From staff-engineer-mode
Routes engineering decisions spanning ideation, design, development, testing, release, operations, and maintenance to the appropriate specialist before producing guidance.
How this skill is triggered — by the user, by Claude, or both
Slash command
/staff-engineer-mode:staff-engineer-modeThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
```
ONE PRIMARY SPECIALIST BY DEFAULT; INFER ROUTING CONTEXT BEFORE WITHHOLDING
Loading many specialists means routing failed.
Engineering surfaces -- architecture, reliability, operations, security, delivery, data, platform, client, AI/ML, accessibility, cost, readiness, rollout, migration, incidents, doc lifecycle controls, control records, API/service contracts, or system design -- start here.
Do not invoke superpowers:brainstorming, superpowers:writing-plans, another broad process skill, or host orchestration first. Route through SEM, load specialist, then use other tools only for sub-decisions.
"Build X", "design X", "make X reliable", "add HA", "plan rollout", "review service", "prep launch", or "investigate incident" are engineering prompts. Doc lifecycle only: owner, truth, freshness, operational accuracy, missing guidance, or archive; routine README/install typo, markdown, link-text, or copy cleanup does not route. Workflow/process/plan are artifacts, not bypasses. Client compatibility, removal, rollout, safety, or readiness route without names.
To load a specialist, Read <specialist-root>/<slug>.md. Resolve <specialist-root> in this order:
SPECIALIST_ROOT= appears in session context (Claude Code, Cursor, OpenCode), use it.specialists directory beside the loaded plugin checkout.~/.codex/staff-engineer-mode/specialistsspecialists directory next to the loaded GEMINI.mdspecialists/ directory at the router skill's install root.Three rules, all mandatory:
Skill staff-engineer-mode:<slug> returns Unknown skill and is a routing failure.Classify by artifact, phase, surface, and risk; users need not know specialist names.
After loading a specialist, show a structured artifact from its Required
Outputs. If a matching template exists under skills/_shared/assets/templates/,
render its headings or tables in the reply, or use the same shape. Keep
templates, checklists, and reviews user-visible.
Command attempts are event-policy exceptions. Before commits/amends, stage in
one shell command, inspect staged diff, read agent-pr-review, show the review
artifact, record the receipt in its own shell command, then commit in another
command.
Do not combine stage/ack/commit/push or add AI attribution.
Before tags, versions, hosted releases, packages, artifacts, or promotions, read
release-build-reproducibility and production-readiness-review, show the
structured review artifacts to the user, record the receipt in its own shell
command, then run the release command in a separate shell command.
Infer from prompt, branch context, conversation, and loaded context. Do not read new repo files before selecting a specialist, and do not ask for intake fields.
Pick primary and secondary only from this exact list. Never invent, shorten, or paraphrase a slug.
accessibility-gates, agent-pr-review, ai-coding-governance, api-design-and-compatibility,
architecture-decisions, backup-and-recovery, caching-and-derived-data,
client-application-security, code-readability-for-agents, configuration-and-automation-safety,
container-runtime-and-orchestration,
cost-aware-reliability, cryptography-and-key-lifecycle, database-operations, data-contracts,
data-lineage-and-provenance, data-pipeline-reliability, dependency-and-code-hygiene, dependency-resilience,
dev-environment-parity, distributed-data-and-consistency, documentation-lifecycle,
edge-traffic-and-ddos-defense, engineering-control-evidence, event-workflows,
experimentation-and-metric-guardrails, feature-flag-lifecycle, fleet-upgrades,
high-availability-design, identity-and-secrets, incident-response-and-postmortems,
infrastructure-and-policy-as-code, input-validation-and-injection-defense,
internal-service-networking, llm-application-security,
llm-evaluation, llm-serving-cost-and-latency, migration-and-deprecation,
ml-reliability-and-evaluation, mobile-release-engineering,
multi-region-and-data-residency, observability-and-alerting,
oncall-health, operational-ownership-transfer, performance-and-capacity,
persistent-connection-systems, platform-golden-paths, privacy-and-data-lifecycle,
production-readiness-review, progressive-delivery, release-build-reproducibility,
resilience-experiments, resilience-requirements, scheduled-job-reliability, secure-sdlc-and-threat-modeling,
service-decommission-and-sunset, slo-and-error-budgets,
software-supply-chain-security, state-machine-correctness, tenant-isolation,
test-data-engineering, testing-and-quality-gates, vulnerability-management,
web-release-gates
agent-pr-review events.primary (and any secondary) verbatim from the Bundled Specialist Slugs list above; if no listed slug fits, withhold routing instead of inventing or paraphrasing one.engineering-control-evidence only for cross-surface mappings, scorecards, exceptions, or control packs.Select one primary when context is enough. Recommend at most one secondary follow-up. Broad requests become a short sequence, not a pile of loaded specialists.
