staff-kit

Plugin staff-kit from marketplace claude-staff-kit in rajib2k5/claude-staff-kit. Commands read from the repo's manifest — 2 days ago.

Skills (39)

• claude-agent-orchestration

  Decide how many Claude calls a task needs and who does what - single call, workflow or agent, with a coordinator, scoped subagents, an error and escalation policy, human gates and SDK session handling. Use when designing or reviewing a multi-step or multi-agent Claude system, or when subagent result

• claude-api-ops

  Run Claude API workloads in production with the right model tier, explicit stop_reason handling, safe retries, placed cache breakpoints, batch routing, measured token cost and a plan for model retirement. Use when shipping or tuning a Claude integration, diagnosing truncated or failed calls, cutting

• claude-code-config

  Decide where each piece of Claude Code configuration belongs — CLAUDE.md scope, settings file, permission rule, hook, skill, subagent or slash command — and run Claude Code safely in CI. Use when setting up Claude Code for a team, when a rule is being ignored, or when a pipeline runs `claude -p`. Tr

• claude-context-management

  Manage what Claude sees across a long task - budget the window, decide what to keep, prune, summarise or persist, and design context editing, compaction and memory so durable facts survive. Use when a session or agent run gets long, quality drifts, tool output bloats the window, or work must survive

• claude-evals

  Evaluate the output of Claude-based systems — measurable success criteria, eval sets, code, model and human graders, regression evals on prompt or model change, fact-checking, and independent reviewer passes. Use when you must know whether a prompt, model or pipeline is good enough, whether a change

• claude-governance

  Set the controls that make an organisation's use of Claude responsible and auditable — data handling, human oversight, transparency, bias checks, untrusted-content defence, enterprise admin controls, fallback planning and audit trails. Use when deciding what Claude may see, decide or do, and how the

• claude-prompt-engineering

  Design Claude prompts that hold up in production — structure, explicit criteria, few-shot examples, multi-pass review, and structured output with validation, retry and defensive parsing. Use when writing or fixing a system prompt, a review or extraction prompt, a prompt template, or any code that pa

• claude-rag-and-citations

  Ground Claude answers in sources by choosing between retrieval, long context and project knowledge, attaching claim-level provenance with citations and search_result blocks, and diagnosing wrong or ungrounded answers layer by layer. Use when building or fixing a RAG pipeline, a cited-answer feature

• claude-solution-design

  Turn a business problem into a Claude solution — frame the need, agree scope, pick the product surface and base pattern, place the integration points, brief stakeholders and plan pilot-to-maintenance rollout. Use when someone brings a request that names a tool or a solution before the outcome is cle

• claude-tool-and-mcp-design

  Design tools and MCP servers that Claude selects and calls correctly — granularity, names and descriptions, input schemas, typed error results, tool_choice, tool-set size and MCP server structure and scope. Use when Claude picks the wrong tool, passes bad arguments, loops on errors, or when building

• cost-conscious-orchestration

  Routes work to the cheapest model that can do it correctly, delegates to subagents to protect the main context, and tracks every task on a todo list that is verified complete before the work is reported done. Use when planning multi-step work, dispatching any subagent, choosing a model, deciding whe

• data-cost

  Investigate and reduce warehouse or cloud spend — establish unit economics, find idle and oversized resources, cut data scanned, and quantify savings as validated ranges. Use for any cost investigation or savings estimate. Triggers - reduce costs, snowflake bill, cloud spend, cost optimization, why

• data-dbt-audit

  Audit a dbt project for layering violations, untested keys, unsafe incremental models, missing contracts, and undocumented marts — then report ranked by which would put a wrong number in front of someone. Use on any dbt project you inherit or before trusting its output. Triggers - audit dbt, review

• data-metric

  Define a business metric completely, or reconcile two conflicting versions of one — grain, filters, time semantics, and additivity made explicit. Use when defining a KPI or when two dashboards disagree about the same number. Triggers - define this metric, KPI definition, two dashboards disagree, whi

• data-model

  Design a data model or dbt model correctly — declare the grain first, choose keys, pick the SCD type and materialisation, then write the tests that prove the grain holds. Use before building any table, mart, or dbt model. Triggers - design a table, data model, new dbt model, star schema, dimensional

• data-pipeline-debug

  Root-cause a data pipeline failure across dbt, Airflow, Snowflake tasks and streams, and ingestion — including the hard case where the run succeeded but the data is stale or wrong. Use for any pipeline incident or unexplained data discrepancy. Triggers - pipeline failed, dbt run failed, airflow task

