agentleFS
techfleetworkstechfleetworks/enterprise-software-AI-skills/.claude/skills/sre-operational-readiness/SKILL.md

sre-operational-readiness

Skill1 starsChanged 17 days agoMITView on GitHub

Use whenever building, launching, or operating a service that runs in production: new services, new endpoints, or any system that must stay up and be debuggable. Applies Site Reliability Engineering: define SLIs and SLOs and error budgets, instrument the four golden signals, write symptom-based alerts that do not page on noise, and prepare incident response, blameless postmortems, runbooks, and a production-readiness review before launch. Trigger proactively on "SLO," "SLA," "uptime," "reliability," "monitoring," "alerting," "on-call," "incident," "postmortem," "runbook," "observability," "metrics," "dashboards," "goes down," "how do we know if it breaks," or any production launch, even without those words. Pairs with release-deployment-safety (the signals a canary watches), enterprise-architecture-standards (observability), and comprehensive-test-strategy.

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 6 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read references/slis-slos-error-budgets.md.

    line 38
  2. Distinctive:

    Keep the SLO (your internal target) apart from the SLA (the external contractual promise, always looser than the SLO).

    line 52
  3. Very specific:

    Read references/observability-golden-signals.md.

    line 57
  4. Very specific:

    Read references/alerting-and-oncall.md.

    line 71
  5. Distinctive:

    Use error-budget burn-rate alerts: a fast burn pages now, a slow burn opens a ticket.

    line 77
  6. Distinctive:

    Every alert must be actionable and link to a runbook.

    line 80
  7. Very specific:

    Mechanical gate: run scripts/check-alert-hygiene.mjs <alerts-dir> in CI so these rules cannot be skipped by accident.

    line 82
  8. Very specific:

    Read references/incident-response-and-postmortems.md.

    line 87
  9. Very specific:

    Read references/runbooks-and-production-readiness.md.

    line 108
  10. Distinctive:

    Write runbooks: how to deploy, how to roll back, the common failure modes and their fixes, the key dashboards, and the escalation contacts.

    line 110
  11. Distinctive:

    Keep enough live copies that losing one still leaves you redundant (for mission-critical data, a primary plus two replicas).

    line 124
  12. Distinctive:

    Every "the service is reliable / we will know when it breaks" statement is a factual claim, so it follows the repo's evidence discipline …

    line 203

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

SRE and Operational Readiness

Why this exists

Software that "works" and software that "runs reliably at scale" are different disciplines. The gap is operational: knowing the service is healthy from the user's point of view, being paged only when it matters, resolving incidents fast, and learning from failure without blame. Teams that run large systems treat reliability as a measured, budgeted engineering property, not a hope. This skill makes a service operable before it launches, not after the first 2 a.m. page.

It works with the enterprise-architecture-standards observability reference (which covers the instrumentation mechanics) by adding the SRE practices around it: what to measure, when to page, how to respond, and how to learn.

The main idea

100 percent reliability is the wrong target. The right target is an explicit SLO (say 99.9 percent) with an error budget, which is the amount of unreliability you are allowed. That budget turns reliability into a shared, measurable decision: if you are inside the budget, ship features fast; if you have burned it, slow down and fix reliability. Everything below serves that loop: measure user-facing health, alert on it, respond, learn, adjust.

Two honest corollaries sit under that idea. First, optimize for how fast you recover, not for never failing. A system that almost never fails is fragile, because nobody knows how to repair it and the automation has rotted (the point Database Reliability Engineering makes as "why MTTR over MTBF"). Second, an SLO is a social choice, not a math problem. Any single availability number can be gamed, so pick one that reflects real user experience and say who measures it and over what window.

