Configures PromQL-based Service Level Objective (SLO) alerting policies for Google Cloud resources registered in App Hub or individually specified. Generates Terraform output. Use when the user asks to configure an SLO or Service Level Objective. Don't use for standard alerting policies.
Configures PromQL-based Service Level Objective (SLO) alerting policies for Google Cloud resources registered in App Hub or individually specified. Generates Terraform output. Use when the user asks to configure an SLO or Service Level Objective. Don't use for standard alerting policies.
SLO Alert Configuration Setup Wizard
This skill guides the user through a structured conversation to configure
PromQL-based Service Level Objective (SLO) alerting policies in Terraform. Your
role is to act as a setup wizard that conceptually models the 4 key components
of an SLO API (Service Scope, Service Level, SLI, and Alert Condition), gathers
the requirements, and outputs a Terraform configuration.
CRITICAL RULES
Structured Conversation: You MUST follow the 4-step wizard workflow
below.
Gather Missing Information: Evaluate all 4 steps below first. Ask the
user for all missing information across all steps in a single response.
DO NOT stop after finding the first missing piece of information.
DO NOT use the ask_question tool. You must ask questions using
plain text in your response and end your turn to wait for the user to
reply.
DO NOT write the Terraform configuration if information is missing.
Skip What Is Known: If the user has already provided information for a
step in their previous messages or initial prompt DO NOT ask them for
it. Move to the next missing piece of information. If ALL information for
Steps 1-4 is provided, call write_to_file to generate the Terraform
configuration without asking for permission to proceed.
Provide Best Practices: Whenever you ask the user a question, you
MUST explicitly state the recommended "Best Practice".
Best Practice Shortcut: If the user asks for "best practices" or
similar, do not overwrite their explicit inputs. SKIP all remaining data
gathering and keep any specific targets or custom metrics they provided. For
all fields left blank, apply the recommended defaults defined in the "SRE
Best Practice Suggestion" of each step.
Terraform Output: Write the generated observability configuration ONLY
as Terraform (.tf) files using the google_monitoring_alert_policy
resource and condition_prometheus_query_language resources.
Alert Strategy: ALWAYS include an alert_strategy block with an
auto_close setting. Leave notification_channels empty unless the user
provides one. Provide plain-English explanations of the PromQL math before
finalizing the conversation.
SETUP WIZARD WORKFLOW
Step 1: Define ServiceScope
Check Context: Identify target resource, service, workload, or
application the user wants to monitor. If you already know, proceed.
Otherwise ask the user to identify it.
Autonomous Investigation: If the user specified a project or general
service name without providing specifics, autonomously use gcloud to
discover the target services in their environment. If multiple services or
workloads are discovered, list all of them and suggest applying SLO ONLY
to the most critical backend services as a best practice.
If you struggle to identify potential resources, ask the user to specify.
Identify Underlying Infrastructure: To resolve the correct PromQL
metric, you MUST know the underlying Google Cloud resource type.
If the user only provides a logical name or an App Hub Service/Workload
name such as projects/.../services/frontend or
projects/.../workloads/backend, you still need to know the underlying
infrastructure.
If the prompt provides the underlying infrastructure, use that
information. Do NOT attempt to discover it.
If you don't know the underlying infrastructure but have a resource
identified, you MUST proactively use gcloud to discover the
infrastructure. If you struggle to identify the resource type, ask the
user to specify.
Label Scoping:
If the user explicitly mentions the resource is in App Hub or provides
an App Hub URI like projects/.../locations/.../applications/..., use
App Hub labels and consult references/app_hub_labels.md to identify
the correct group-by fields.
Otherwise, assume it is a standard Google Cloud resource and use
standard grouping labels such as project_id, location, service_name
for Cloud Run.
Example gcloud commands:
gcloud --quiet apphub applications services list --application=- --location=-
gcloud --quiet apphub applications workloads list --application=- --location=-
Check Context: If the user has already provided a Service Level Target
percentage, an SLI condition/threshold, and a measurement period proceed to
the next step. Otherwise, if any are missing, you MUST ask for them.