• data-quality

  Hunt the silent data failures — grain violations and join fan-out, freshness gaps, reconciliation drift, absence-vs-zero, broken SCDs, and untested keys. Use when a number looks wrong, when a pipeline succeeds but data seems stale, or before a dataset is trusted. Triggers - data quality, this number

• data-snowflake-audit

  Audit a Snowflake account for RBAC hygiene, warehouse topology, credit waste, missing recoverability, and unmasked PII — every finding backed by a query you ran with its measurement window. Use for platform review or before taking ownership of an account. Triggers - audit snowflake, snowflake health

• data-snowflake-tune

  Tune a slow or expensive Snowflake query using the Query Profile — diagnose pruning, then spill, then idle compute, in that order, with before and after numbers. Use for any slow warehouse query or unexpected credit burn. Triggers - slow snowflake query, expensive query, tune this query, warehouse i

• destructive-ops-guard

  Use before running, generating or approving a command or statement that deletes, drops, truncates, overwrites, force-pushes or rotates something the user owns — a repository, the user's own folders, a live database (Snowflake or any RDBMS server, through SQL of any dialect, dbt, an ORM or a script)

• evidence-gated-validation

  Routes a validation request to the right method first (references/validation-router.md), then runs dual-channel EGV when that is the route. Validates AI-generated code in any programming language, plus ML/AI algorithm implementations, using evidence-gated validation with two independent audit channe

• learn-as-you-build

  Teach the user the code while you build it, so they can explain the architecture and the code in an interview. Builds one milestone at a time, writes a walkthrough doc, then runs a checkpoint of short lessons, each followed by its own quiz question. Has five modes (full, queued, lite, deferred, off)

• simple-english

  Write or rewrite technical text with the rules of ASD-STE100 Simplified Technical English so it is clear, unambiguous, and free of AI slop. Use for documentation, READMEs, runbooks, procedures, error messages, release notes, incident reports, and API guides. Also use when the user says "STE", "Simpl

• snowflake-sql

  Run Snowflake SQL and inspect a Snowflake account with a programmatic access token (PAT) through the Snowflake SQL API, using the bundled stdlib-only Python CLI sf.py. Use for any task that has to execute something in Snowflake - queries, SHOW / DESCRIBE / GET_DDL, grants, query history, warehouse c

• staff-adr

  Write an Architecture Decision Record capturing a decision, the alternatives rejected, and the conditions to revisit it. Use after any significant technical decision. Triggers - write an ADR, document this decision, architecture decision record, why did we choose.

• staff-bootstrap

  Install the global engineering standards into a new or existing project — writes a project CLAUDE.md that inherits the rules library, sets up the agent roster, quality gates, and CI checks. Use when starting a new project, or when an existing repo has no standards wiring. Triggers - new project, boo

• staff-debug

  Root-cause a bug systematically — reproduce, bisect the change, form one hypothesis at a time, and prove the cause before fixing. Use for any bug, test failure, or unexplained behaviour, before proposing a fix. Triggers - debug this, why is this failing, root cause, this is broken, unexpected behavi

• staff-design

  Run a staff-level design process for a feature or subsystem — constraints, non-goals, two or three real alternatives, a recommendation, failure modes, and an incremental migration path. Use before building anything non-trivial. Triggers - design this, architecture for, how should we build, system de

• staff-incident

  Drive a live incident and write the postmortem — mitigate before diagnosing, keep a timeline, then produce a blameless writeup with detection-time analysis and owned actions. Use during an outage or immediately after. Triggers - incident, outage, production is down, postmortem, we have an incident.

• staff-migrate

  Plan and execute a schema, data, or API migration using expand/contract with a resumable backfill, a real rollback plan, and a production-scale dry run. Use before any migration, backfill, or breaking interface change. Triggers - migration, migrate the schema, backfill, breaking change, alter table,

• staff-onboard

  Rapid reconnaissance of an unfamiliar codebase — entry points, architecture, data flow, test and deploy commands, conventions, and the risky areas. Use when starting work in a repo you do not know. Triggers - explain this codebase, onboard me, what does this repo do, get me up to speed, understand t

• staff-perf

  Investigate a performance problem from a measurement rather than intuition — baseline, profile, one change, re-measure, report both numbers with their window. Use when something is slow and the actual cause is needed. Triggers - this is slow, optimize this, performance problem, why is this taking so

