b08x/ruby-dev

Use when debugging Ruby code, investigating root causes, identifying waste (dead code, over-engineering), or understanding unfamiliar codebases. Applies four diagnostic methods (Gemba Walk, Muda Analysis, Root-Cause Tracing, Five Whys) based on problem type.

Use when processing, transforming, or analyzing data in Ruby - CSV/JSON parsing, ETL pipelines, data validation, batch processing, and file format conversion.

Use when evaluating the ruby-dev plugin itself — establishing or extending a reproducible baseline, designing eval challenges with planted ground truth, building or running the ruby-dev-evals harness, scoring specialist output, checking rubyist gateway protocol compliance, calibrating an LLM judge, or turning eval findings into prompt fixes. Trigger on 'baseline eval', 'evaluate the plugin', 'eval harness', 'does the plugin beat bare', 'gateway compliance', 'regression test the prompts', 'ruby-dev-evals'.

Use for Ruby AI/LLM work - retrieval pipeline architecture (RAG, RRF hybrid retrieval, clause-level SFL chunking, pgvector), LLM agents, MCP servers, the RubyLLM client (multi-provider chat, tool calling, streaming, embeddings, structured output, acts_as_chat, MCP client), and typed dspy.rb programs (Sorbet signatures, Predict/ChainOfThought/ReAct/CodeAct, DSPy::Module, Toolsets, DSPy::Evals, MIPROv2/GEPA). Trigger on 'RAG', 'LLM agent', 'semantic search', 'ruby_llm', 'RubyLLM.chat', 'tool calling', 'structured output', 'acts_as_chat', 'MCP client', 'dspy', 'DSPy::Signature', 'typed LLM output', 'prompt optimization', 'GEPA', 'MIPROv2'. For tokenization/tagging/lexical tools see ruby-nlp.

Use when building desktop GUI applications in Ruby with glimmer-dsl-libui - native cross-platform windows, forms, tables, drawing areas, custom controls, and data-bound MVP architectures. Trigger on 'desktop app', 'GUI', 'native window', 'glimmer', or 'libui' in a Ruby context.

Use for Ruby projects storing data across more than one database - Ohm (Redis) and Sequel (PostgreSQL/pgvector) model design, picking between them, porting a model between ORMs, and dual-database storage/retrieval patterns (multi-table payload separation, hybrid RRF retrieval, scalar filters). Trigger on 'Ohm model', 'Sequel model', 'Redis model', 'pgvector store', 'multi-database', 'dual-database', 'ORM pattern', 'clause store', 'hybrid retrieval filters'.

Use when Ruby code is slow, allocating too much, or needs optimization - profiling with stackprof, benchmarking with benchmark-ips, reducing allocations, and hoisting lookups. Trigger on 'slow', 'performance', 'optimize', 'profile', 'benchmark', 'memory bloat', or GC pressure in a Ruby context.

Use when applying named refactoring patterns to Ruby code. Matches code issues to known transformations (Zeitwerk, async, frozen strings, etc.) and applies surgical fixes.

Internal phase skill, loaded by the cleaner subagent; do not trigger directly. Use as the Cleaner phase of a Ruby build - diagnoses first, then refactors until the CRAP score meets a stated threshold. Also covers bug diagnosis, named refactorings, and measured performance work.

Internal phase skill, loaded by the hardener subagent; do not trigger directly. Use as the Hardener phase of a Ruby build - runs mutation testing and adds tests until surviving mutants are killed or explicitly justified, and scans dependencies and Rails code for vulnerabilities.

Internal phase skill, loaded by the implementer subagent; do not trigger directly. Use as the Implementer phase of a Ruby build - writes unit tests and production code until the Gherkin acceptance scenarios and unit tests pass, loading the domain skill that matches the task (scaffolding, data pipelines, Ohm/Sequel models, TUI, GUI, NLP, RubyLLM/DSPy, YARD docs).

Internal phase skill, loaded by the specifier subagent; do not trigger directly. Use as the Specifier phase of a Ruby build - turns human-written requirements into Gherkin acceptance scenarios (Given/When/Then) and written QA procedures, and writes no production code. For AI/retrieval systems it also produces the component plan (store, text, client, typed-program layers).

Internal phase skill, loaded by the verifier subagent; do not trigger directly. Use as the Verifier phase of a Ruby build - in a fresh context, runs the dependency rule checker, a SIFT audit, and the acceptance tests, and writes no code. Reports a verdict with Toulmin evidence and severity levels.

Use for classical NLP and text-linguistics tasks in Ruby - sentence/word tokenization, POS tagging, dependency parsing, WordNet lexical lookup, fuzzy/string similarity, TF-IDF/BM25 ranking, topic modeling, and local or hosted transformer inference. Trigger on 'tokenize', 'sentence segmentation', 'POS tagging', 'dependency parsing', 'WordNet', 'fuzzy match', 'TF-IDF', 'BM25', 'topic modeling', 'spaCy', 'transformer inference'.

Use when scaffolding new Ruby projects or gems. Applies rubysmith/gemsmith flag presets by project archetype and runs convention passes for project hardening.

Use when assessing Ruby code quality, performing architectural reviews, or generating tech advisory reports. Applies SIFT Protocol (Structure, Idioms, Functionality, Testing) with Toulmin evidence framework and weighted rubrics. Four report modes - Full, System Design, Tech Advisory, Backlog.

Use when building terminal user interfaces in Ruby - interactive prompts, tables, progress bars, color output, and multi-step workflows using the TTY toolkit and related gems.

Use when generating, adding, or improving documentation comments for Ruby files, including requests phrased as 'YARD docs'. Writes decode-style tags (@parameter, @returns, @raises, @yields, @option, @attribute, @example), converts existing YARD tags to decode tags, and measures documentation coverage with bake decode:index:coverage.