Claude Skills by ArieGoldkin

Build AI-first applications with RAG pipelines, embeddings, vector databases, agentic workflows (ReAct, multi-agent, Opus 4.5), LLM integration, prompt engineering, streaming, and cost optimization. Use when: building an AI feature, integrating an LLM, setting up vector search, or designing agent architectures. Triggers on: AI app, LLM integration, RAG, vector database, agentic, prompt engineering, AI-native, semantic search, multi-agent, embeddings pipeline, cost optimization

AI coaching & conversational support — empathetic interactions, safety guardrails, motivational interviewing, SMART goals, crisis handling. Applicable to coaching, therapy, education, customer support.

Text embeddings for semantic search and similarity. Covers model selection (OpenAI text-embedding-3, nomic-embed), chunking strategies, batch processing, cosine similarity, and vector DB integration. Use when: converting text to vectors, choosing embedding models, implementing chunking, or setting up semantic search. Triggers on: embeddings, text-embedding, vector, chunking, cosine similarity, semantic search vectors, embedding model, batch embed, dimension reduction

LLM function calling and tool use patterns. Covers JSON tool schemas (strict mode), execution loops, structured output with Pydantic, parallel tool calls, and LangChain tool binding. OpenAI and Anthropic patterns. Use when: enabling LLMs to call external tools, defining tool schemas, building tool execution loops, or returning structured JSON. Triggers on: function calling, tool use, tool schema, structured output, Pydantic validation, tool execution, parallel tool calls, bind tools

Golden dataset curation, backup/restore, validation with schema checks, duplicate detection, and coverage analysis

LLM observability with Langfuse — tracing, evaluation, prompt versioning, cost tracking, and LLM-as-judge scoring

LangGraph workflow patterns for stateful multi-step LLM workflows. State schemas (TypedDict/Pydantic), conditional routing, parallel fan-out/fan-in, human-in-the-loop gates, checkpointing, and supervisor agents. Use when: building LangGraph workflows, adding approval steps, coordinating agents, or persisting state. Triggers on: LangGraph, workflow graph, conditional routing, human-in-loop, checkpointing, supervisor agent, state machine, fan-out

LLM application patterns for evaluation, streaming, and testing. Evaluation: LLM-as-judge, multi-dimension scoring, hallucination detection, Langfuse integration. Streaming: SSE, FastAPI endpoints, tool calls in streams, backpressure. Testing: mocking LLM responses, VCR.py recording, structured output validation. Use when: evaluating LLM quality, adding streaming, or testing AI features. Triggers on: LLM evaluation, LLM-as-judge, quality gate, streaming responses, SSE, test LLM, VCR, mock LLM

Multi-agent fan-out/fan-in workflows — supervisor delegation, conflict resolution, and parallel research pipeline synthesis

Google NotebookLM programmatic access — create notebooks, manage sources (URLs, PDFs, YouTube, Drive), generate audio/video/slides/quizzes, and chat with documents. Triggers on notebooklm, podcast from docs, research notebook.

Local LLM inference with Ollama. Use when setting up local models for development, CI pipelines, or cost reduction. Covers model selection, LangChain integration, and performance tuning.

(Aspirational) Production hybrid search with PGVector + BM25 using Reciprocal Rank Fusion, metadata filtering, and performance optimization for semantic retrieval

Provider-native prompt caching for Claude and OpenAI. Use when optimizing LLM costs with cache breakpoints, caching system prompts, or reducing token costs for repeated prefixes.

RAG pipeline patterns for grounded LLM responses. Covers basic retrieval, citations, hybrid search (semantic + keyword with RRF), context window management, and sufficiency checks for hallucination prevention. Use when: building a Q&A system, adding citations, implementing knowledge bases, or preventing hallucinations. Triggers on: RAG, retrieval augmented, knowledge base, Q&A pipeline, citations, hybrid search, context retrieval, hallucination prevention, grounded responses

Redis semantic caching for LLM apps — vector similarity matching, multi-level cache, TTL strategies, cache warming

Real-time streaming with SSE, WebSockets, ReadableStream — backpressure, reconnection, and LLM streaming patterns

Archive old handoff files to maintain a clean continuity directory. Use when handoffs accumulate.

Archive old ledger sections to reduce file size. Use when ledger exceeds 500 lines.

Check continuity system health and file integrity. Use for periodic maintenance checks.