• staff-review

  Deep multi-lens review of a diff or PR — correctness, security, data integrity, silent failure, and rollback — with every finding labelled by severity and backed by a concrete failure scenario. Use before merging anything non-trivial. Triggers - review this, review the PR, review my changes, is this

• staff-security

  Threat-model and audit code for exploitable security defects — broken authorization, injection, secret exposure, supply-chain risk, crypto misuse, and prompt injection. Use before shipping anything touching auth, untrusted input, secrets, payments, or PII. Triggers - security review, security audit,

• staff-ship

  Run the definition-of-done gate before declaring work complete — executes build, tests, lint, and types, checks rollback and observability, and reports honestly what was skipped. Use before saying "done", opening a PR, or merging. Triggers - is this done, ready to ship, pre-merge check, definition o

• staff-tdd

  Drive a feature or bugfix test-first — reproduce, observe RED, minimal GREEN, then refactor under a green suite. Use before writing implementation code for any feature or fix. Triggers - write tests, TDD, test first, add test coverage, reproduce this bug.

• staff-validate

  Route a claim, number, migration or fix to the right validation method (validation router), then run it — an Opus validator when routed, never downgraded; deterministic gates plus a Sonnet verifier when that is the route — and empirically confirm any blocker before acting. Use before reporting anyth

• code-mods-essentials

  Build custom hooks and extensions for Claude Code. Use when writing mods that monitor context/tokens, validate operations, or create review interfaces. Covers hook patterns, state management, UI rendering, testing, and marketplace distribution.

• graphify

  Use for any question about a codebase, its architecture, file relationships, or project content — especially when graphify-out/ exists, where the question should be treated as a graphify query first. Turns any input (code, docs, papers, images, videos) into a persistent knowledge graph with god node

Slash commands (12)

• /staff-kit:adr

  Write an Architecture Decision Record capturing the decision, rejected alternatives, and revisit conditions

• /staff-kit:bootstrap

  Wire this project onto the global engineering standards (detect stack, write project CLAUDE.md, verify commands)

• /staff-kit:dbt-audit

  Audit a dbt project — layering violations, untested keys, unsafe incrementals, missing contracts, undocumented marts

• /staff-kit:deep-review

  Multi-lens staff review of a diff or PR — correctness, security, data integrity, silent failure, rollback

• /staff-kit:design

  Run a staff-level design process — constraints, non-goals, real alternatives, recommendation, migration path

• /staff-kit:dq

  Hunt silent data failures — grain/fan-out, freshness, reconciliation, absence-vs-zero, broken SCDs

• /staff-kit:rca

  Root-cause a bug or failure systematically — reproduce, bisect, one hypothesis at a time, prove the cause

• /staff-kit:sf-audit

  Audit a Snowflake account — RBAC hygiene, warehouse topology, credit waste, recoverability, unmasked PII

• /staff-kit:sf-tune

  Tune a slow or expensive Snowflake query from the Query Profile — pruning, then spill, then idle compute

• /staff-kit:ship

  Run the definition-of-done gate — execute build/tests/lint/types, check rollback and observability, report honestly

• /staff-kit:standards

  Print the engineering standards library index — rules, agents, skills, and how they compose

• /staff-kit:validate

  Route a claim, number, migration, or fix to the right validation method and run it (Opus adversarial validator only when routed, never downgraded)

Subagents (19)

• adversarial-validator

  Use when the validation router routes the adversarial method (rows 6–9, 12 when T2+ or concurrency/retries/timezone/currency, 13; row-5 procedure claims) — SQL, schema, migrations, security, "fixed" claims, cost models, or numbers destined for a decision-maker. Its job is to BREAK the work, not conf

• analytics-engineer

  Owns metric definitions and the semantic layer — makes sure a business term means one thing everywhere, defines metrics with grain and filters explicit, and traces a reported number back to its source. Use when defining or reconciling metrics, or when two dashboards disagree. Runs on Opus.

• codebase-explorer

  Read-only reconnaissance across a codebase — locate symbols, map call sites, inventory files matching a pattern, list which tests cover a function, trace an import graph. Use for broad fan-out searches where you want the conclusion, not the file dumps. Mechanical work, runs on Sonnet.

