Claude Skills by grahama1970

Search and extract clips from ingested movies. Query SRT subtitles, extract video clips at timestamps, and track Horus notes on scenes. Use after /ingest-movie has processed content.

Search and annotate music from ingested YouTube history. Uses Horus Music Taxonomy (HMT) for bridge-based memory integration. Enables episodic associations linking music to lore events.

Search and annotate YouTube transcripts already ingested by /ingest-youtube. Supports per-channel indexing and Horus notes with /memory integration.

Fetch up-to-date library documentation from Context7 API. Use when user asks "how do I use this library", "show me docs for", "library documentation", "API reference for", or needs current documentation for any code library.

Self-improving conversation convergence loop for SPARTA stress tests. Diagnoses unsatisfied sessions, re-runs until personas say satisfactory, outputs structured JSON for /assess consumption. Use when conversation quality is low, sessions are partial, or you need to converge conversations.

Real-time two-way voice conversation with persona. Full-duplex audio via PersonaPlex (Moshi-based) with barge-in, backchannels, emotional state tracking, humming (Federated Taxonomy-driven song selection from /consume-music, episodic memory, and dream motifs), and conversation steering. Supports Discord voice, KDE PipeWire, and Horus WebRTC surfaces.

Report on the continuous PDF extraction learning system. Shows corpus statistics, S00/S04 section estimation quality, failure patterns, pipeline bottlenecks, and trends over time. Reads manifest.jsonl and pattern_registry.json from the 12TB extractor corpus.

Generate annotated PDFs from extraction pipeline run data. Overlays color-coded bounding boxes (text, tables, figures, equations) from S02/S05/S06 onto the original PDF. Supports single PDF, run directory, batch from blacklist, and interactive review server.

Explain a module or codebase with source-grounded architecture diagrams. Read the implementation, choose the view that answers the question, then compose existing diagram skills. Use for system structure, dependencies, execution flow, lifecycle, and assurance explanations.

Multi-round character casting orchestrator for the Horus movie pipeline. Extracts characters from scripts, discovers reference actors, generates identity images, and assigns voices. Runs as Phase 2.5 in create-movie.

Train task-specific classifiers for the extractor pipeline. Supports vision, text, and hybrid classifiers with GRPO training and execution feedback. Includes data collection, confidence-based routing, and shadow deployment.

Orchestrate the end-to-end coding process for the Horus persona. Composes research (/dogpile), isolated execution (/battle), structured implementation (/task, /orchestrate), and brutal code review (/review-code).

Generate comprehensive CONTEXT.md for agent handoff. Captures current state, decisions, lessons learned, and next steps so another agent can continue with zero prior context.

Generate and maintain iterative design boards from image directories. Composite PNG grids, markdown boards with size previews, round tracking, and side-by-side comparison tables.

Start and manage a research-backed UX/design collaboration loop from a natural language design request to brief, references, mockups, structured human review, ux-lab implementation, screenshot/CDP proof, reviewer acceptance, and memory learning. Use when the human asks to create, redesign, research, review, or implement a UX surface; when a design task needs memory, dogpile, ask/scillm collaborators, interview HTML review, best-practices-* routing, mockup-first workflow, or ux-lab handoff.

Build structured evidence cases using Claims-Arguments-Evidence (CAE) trees. AGENT-DRIVEN composable orchestrator: the agent decomposes the question, calls existing skills for data collection and verification, then DECIDES the verdict. Python runner.py is a thin data collector and persistence layer.

Create publication-quality figures, charts, and diagrams. Multi-backend: Graphviz, Mermaid, NetworkX/D3, matplotlib, plotly, seaborn. 50+ visualization types for any domain.

Train task-specific small GPTs (0.5B-1.7B) for Tier 1.5 inference. Supports QLoRA SFT, GRPO with pluggable rewards, Karpathy microgpt from-scratch training, Optuna HP search, iterative self-improvement with holdout gates, and GGUF export for production inference.

GSN (Goal Structuring Notation) diagrams + OMG SACM 2.2 XML export from compliance graph. Supports taxonomy context mapping (Mind tags → SACM Context nodes), OSCAL evidence linking, and Lean4 proof references.

Create icons for Stream Deck or other uses. Fetches from web, resizes to 72x72, creates white/active states.

Create images using AI generation (FLUX.1-schnell, Ollama), Mermaid diagrams, or placeholders. Supports multiple backends with automatic fallback.

Train LoRA intent mappers for structured query generation using GRPO with execution feedback. Uses DeepSeek-R1-Distill-Qwen-7B as base model. Supports SFT warmup, GRPO training with ArangoDB rewards, and iterative improvement with automatic retry on eval failure.

Nightly journal entry creation for all personas. Wraps persona-journal generation with dream pipeline integration -- journal entries feed into create-movie dream residue and dreams feed back into future journals.

Compose a Kling video request from a pydantic-validated scene table and per-character reference images, with no bespoking. Use when a user says create a kling scene, make a kling video from a scene table, compile a scene script to kling, or wants the scene-writing -> kling pipeline run as one gated command instead of hand-built requests.

Convert project manifests into Kling Shot Markup Language (KSML) for video generation.

Generate or bake cinematic 3D LUTs (.cube) from reference/source pairs or presets.

Compose MIDI arrangements and convert between JSON spec and MIDI files. Two modes: 'compose' uses LLM (via /prompt-lab prompt) to create a full arrangement from annotated lyrics + reference MIDI fragments + heart tags. 'from-spec' and 'to-spec' are mechanical JSON↔MIDI conversion via pretty_midi. The arrangement is the creative heart of the music pipeline — it decides what every instrument plays at every beat.