The workflow

Step 1: Define SLIs and SLOs before launch

Read references/slis-slos-error-budgets.md. For each user-facing service:

  • Pick SLIs (the measured indicators) that reflect user experience: availability (good requests over total requests), latency (share of requests under a threshold), correctness, freshness. Measure as close to the user as you can. Keep it to three or fewer; more than that usually means you are tracking symptoms of one real indicator.
  • Set SLOs (the targets) that are realistic and tied to user need, not a vanity 99.999 percent.
  • Express latency as a distribution, never an average. Latency is multimodal, so a mean hides the slow tail. State a percentile over a window (for example, 99 percent of requests between 25 and 100 ms over one minute), and store raw values, not pre-averaged ones.
  • State availability with its window, a cap on a single incident, and the share of users affected before you call it downtime.
  • Derive the error budget (100 percent minus the SLO) and the policy for what happens when it is spent.
  • Keep the SLO (your internal target) apart from the SLA (the external contractual promise, always looser than the SLO).

Step 2: Instrument the four golden signals

Read references/observability-golden-signals.md. Every service emits at least:

  • Latency (split successful from failed requests, and track p50, p95, p99, not just an average).
  • Traffic (requests per second, throughput).
  • Errors (rate and type).
  • Saturation (how full the system is: CPU, memory, queue depth, connection pools).

Plus the three pillars: metrics (aggregate health), logs (structured and correlated, see the owasp and comprehensive-test-strategy logging guidance), and traces (a request's path across services). Add a correlation ID so one request can be followed end to end. Store and reason about these by role (a pool, a tier), not by host, so continuity survives a failover or a rebuild.

Step 3: Alert on symptoms, not causes, and do not page on noise

Read references/alerting-and-oncall.md. The discipline:

  • Page a human only for a user-impacting problem that needs action now (SLO burn, hard-down). Everything else is a ticket or a dashboard, not a page.
  • Alert on symptoms ("checkout error rate above the SLO burn rate"), not causes ("CPU high"). CPU can be high while users are fine, or fine while users are broken.
  • Use error-budget burn-rate alerts: a fast burn pages now, a slow burn opens a ticket. Alert before the budget is gone, not after the breach (know the rapids are downstream while you are still in calm water).
  • Every alert must be actionable and link to a runbook. An alert nobody acts on trains people to ignore alerts, and alert fatigue is itself a reliability risk.
  • Mechanical gate: run scripts/check-alert-hygiene.mjs <alerts-dir> in CI so these rules cannot be skipped by accident. See the Mechanical gate section below.

Step 4: Prepare incident response before you need it

Read references/incident-response-and-postmortems.md. Have these ready in advance:

  • Severity levels (SEV1 through SEV3 or SEV4) with clear definitions and response expectations.
  • Roles: an Incident Commander (coordinates), a Communications Lead, and Operations or subject experts. One person does not do all three.
  • A comms plan: a status page, a cadence for stakeholder updates, and where the team coordinates.
  • Mitigate first, diagnose later. Restore service (roll back, flip a flag, fail over) before you root-cause. This ties directly to release-deployment-safety rollback.

Step 5: Run a blameless postmortem for every significant incident

Same reference. After a SEV, write a blameless postmortem: a timeline built from logs and telemetry (not from memory), the impact, the contributing factors, what went well, what did not, and concrete action items with an owner and a due date. Blameless means focusing on the system and the contributing factors, never on individual blame. People act reasonably given what they knew, so fix the system that let a mistake become an outage. Human error is a symptom of a fragile system, not a root cause, so prefer a prioritized list of contributing factors over a single "root cause." Track the action items to completion, or the postmortem is theater.

Step 6: Write runbooks and pass a production-readiness review

Read references/runbooks-and-production-readiness.md. Before launch:

  • Write runbooks: how to deploy, how to roll back, the common failure modes and their fixes, the key dashboards, and the escalation contacts. A runbook is what lets a tired on-call engineer who did not write the code recover it at 3 a.m.
  • Pass a Production Readiness Review: SLOs defined, monitoring and alerting in place, runbooks written, rollback tested, capacity planned, dependencies and failure modes understood, on-call staffed. Do not launch a service nobody can operate. (Google runs two such gates, a Launch Readiness Review before public traffic and a stricter Handoff Readiness Review before a team takes over an operational burden.)

Step 7: Confirm the data is safe, not just the service

For any stateful service, add three checks the operational-visibility work in Database Reliability Engineering calls the floor: is the data safe, is the service up, are consumers in pain?