Multi-session state preservation and context management. Covers save-state, create-handoff, resume-session, check-maintenance, archive, metrics, and setup. Append-until-handoff model with dirty tracking and session heartbeat. Use when: saving progress, ending a session, resuming work, or managing continuity health. Triggers on: continuity, save state, handoff, resume session, session management, context preservation, maintenance check, ledger, archive, continuity setup

View session status and continuity metrics. Use to check current session state.

End session properly by creating a handoff document. Use when finishing work or before long breaks.

Cross-plugin system diagnostics for the claude-forge ecosystem. Checks installed plugins, hook compilation, duplicate hook detection, continuity system health, environment (Node, npm, VCS CLI), and log sizes. Use when: plugins seem broken, hooks are not firing, skills or agents have gone missing, after installing or updating plugins, or for periodic health checks. Triggers on: doctor, diagnose plugins, plugin health, hooks not working, skills missing, plugin not loading, system check, plugin ...

Load previous context from handoffs and ledger. Use when starting a new session or returning after a break.

Update project ledger with current state. Use periodically to preserve context before compaction.

Configure the StatusLine-based context percentage monitor. Use when setting up context window warnings for the first time.

Initialize the continuity system for a new project. Use when setting up continuity for the first time.

Salesforce CRM integration — bidirectional sync, case management, webhooks, bulk operations, SOQL queries

Set up Husky pre-commit hooks with lint-staged (Prettier or Biome), type checking, and tests in the current repo. Use when user wants to add pre-commit hooks, set up Husky, configure lint-staged, add commit-time formatting/typechecking/testing, or says "add pre-commit", "setup hooks for this repo", "wire up husky". Triggers on pre-commit, husky, lint-staged, prettier hook, biome hook, commit hook, setup hooks, format on commit.

Audit a SKILL.md against the repo's Skill Authoring Rules and emit candidate flags for human review (CSO, size, no-op/sediment, completion criteria)

Agentic workflow patterns — ReAct agents, plan-and-execute, self-correcting validation, sliding-window memory, and replanning. Triggers on agent loop, ReAct, plan-and-execute, reasoning loop, self-correcting agent, replanning, agentic workflow

Architecture Decision Records — Nygard templates, context/decision/consequences, lifecycle, and trade-off evaluation. Triggers on ADR, architecture decision, design decision, Nygard, record a decision, decision record, document a trade-off

Jira and Confluence integration via MCP (28 tools). Create/update/transition issues, manage Confluence pages, JQL/CQL queries, sprint planning, and URL parsing. Use when: working with Jira tickets, creating issues, updating status, searching Confluence, editing wiki pages, or sprint management. Triggers on: Jira, Confluence, ticket, issue, sprint, JQL, CQL, atlassian, PROJ-123, wiki page, board, backlog

Audit a SKILL.md against the repo's Skill Authoring Rules and emit candidate flags for human review — CSO compliance, >150-line progressive-disclosure review, no-op/sediment/duplication, and completion criteria. Never edits or deletes. Use when: reviewing or pruning a skill before a version bump, vetting a new skill's quality, or sweeping the corpus for sediment. Triggers on: audit skill, audit-skill, skill quality, prune skill, sediment, no-op, skill review, lint skill

Autonomous goal-driven orchestrator. Classifies a natural-language goal, routes it to the right etk/ctk skill, confirms the plan, and executes. Use when: user describes a goal not a method, the right skill is unclear, or you want the agent to pick the approach. Triggers on: auto-research, figure out, fix the, improve the, get coverage, design a, build the, make sure, optimize the, why isn't

Refine ideas into actionable designs via Socratic questioning (simple mode) or parallel multi-agent analysis (--deep). Includes scoring matrix, tier detection, and security review. Use when: exploring a feature idea, making architectural decisions, comparing design approaches, or starting a complex project. Triggers on: brainstorm, help me design, explore options, architect, how should we, design a, what approach, trade-offs, compare approaches

Ultra-compressed communication mode. Cuts token usage ~75% by dropping filler, articles, and pleasantries while keeping full technical accuracy. Use when user says "caveman mode", "talk like caveman", "less tokens", "be brief", "terse mode", invokes /caveman, or asks for compact responses. Triggers on caveman, caveman mode, talk like caveman, less tokens, be brief, terse, compact response.

