Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Prometheus instrumentation discipline: right metric type, right name, right labels. Invoke whenever task involves any interaction with Prometheus metrics โ instrumenting application code, writing PromQL queries, defining alerting or recording rules, choosing metric types, managing label cardinality, building exporters, or reviewing monitoring configuration.
Prometheus
Choose the right metric type, name it clearly, label it sparingly. Prometheus is a pull-based monitoring system built on
a dimensional data model โ every metric is a time series identified by a name and key-value label pairs. Getting this
right at instrumentation time prevents rework later.
Exporters โ [${CLAUDE_SKILL_DIR}/references/exporters.md] Exporter architecture, collectors, help strings,
push-based sources
Metric Type Selection
Choose correctly at instrumentation time โ changing later requires migrating dashboards, alerts, and recording rules.
Question
Answer
Type
Can the value decrease?
No
Counter
Is it a snapshot of current state?
Yes
Gauge
Observing a distribution needing cross-instance aggregation?
Yes
Histogram
Need accurate quantiles from a single instance, known at instrumentation time?
Yes
Summary
None of the above
โ
Gauge
Counter
Monotonically increasing value โ resets to zero only on restart.
Use for: requests served, errors occurred, bytes transferred, tasks completed
API: inc(), inc(v) where v >= 0
Always suffix with _total: http_requests_total
Always apply rate() or increase() in queries โ raw values are meaningless
Prometheus handles counter resets automatically in rate()
Never use a counter for values that can decrease โ that is a gauge
Gauge
Value that goes up and down arbitrarily.
Use for: temperature, memory usage, in-progress requests, queue depth, timestamps
API: inc(), dec(), set(v), set_to_current_time()
No _total suffix
Never apply rate() to a gauge โ use deriv() or delta()
Timestamp pattern: store Unix epoch seconds as myapp_last_success_timestamp_seconds; compute elapsed time with
time() - metric in PromQL
Info metric pattern:myapp_build_info{version="1.2.3", commit="abc"} 1 โ metadata as labels with constant value
1
Histogram
Samples observations into configurable buckets. Produces _bucket{le="..."}, _sum, _count.
Use for: request latencies, response sizes, any distribution needing percentiles or cross-instance aggregation
API: observe(v)
Buckets are cumulative โ le="0.5" includes all observations <= 0.5
Must include +Inf bucket (equal to _count)
Choose buckets matching expected value range; place more buckets near SLO boundaries
Buckets cannot be changed after metric creation
Use histogram_quantile() in PromQL to calculate percentiles
Aggregatable across instances โ the primary advantage over summary
Summary
Calculates streaming quantiles on the client side. Produces {quantile="..."}, _sum, _count.
Cannot be aggregated across instances โ avg(x{quantile="0.95"}) is statistically invalid
_sum and _count without quantiles is a valid and useful configuration
Use summary over histogram only when ALL of these are true:
You need accurate quantiles (not approximate)
From a single instance (no cross-instance aggregation)
You know the exact quantiles at instrumentation time
You accept that adding new quantiles requires code changes
Default choice: histogram. See ${CLAUDE_SKILL_DIR}/references/metric-types.md for detailed comparison.
Naming
Format: <namespace>_<subsystem>_<name>_<unit>_<suffix>. Not all parts required โ minimum is namespace + meaningful
name + unit/suffix.
Naming Rules
Use snake_case โ lowercase with underscores, matching [a-zA-Z_:][a-zA-Z0-9_:]*
Colons (:) are reserved for recording rules โ never use in direct instrumentation
Double underscore prefix (__) is reserved for Prometheus internals
Every metric MUST have a namespace prefix identifying its origin
Always use base units โ seconds not milliseconds, bytes not megabytes. Let visualization tools handle conversion.
Append unit to metric name in plural form: http_request_duration_seconds
Suffix counters with _total, counter-with-unit as _<unit>_total (e.g., process_cpu_seconds_total)
Suffix info metrics with _info, timestamps with _timestamp_seconds
A metric MUST represent the same logical thing across all its label dimensions โ sum() or avg() across all
dimensions should be meaningful. If nonsensical, split into separate metrics.
See ${CLAUDE_SKILL_DIR}/references/naming.md for base units table, component ordering, and full examples.
Labels
Use labels for dimensions you will filter or aggregate by: http_requests_total{method="GET", status="200"} โ not
separate metrics per status. Do not put label names in metric names.
When NOT to Use Labels
Unbounded values โ user IDs, email addresses, full URLs, query strings
High cardinality โ anything above ~100 unique values per metric
Cardinality
Every unique label combination is a new time series โ each costs RAM, CPU, disk, and network. Cardinality math:
total series = metric cardinality x number of targets.
< 10: Safe for most metrics
10-100: Acceptable, monitor growth
100-1000: Investigate alternatives
> 1000: Move analysis out of Prometheus
Label Best Practices
Start with no labels. Add as concrete use cases emerge.
Keep cardinality below 10 per metric as a default target.
Initialize all label combinations you know upfront to avoid missing metrics โ export 0 for known label sets.
Use stable label values. Avoid labels that change frequently.
Never include a "total" label value โ rely on Prometheus sum().