  • Keep enough live copies that losing one still leaves you redundant (for mission-critical data, a primary plus two replicas).
  • Monitor backups for success and for restorability. A backup you have never restored is not known to be safe. Replication is not a backup and RAID is not a backup, because both cascade a bad write or a corruption to every copy.
  • Monitor replication for lag, for a stopped thread, and for silent drift (divergence between copies), which is the most dangerous because nothing obvious breaks.

Step 8: Encode operational guarantees as tests where you can

Some reliability properties are testable. Add them to the suite (an @reliability tag) so they do not silently regress:

  @reliability
  Scenario: Health check reflects dependency failure
    Given the database is unreachable
    When the readiness endpoint is polled
    Then it reports "not ready" so the load balancer stops sending traffic

  @reliability
  Scenario: Requests carry a correlation ID through the call chain
    When a request enters the system without a correlation ID
    Then one is generated and propagated to all downstream service calls and logs

  @reliability
  Scenario: Graceful degradation when a non-critical dependency is down
    Given the recommendations service is unavailable
    When a user loads the product page
    Then the page renders without recommendations rather than failing

Mechanical gate

scripts/check-alert-hygiene.mjs is the structural half of the alerting discipline in Step 3 (ADR-0020). Point it at your alerting-rule directory (the Prometheus-style YAML used by Prometheus, Alertmanager, Grafana-managed alerts, Mimir, and Thanos) in CI. It fails the build when an alert is missing a severity, a summary or description, or a runbook link, and when a paging alert watches only a resource cause (CPU, memory, disk, load) with no user-facing symptom in its expression. A deliberate exception, such as a genuine capacity page, is suppressed only with an inline # alert-hygiene:ignore <reason> inside that alert block; a bare ignore with no reason does not suppress. It is a block linter, not a YAML parser: it over-flags rather than hiding a real gap, and the cause-based-page rule reads only the expr: value so a symptom word in a summary cannot wrongly clear it. Its discriminating test is test/alert-hygiene.test.mjs.

The rest of this skill (SLOs that reflect real user experience, incident roles, blameless postmortems, tested restores) is enforced by judge-arch review, by the production-readiness review, and by the @reliability scenarios, not by a linter. Say honestly which properties are gated mechanically and which rest on review.

Before you call it done

  • Each user-facing service has SLIs that reflect user experience, an SLO with a window, an error budget, and a policy for when the budget is spent.
  • The four golden signals are instrumented, with a correlation ID threaded through, stored by role.
  • check-alert-hygiene is green on the alerting rules in this change (or an exception carries a written inline reason): every alert has a severity, a summary, and a runbook, and nothing pages on a bare cause.
  • Incident severities, roles, and a comms plan exist, and the service has runbooks and passes a production-readiness review before launch.
  • For a stateful service, backups are monitored for a tested restore, and replication is monitored for lag, breakage, and drift.
  • The evidence ledger below is filled in: each reliability claim has a state and a re-runnable source, and the central claim ("we will know, and be paged correctly, when this breaks") is at observed, not inferred.

Reference files

TopicFile
SLIs, SLOs, SLAs, error budgets, burn-rate policyreferences/slis-slos-error-budgets.md
Golden signals, metrics, logs, traces, correlation, dashboardsreferences/observability-golden-signals.md
Symptom-based alerting, burn-rate alerts, on-call, alert fatiguereferences/alerting-and-oncall.md
Incident severities, roles, comms, mitigation, blameless postmortemsreferences/incident-response-and-postmortems.md
Runbooks, production-readiness review checklist, capacity, toilreferences/runbooks-and-production-readiness.md
Mechanical gate: alert-hygiene linter (severity, summary, runbook, symptom-not-cause)scripts/check-alert-hygiene.mjs

Evidence & skeptical assessment (required)

Every "the service is reliable / we will know when it breaks" statement is a factual claim, so it follows the repo's evidence discipline (canonical contract: skeptical-audit/references/evidence-discipline.md). A production-readiness review is an evidence ledger: each reliability property is a claim with a state and a re-runnable source, never a bare assertion. Hunt the disconfirming case first (what fails silently, with no page?) before you sign off.