Drive the cmux native macOS terminal app (third-party, manaflow-ai/cmux) from its CLI or Unix socket — workspaces, panes, surfaces, WKWebView browser automation, event-driven agent waiting, notifications, sidebar metadata, session restore. Use when the user mentions cmux, wants an agent to control terminal/pane layout, automate a browser panel on macOS, drive the markdown viewer, orchestrate agents in panes, or wire AI-agent hooks. macOS 14+ only and requires the cmux app installed. Triggers ...

cmux Conductor — routes a work item to the right execution topology. Use when: you have a work item or ticket and want the right amount of cmux (including none), you're choosing between fleet/race/solo execution, you want agents opened under your current workspace, or you want /etk:develop run over cmux panes. Triggers on: conduct, cmux run, run this in cmux, fleet this, race this, orchestrate execution, how many agents, dev fleet, open under this workspace

Generate, execute, and heal test suites across unit, integration, and E2E tiers. Automated failure healing, coverage comparison, fingerprint caching, and ARIA accessibility diffing. Supports --target for autonomous coverage improvement and --streak to require N consecutive green runs before a test is kept. Use when: adding tests, increasing coverage, generating test files, or validating untested code. Triggers on: cover, generate tests, test coverage, write tests, add tests, untested, increas...

Guided 6-phase development pipeline: Gate, Design, Hypothesize, Plan, Build (TDD), Verify. Chains brainstorming, quality gates, and testing skills with human checkpoints. Supports greenfield, brownfield, bugfix, and refactor modes. Use when: building a feature end-to-end, implementing a ticket, or following a structured dev process. Triggers on: develop, build feature, implement, create feature, from ticket, structured development, pipeline, end-to-end

Evidence-based verification — test results, coverage metrics, build artifacts, deployment health checks before marking tasks done. Triggers on evidence, verify done, proof of completion, definition of done, is it actually done, mark complete

Autonomous metric-driven iteration loop. Modify code, measure a metric, keep improvements, discard regressions via git rollback. Works with any numeric metric and any direction (minimize/maximize). Use when: optimizing performance, reducing latency, minimizing bundle size, improving Lighthouse scores, or tuning any measurable metric. Triggers on: experiment, optimize, reduce latency, improve performance, minimize, maximize, benchmark, iterate, metric-driven, autoresearch

Observation-driven debugging loop (OHAOI): Observe, Hypothesize, Act, Iterate. Investigates root cause before fixing. Integrates with Jira tickets, GitHub issues, and MR/PR creation. Use when: debugging a bug report, fixing a failing test, tracing unexpected behavior, diagnosing production errors. Triggers on: fix bug, debug, failing test, broken, error, crash, regression, flaky, investigate, root cause, stack trace, reproduce

HIPAA compliance validation — health data handling, API security, data privacy, and security gap identification. Triggers on HIPAA, PHI, health data, compliance check, data privacy, protected health information, BAA

Investigate and triage Sentry issues with structured analysis. Fetches issue data via Sentry API (1Password CLI for auth), aggregates event patterns (releases, browsers, users, environments), cross-references AWS CloudWatch Lambda metrics, and produces an assessment document with root cause hypotheses, decision matrix, and open questions. Use when: a Sentry issue needs investigation before deciding on a fix, you want to understand event distribution patterns, you need server-side AWS cross-co...

Author a standardized, structured MR/PR description — Background (need / before / after / related flows), High-Level Design (API/Infra/Schema/UI/Data change table + mermaid sequence), Pitfalls & Regressions — from the branch diff, then open the MR/PR (draft-to-file, adoption-marked) and hand off to /etk:review-mr. Works on GitLab (glab) and GitHub (gh). Use when: opening a merge request or pull request, preparing an MR/PR, writing an MR/PR description, or wrapping a feature/fix before review....

Complexity assessment and quality gates for task planning — 1-5 scoring, information gathering, escalation triggers. Triggers on quality gate, complexity assessment, task scoring, escalation, planning gate, readiness check, how complex is this

Comprehensive MR/PR review with quality checks, domain-aware parallel agents, structured confidence scoring, evidence gating, composite grading, and YAML+MD findings artifacts handed to /etk:post-mr-comments for inline-anchored posting. Use when: reviewing a GitLab MR or GitHub PR, auditing a teammate's diff before merge, gating CI on review quality, or generating draft inline comments for a large multi-domain change. Triggers on: review-mr, review pr, code review, MR review, audit MR, gate t...

Enumerate ticket acceptance criteria before code changes — fetches ticket via Jira/GitHub MCP, extracts AC bullets/sub-tasks/checkboxes, emits a checklist to restate after each commit. Use when: user mentions a Jira/GitHub ticket ID (NAPP-1234, PROJ-456, gh issue), asks 'what's the scope of NAPP-X', references 'acceptance criteria', starts /etk:fix-bug, /etk:develop, or /etk:review-mr against a ticketed task. Triggers on: ticket ID, NAPP-, PROJ-, acceptance criteria, scope of ticket, AC for, ...