• cost-optimizer

  Finds and quantifies cloud and warehouse spend reduction — unit economics, idle and over-provisioned resources, data scanned, egress, and commitment options. Use for any spend investigation or savings estimate. Every figure it produces must survive adversarial validation. Runs on Opus.

• data-quality-auditor

  Audits datasets and pipelines for the silent failures — grain violations, join fan-out, freshness gaps, reconciliation drift, absence-vs-zero errors, and untested keys. Use when a number looks wrong, when a pipeline succeeds but the data is stale, or before a dataset is trusted for reporting. Runs o

• dbt-engineer

  Builds and audits dbt projects — layering, model design, tests, contracts, incremental strategies, macros, and slim CI. Use for any dbt model work, or when a dbt project needs its layering and incremental logic attacked. Runs on Opus.

• dependency-auditor

  Inventories dependencies and reports versions, licences, known CVEs, unpinned or unlocked entries, and unused or duplicated packages. Mechanical inventory work, runs on Sonnet. Escalate judgement calls about whether to upgrade to the caller.

• docs-writer

  Drafts READMEs, runbooks, ADRs, design docs, and dbt/model documentation from material you provide. Use when the decisions are already made and the task is writing them down accurately. Drafting from given material — runs on Sonnet.

• incident-responder

  Drives a live incident — mitigate first, diagnose second — then produces a blameless postmortem with detection-time analysis and owned action items. Use during an active outage or immediately after one. Runs on Opus.

• migration-planner

  Plans schema, data, and API migrations using expand/contract — with a backfill strategy, a real rollback plan, and a dry run at production scale. Use before any migration, backfill, or breaking interface change. These are tier T3 changes and require adversarial validation. Runs on Opus.

• performance-engineer

  Diagnoses and fixes performance problems from a measurement, never from intuition — profiling, N+1 and query plans, tail latency, concurrency, caching, and complexity budgets. Use when something is slow and you need the actual cause with before/after numbers. Runs on Opus.

• pipeline-debugger

  Root-causes data pipeline failures across dbt, Airflow, Snowflake tasks/streams, and ingestion — including the hard case where the pipeline reports success but the data is wrong or stale. Use for any pipeline incident or unexplained data discrepancy. Runs on Opus.

• principal-code-reviewer

  Deep code review for correctness, security, data integrity, and silent failure — the defects tests do not catch. Use before merging non-trivial changes or when you want a diff attacked rather than approved. Labels every finding by severity with a concrete failure scenario. Runs on Opus.

• release-manager

  Assembles a release — verifies CI status, collects the changelog from commits, checks version bumps and migration ordering, and confirms the rollback path is stated. Mechanical assembly and checking against a checklist, runs on Sonnet. Escalate any judgement call about whether to ship.

• security-auditor

  Threat-models and audits code for injection, broken authorization, secret exposure, supply-chain risk, crypto misuse, and prompt injection. Use before shipping anything handling auth, untrusted input, secrets, payments, or PII. Produces exploitable findings, not a checklist. Runs on Opus.

• snowflake-architect

  Designs and audits Snowflake platforms — warehouse topology, RBAC hierarchy, pruning and clustering, streams/tasks/dynamic tables, Time Travel and cloning, and credit governance. Use for Snowflake design, query tuning, cost investigation, or access-model review. Runs on Opus.

• sre-reliability

  Reviews and designs for reliability — SLIs/SLOs and error budgets, timeouts and retries, circuit breakers, backpressure, graceful degradation, health checks, progressive delivery, and alerting that pages on user pain. Use before shipping a service change or when a system is fragile under load. Runs

• staff-engineer

  Use for system design, architecture decisions, and cross-module change planning — when the question is "how should this be built" or "what will this cost us later", not "write this function". Produces designs with named failure modes, rejected alternatives, and a migration path. Runs on Opus.

• test-engineer

  Designs and writes test suites — reproduction tests for bugs, boundary and error-path coverage, property-based tests, contract tests, and de-flaking. Use when a change needs tests that would actually catch a regression, or when a suite is green but untrustworthy. Runs on Opus.

Hooks (3)

• PreToolUse

  node "${CLAUDE_PLUGIN_ROOT}/hooks/evolv-destructive-guard.js"

• SessionStart

  node "${CLAUDE_PLUGIN_ROOT}/hooks/evolv-skills-enforcer.js"

• Stop

  node "${CLAUDE_PLUGIN_ROOT}/hooks/evolv-skills-enforcer.js"

Similar plugins