ClaimReaches observed byBecomes proven when
"We alert on symptoms, and every alert is actionable"check-alert-hygiene runs clean on the alerting rules in CIa burn-rate alert fires in a game day and a human follows its runbook to resolution
"We will be paged when users are affected"a fault injected in staging trips the SLO-based alertan error-budget burn-rate alert has caught a real incident before users reported it
"The service degrades gracefully, it does not fall over"the @reliability degradation scenario passes in CIa dependency is failed in production (a drill) and the service still serves
"Backups are safe"a monitor shows the most recent backup succeededa scheduled restore of that backup is validated on a loop, not just taken
"We can operate this at 3 a.m."runbooks exist and the production-readiness review is completethe on-call rotation has recovered the service from its runbook without the author

Definition of done: the ledger covers every reliability property the service needs with a verdict and a re-runnable source; a claim that could be measured (does the alert fire, does the restore work) is observed or proven, never inferred; and no "ready to launch" verdict is given while the central "we will know when it breaks" claim is below observed. A green alert-hygiene gate is necessary, not sufficient.

Why this holds for every developer and every agent

It binds everyone because the gate, the scenarios, and the runbooks live in CI and git, not in memory. Agents get the discipline through the always-on hook, which lists this skill in always-on/enterprise-skills-always-on.md. People and agents alike are held by check-alert-hygiene (a blocking gate that fails an un-actionable alert) and by the @reliability scenarios in the shared suite. It survives any one person because the gate script, the reference files, and the scenarios all live in the repo. The honest limit: the alert gate proves an alert is shaped to be actionable, not that the SLO reflects real user experience or that the postmortem culture is blameless. Those rest on the production-readiness review and judge-arch, and a missed page or a silent failure is treated as a missing alert or a missing signal, to be added the way Rob Ewaschuk's alerting philosophy describes.

Sources and attribution

  • Betsy Beyer, Chris Jones, Jennifer Petoff, Niall Richard Murphy, editors, Site Reliability Engineering, O'Reilly, 2016: service level objectives and indicators (https://sre.google/sre-book/service-level-objectives/), embracing risk and error budgets (https://sre.google/sre-book/embracing-risk/), the four golden signals and symptom-based monitoring (https://sre.google/sre-book/monitoring-distributed-systems/), managing incidents (https://sre.google/sre-book/managing-incidents/), and blameless postmortem culture (https://sre.google/sre-book/postmortem-culture/). [documented]
  • Betsy Beyer, Niall Richard Murphy, David K. Rensin, Kent Kawahara, Stephen Thorne, editors, The Site Reliability Workbook, O'Reilly, 2018: implementing SLOs (https://sre.google/workbook/implementing-slos/) and alerting on SLOs with multi-window, multi-burn-rate alerts (https://sre.google/workbook/alerting-on-slos/). [documented]
  • Rob Ewaschuk, "My Philosophy on Alerting," Google (https://docs.google.com/document/d/199PqyG3UsyXlwieHaqbGiWVa8eMWi8zzAn0YfcApr8Q/edit): page on symptoms, make every page actionable, and add only the alerts that would have caught a real outage. [documented]
  • Laine Campbell and Charity Majors, Database Reliability Engineering, 1st edition, O'Reilly, 2017 (copyright 2018; ISBN 978-1-491-92594-2): why MTTR over MTBF; an SLO is a social choice and any number will be gamed; latency as a distribution, store raw values; keep SLIs to three or fewer; store and reason by role, not host; alert before the budget is spent; is-the-data-safe (enough live copies, backups monitored for restorability, replication lag and breakage and drift); replication and RAID are not backups. [documented]
  • Gene Kim, Jez Humble, Patrick Debois, John Willis, The DevOps Handbook, 1st edition, IT Revolution Press, 2016 (ISBN 978-1942788003): alert fatigue and deleting noisy alerts; overlaying deploys on metric graphs because most outages follow a change; developers sharing the on-call rotation; the Launch Readiness and Handoff Readiness reviews; blameless postmortems and injecting production failure (game days, chaos) to build resilience. [documented]

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.