Canonical front door for design creation and redesign. Routes new mockups, multi-model bakeoffs, design improvement, and ship-stage verification through mockup-lab, ask, review-design, create-image, ux-lab, best-practices-react, and test-interactions instead of making users choose among fragmented entrypoints.

Orchestrated movie creation for Horus persona. Guides through phases: Research → Script → Build Tools → Generate → Assemble. Uses Docker-isolated coding environment, free/open-source tools only, with full memory integration.

AI-assisted music creation for Horus persona. Voice conversion (RVC) and music generation (MusicGen/AudioCraft). Stem separation uses create-stems skill. Local-first with GPU acceleration.

Orchestrate paper writing from project analysis. Composes assess, dogpile, arxiv, and review-code skills. Interview-driven workflow with frequent human collaboration to resolve ambiguity and validate each stage.

Create complete PDF test fixtures combining ReportLab tables and AI-generated images. Orchestrates create-table and create-image skills. Creates PDFs designed to test/expose extractor bugs.

Automated peer review of research papers using Shadow-LEGO's self-improving cascade. Multiple fictional reviewer personas score papers against arxiv exemplars on standard rubrics (soundness, novelty, clarity, significance, presentation). The cascade learns from teacher reviews to produce increasingly accurate local reviews at zero marginal cost.

Deliberate persona creation for client modeling, expert profiles, and stakeholder mapping. Integrates with /interview for collaborative creation and /ask for knowledge enrichment. Supports Theory of Mind (BDI), voice training, and simulacrum validation.

Generate QRA (Question-Reasoning-Answer) pairs from controls, documents, or text. Three modes: native (framework definitions), relationship (crosswalk chains), standalone (documents).

Create production-grade React/Tailwind/ShadCN designs with a multi-persona feedback loop. Generates "ready to connect" front-ends, uses a Persona Council (Simulacrum + Taxonomy) to critique designs, consults experts if confidence is low, and iterates until production quality is reached.

Train task-specific regression models from tabular data. Supports linear, ridge, lasso, gradient boosting, random forest, and XGBoost. Auto-selects best model via cross-validation. JSONL/CSV input, joblib output.

Create project-aware evidence reports and report UIs from a given project, surface, artifact, or question. Use when the user asks to create a report, HTML/CSS report, project status report, page-purpose report, implementation/readiness report, source-of-truth audit, application inventory, review readout, or report-to-plan handoff that must apply best-practices-report while also using project-specific sources such as project-knowledge, monitor-sparta, test-interactions, ask persona reviews, do...

Generate scene-specific music using ACE-Step 1.5 via Dockerized FastAPI. Supports reference audio for theme continuity across scenes. Integrates with Federated Taxonomy (HMT) for multi-hop graph traversal in /memory.

Grammarly-like sentence decomposition with NVIS confidence colors. Decomposes question text into annotated fragments showing which entities resolved, which are misspelled, which are fabricated, and which are unknown. Composes /extract-entities for grounding data and /interview for "Did you mean?" clarify prompts.

Orchestrate sound effects (SFX) selection and placement per scene. Integrates with sfx-catalog for library access and memory for learning. Multi-round collaborative workflow with auto-approve for pipeline mode.

Create receipt-backed status surfaces and Tau DAG handoff contracts from local progress, gate, and checker receipts. Use when a workflow needs a global progress bar, evidence status board, local readiness surface, or robust Tau connection without claiming provider readiness or live completion.

Audio stem separation using Demucs (4-stem or 6-stem) and optional UVR ensemble. Uses the Python API entrypoint for clean error handling. GPU-accelerated, local-first.

Creative writing orchestrator for Horus persona. Multi-phase workflow: Initial Thought → Research (movies, books, memory, past) → Dogpile Context → Iterative Writing (2-3 drafts with critiques). Supports all formats: short stories, screenplays, podcast scripts, novellas, flash fiction.

Transform screenplay output from /create-story into animatics with camera framing, visual panels, and timing. Highly collaborative - the skill acts as a creative cinematographer, making suggestions and asking questions throughout the process. Integrates with /memory for learned techniques.

Create, iterate, and deploy dynamic Stream Deck button pages. Design context-aware layouts that auto-switch based on active window, voice mode, or running skills. Evaluate layouts against ground truth, self-correct, and store winners as memory-backed templates with taxonomy tags. Triggers: create streamdeck page, design page layout, optimize deck layout.

Per-surface visual styleguide lifecycle — capture, annotate with token callouts, audit via /review-design, track token debt, and detect visual drift between rounds. Composes /create-design-board and /review-design into a repeatable pipeline.

Create, restyle, preview, and validate self-contained themed SVG diagrams and animations for any surface — READMEs, websites, pitchdecks, docs, artifacts. Use when a user asks for an animated SVG, an animated README SVG, a reusable SVG style preset, a dark neon technical diagram, deterministic CSS-keyframe animation, or a README-safe SVG generated from a semantic scene. README-grade constraints (no JS, self-contained, reduced-motion fallback) are the universal floor, which makes every artifac...

Train vision classifiers for Camelot table extraction strategy prediction. Uses MobileNetV2 with GRPO training and Camelot execution feedback. Integrates with Federated Taxonomy for preset-aware predictions.

Create PDF tables using ReportLab Table class for proper structure detection. Creates tables that Marker and other extractors can parse. Use when you need detectable tables in PDFs.