | low | info
• confidence: high | medium | low
• audience: operator | owner | team | system
• timing: immediate | next-digest | business-hours | maintenance-window

Workflow

1. Normalize the signal

Convert the raw notification into a compact event record:

• source type
• event summary
• likely impact
• affected system or workflow
• first seen time
• repeat count if known
• current evidence

If the source is noisy or ambiguous, rewrite it into one sentence before classifying it.

2. Check actionability

Ask:

• Can someone do something useful now?
• Does delay make the outcome materially worse?
• Is there a clear owner or audience?

If not actionable, prefer ignore or batch-later.

3. Score urgency

Urgency increases when:

• revenue, uptime, data safety, or customer experience is at risk
• the problem is time-sensitive or irreversible
• the signal indicates active degradation, not just a threshold crossing
• the event affects many users or critical workflows

Urgency decreases when:

• the event is informational only
• the issue is self-healing or reversible
• the metric fluctuation is within normal variance
• the event is outside service relevance or ownership

4. Check trust and evidence

Before escalating, check whether the alert is:

• from a trustworthy source
• supported by more than one signal
• already acknowledged elsewhere
• likely a false positive or transient blip

Low-trust alerts should usually not page people unless impact is potentially severe.

5. Deduplicate and suppress

Treat an alert as a duplicate when it repeats the same underlying issue within the same suppression window.

Use a suppression key based on the smallest stable combination that identifies the problem, for example:

• source
• affected component
• error family
• environment
• severity bucket

Suppress duplicates when the new event adds no meaningful information.

Do not suppress when the event shows:

• a severity increase
• wider blast radius
• longer duration than expected
• a new affected component
• failure of a previous recovery attempt

6. Route by audience

Route based on who can act, not who might be interested.

Default pattern:

• critical and actionable, immediate owner plus escalation path
• high, owner or operating team quickly
• medium, working queue or next digest unless time-sensitive
• low and info, digest or ignore

7. Apply timing policy

Use timing rules such as:

• immediate if impact is active and delay is costly
• next digest if useful but not urgent
• business hours if action can safely wait
• maintenance window if the alert is expected during change work

Quiet hours should reduce noise, but not hide critical actionable events.

8. Produce final policy output

Return a concise table or bullet list with:

• normalized alert
• severity
• final outcome
• audience
• timing
• rationale
• suppression key or digest bucket when relevant

Recommended output template

## Alert triage result

| Alert | Severity | Outcome | Audience | Timing | Reason |
|------|----------|---------|----------|--------|--------|
| [normalized alert] | high | send-now | operator | immediate | customer-facing outage with clear action |
| [normalized alert] | low | batch-later | owner | next-digest | useful trend, no urgent action |
| [normalized alert] | medium | suppress-as-duplicate | system | current window | same root issue, no new information |

Decision heuristics

Prefer send-now when all are true:

• actionable
• time-sensitive
• trusted enough
• meaningful impact

Prefer batch-later when:

• action exists, but delay is acceptable
• the alert is useful as pattern context
• the event is better understood in aggregate

Prefer ignore when:

• no action is needed
• the event is vanity noise
• the source is too weak without corroboration

Prefer suppress-as-duplicate when:

• it is the same incident inside the suppression window
• nothing important changed

Prefer escalate when:

• severity increased
• the response owner did not act in time
• business impact crossed a threshold
• the event exceeded duration or repetition limits

Privacy and portability rules

Keep outputs reusable and marketplace-safe.

Do not include:

• personal phone numbers
• private email addresses
• real user names
• internal channel IDs
• device nicknames
• company secrets
• private endpoints or tokens

Use abstract placeholders instead:

• primary-on-call
• ops-channel
• business-owner
• customer-alerts
• critical-systems

References

Read these when needed:

• references/policies.md for reusable policy patterns
• references/examples.md for worked examples