Service level target percentages include P-values such as PXX, decimals
such as 0.XX, and percentages like XX%.
Example SLI conditions and thresholds include latency < 500ms or
non-5XX responses.
Prompt: Ask the user for their target reliability, condition/threshold
(if applicable), measurement period, and evaluation intervals ONLY if
they are missing.
SRE Best Practice Suggestion: "SRE Best Practice recommends starting
with a 99.9% (3 nines) slo_target measured over a rolling 28-day
rolling_period, as this aligns well with typical release cycles and
provides a reasonable error budget."
Step 3: Define ServiceLevelIndicator / SLI
Check Context: Has the user specified the exact metric name such as
run.googleapis.com/request_count? If yes, proceed to the next step.
Otherwise, if the user only says "availability" or "latency" without
specifying the EXACT metric name, you may infer the name from the
service type provided a metric for that type is defined in the references.
If the user provides a custom metric and a threshold, assume it is a
Distribution metric and do not ask for further metric details.
You MUST output valid metrics defined in
references/service_metrics.md. If the exact resource type and metric
is not listed, check the public documentation in
references/service_metrics.md to find the exact metric. If you still
cannot find it, you MUST stop and ask the user to provide the custom
metric.
Prompt: Ask the user what specific metric they want to use. You MUST
suggest the inferred standard metric as the recommended best practice. When
interpreting incomplete requests, you MUST explicitly propose the
specific metric string and describe the ratio-based or window-based
definition to the user for confirmation before proceeding.
Metric Mapping: Consult references/service_metrics.md to find the
exact PromQL metric string for the Resource Type identified in Step 1
section 3. If the requested metric type does not exist for the resource in
the references or the primary public documentation, you MUST explicitly
inform the user that there is no default metric and ask them to provide the
specific custom metric name. You MUST provide guidance on how a custom
latency metric might be structured.
CRITICAL: If the primary documentation does not list a default
metric, you MUST NOT try to piece together advanced metrics. Ask the
user to provide the custom metric.
Evaluation Method: Default the EvaluationType to REQUEST_BASED
unless the user specifically describes a window-based requirement,
typically denoted by "good minutes" or "bad minutes".
Window-Based Lookback Period: If the user indicates a window-based
evaluation, you need to know the duration of the lookback windows and
the evaluation interval for each window. You MUST ask the user to
specify both the lookback duration and the evaluation interval if they
have not already provided them. You generate an alerting
policy without this configuration.
Step 4: Define Alerting Policy
Check Context: Has the user specified burn rates? If yes, proceed to the
next step. Otherwise, ask the user to specify a burn rate strategy and
provide a best practice suggestion.
SRE Best Practice Suggestion: SRE Best Practice recommends both a
multi-window fast burn and multi-window slow burn.
Multi-Window Fast Burn: Factor 14.4 over 1h and 5m windows, catching
severe outages quickly without false positives.
Multi-Window Slow Burn: Factor 1 over 3d and 6h windows, catching
system degradation.
Step 5: Generate Configuration
Look up the corresponding PromQL template from
references/promql_templates.md based on the user's choices. Use a
Window-Based template for window-based SLOs.
Populate the template with the ServiceScope labels, ServiceLevel
targets, and ServiceLevelIndicator metrics.
Wrap it in Terraform (google_monitoring_alert_policy).
Present the .tf block with a plain English explanation of the math.
CRITICAL: Explicitly warn the user in the final summary if no
notification channels are configured. Inform them that you can assist with
setting those up if they would like.
Graceful Fallback: If a command exits with an error such as API not
enabled or permission denied, DO NOT try to troubleshoot it and DO
NOT use the schedule tool to wait. Immediately fall back to asking the
user to provide the missing information.
CANNOT
SRE Best Practice Suggestion: SRE Best Practice recommends starting with
two SLIs:
Availability: a Ratio SLI comparing successful requests typically
defined as non-5XX responses, to total requests evaluated as
REQUEST_BASED.
Latency: a Distribution SLI evaluated as WINDOW_BASED such as
99% of 5-minute windows must meet a 300ms threshold.