Instrumentation Patterns
Online-Serving Systems (HTTP servers, APIs, databases)
Count requests at completion (not start) โ aligns with error and latency stats
Always have a total requests counter alongside error counters (for ratio calculation)
Offline Processing (queues, pipelines, ETL)
Key metrics per stage: items in (_total), items out (_total), in progress (gauge), last processed timestamp (gauge),
processing duration (histogram).
Export heartbeat timestamps to detect stalled processing
Batch Jobs (cron, scheduled tasks)
Key metrics (push to Pushgateway): last success timestamp (gauge), last completion timestamp (gauge), duration (gauge โ
single run, not distribution).
Batch job durations are gauges (single event), not histograms
Jobs running more often than every 15 minutes should be converted to daemons
Libraries
Instrument transparently โ users get metrics without configuration. Minimum for external resource access: request count
(counter), error count (counter), latency (histogram).
Subsystem Patterns
Logging: maintain log_messages_total{level="..."} counter per log level
Failures: always pair failure counter with total attempts counter for ratio calculation
Caches:cache_requests_total{result="hit|miss"}, evictions (counter), size (gauge), lookup latency (histogram).
Also instrument the downstream system.
See ${CLAUDE_SKILL_DIR}/references/instrumentation.md for threadpool patterns, custom collectors, and performance
tuning in hot paths.
PromQL
Rate and Increase (Counters Only)
rate(counter[5m]) โ per-second rate. Use for alerts and dashboards.
increase(counter[5m]) โ total increase. Sugar for rate() * range_seconds.
irate(counter[5m]) โ instant rate from last two samples. Only for graphing volatile counters.
rate() first, then aggregate:sum(rate(x[5m])) โ never rate(sum(x)[5m]). Rate must see individual counter
resets.
Never rate() a gauge โ use deriv() or delta().
Histogram Quantiles
histogram_quantile(0.95, rate(metric_bucket[5m])) โ single histogram
When aggregating histogram buckets, always preserve le in the by clause:
histogram_quantile(0.95, sum by (job, le) (rate(metric_bucket[5m])))
Average duration: rate(metric_sum[5m]) / rate(metric_count[5m])
PromQL Gotchas
Staleness: most recent sample within lookback period (default 5 min). Series disappears if not scraped within that
window.
Rate window size:rate() needs at least 2 samples. Range should be at least 4x scrape interval. With 15s scrape
use rate(x[5m]) or wider.
Expensive queries: bare metric names can expand to thousands of series. Always filter or aggregate before
graphing. Use recording rules for expensive expressions.
See ${CLAUDE_SKILL_DIR}/references/promql.md for data types, selectors, aggregation operators, vector matching,
over-time aggregation, binary operators, and operator precedence.
Alerting Rules
Alert on symptoms (user-visible impact), not causes. Use dashboards to pinpoint causes after an alert fires.
Online-serving โ High latency, high error rate (user-facing, high in the stack)
Offline processing โ Data taking too long to get through the system
Batch jobs โ Job has not succeeded recently enough (>= 2x normal cycle)
Capacity โ Approaching resource limits that will cause outage without intervention
Only page on latency at one point in the stack โ if overall user latency is fine, don't page on a slow sub-component. If
an alert fires and there's nothing to do, remove it.
See ${CLAUDE_SKILL_DIR}/references/alerting-and-rules.md for alert design, naming conventions, and recording rule
details.
Recording Rules
Pre-compute frequently used or expensive expressions. Format: level:metric:operations.
Aggregate ratios correctly โ aggregate numerator and denominator separately, then divide. Never average a ratio.
Use without for aggregation โ preserves all labels except those removed. Prefer over by.
Use recording rules for expensive dashboard queries, expressions used in multiple alerts, or complex multi-step
aggregations.
See ${CLAUDE_SKILL_DIR}/references/alerting-and-rules.md for full naming convention and recording rule anti-patterns.
Exporters
Write an exporter when the target system does not expose Prometheus metrics natively. For your own code, use a client
library directly.
Prefix all metrics with exporter name: haproxy_up, mysql_global_status_threads_connected
Create fresh metric instances per scrape โ do NOT use global metric variables updated each scrape (race conditions,
stale labels)
Drop pre-computed rates, min/max since start, and stddev from source systems โ export raw counters and current values;
let Prometheus rate() handle the rest
See ${CLAUDE_SKILL_DIR}/references/exporters.md for architecture, collectors, help strings, label rules, and
push-based sources.
Application
When writing Prometheus instrumentation:
Apply all conventions silently โ don't narrate each rule.
Choose metric types based on the decision criteria above.
If an existing codebase contradicts a convention, follow the codebase and flag the divergence once.
When writing PromQL queries:
Always wrap counters in rate() or increase() before further operations.
Prefer without over by for aggregation.
When writing alerting or recording rules:
Follow level:metric:operations naming.
Alert on symptoms, not causes.
Aggregate ratios correctly (numerator and denominator separately).
When reviewing Prometheus code:
Cite the specific violation and show the fix inline.
State what's wrong and how to fix it.
Integration
The coding skill governs workflow; this skill governs Prometheus implementation choices.