production-readiness-review.incident-response-and-postmortems first, even if root cause seems elsewhere.Treat "review" as a verb until the artifact proves otherwise.
agent-pr-review; general PR, branch, patch, last commit, staged change, or diff review before merge routes there, including tests-pass or deletion-behavior checks.dependency-and-code-hygiene.production-readiness-review.single_primary: output has one primary specialist unless routing is withheld.secondary_cap: output has no more than one secondary specialist.capability_translation: tool, vendor, or framework names are translated into capability language before routing and not repeated in route fields.scope_check: out-of-scope requests are reframed or declined without specialist names.ambiguity_check: ambiguous prompts infer the discriminating artifact before routing; withheld routes expose no specialist names, candidate routes, confidence labels, drafts, or intake questions.intent_inference: rationale identifies the requested artifact and phase before naming a skill.Load references/routing-matrix.md.
production-readiness-review; mobile startup/crash/offline -> mobile-release-engineering; canary metrics -> progressive-delivery; incidents -> incident-response-and-postmortems.agent-pr-review; surface-specific PRs route narrow.architecture-decisions; ownership transfer/handoff -> operational-ownership-transfer; AI repo legibility -> code-readability-for-agents; retry/timeout/fallback/overload -> dependency-resilience.high-availability-design; residency/geo-routing/replication-aware region placement -> multi-region-and-data-residency; fault injection -> resilience-experiments; telemetry -> observability-and-alerting; alert toil or recurring manual runbook work -> oncall-health.resilience-requirements; game days -> resilience-experiments; proven topology -> high-availability-design.container-runtime-and-orchestration; reconnect/heartbeat/fanout -> persistent-connection-systems; raw headroom -> performance-and-capacity.distributed-data-and-consistency; in-process states/invariants -> state-machine-correctness; restore/corruption recovery -> backup-and-recovery; DB execution/query/schema regression -> database-operations.event-workflows; schema evolution -> data-contracts; pipeline freshness/replay -> data-pipeline-reliability; reported-data provenance -> data-lineage-and-provenance.feature-flag-lifecycle; runtime config mutation -> configuration-and-automation-safety; desired-state drift/reconcile -> infrastructure-and-policy-as-code.release-build-reproducibility; env drift -> dev-environment-parity; provenance/signing/builder isolation -> software-supply-chain-security.migration-and-deprecation; terminal teardown/no-resurrection -> service-decommission-and-sunset; model promotion/drift -> ml-reliability-and-evaluation.production-readiness-review; staged exposure/rollback -> progressive-delivery; build artifact identity -> release-build-reproducibility; browser/mobile gates route client-specific.llm-application-security; eval, retrieval-grounded, or agent task-run checks -> llm-evaluation; serving cost/latency/token/cache/fallback budgets -> llm-serving-cost-and-latency; generic model-provider retry, timeout, circuit-breaker, or overload policy -> dependency-resilience; ML serving reliability -> ml-reliability-and-evaluation.edge-traffic-and-ddos-defense; private service routing -> internal-service-networking; dependency-call policy -> dependency-resilience.test-data-engineering; CI/merge gates -> testing-and-quality-gates; environment drift -> dev-environment-parity.dependency-and-code-hygiene; fleet waves/support windows -> fleet-upgrades; supply-chain trust stays separate.cost-aware-reliability; raw headroom -> performance-and-capacity; LLM token/tail cost -> llm-serving-cost-and-latency.documentation-lifecycle; routine doc copy edits do not route; AI agent rules -> ai-coding-governance; control packs -> engineering-control-evidence.production-readiness-review is used for any broad prompt without a readiness event.| Mistake | Correction |
|---|---|
| Keyword matching | Infer artifact, phase, surface, and risk. |
| Loading every related specialist | Choose one primary; list at most one follow-up. |
| Treating tools as domains | Translate tools to capabilities. |
| Asking intake too soon | Infer from prompt, repo, files, branch context, and conversation first. |
Provides CDSS development patterns for drug interaction checking, dose validation, clinical scoring (NEWS2, qSOFA), and alert classification integrated into EMR workflows.
npx claudepluginhub sirmarkz/staff-engineer-mode --plugin staff-engineer-mode