Claude Skills by grahama1970

Turn vague, drifting, long-running, overnight, or repeatedly failing work into a proof-driven Codex goal with one primary proof command or artifact, explicit completion criteria, a bounded retry limit, operational status snapshots, and fail-closed stop conditions. Use when the user asks for goal help, anti-spiral controls, overnight goal reliability, measurable Definition of Done, blocker reports, or when an agent is drifting from the real task.

Enforce a deterministic pre-plan self-improvement loop before /plan. Use when a task needs proof that the agent read the relevant files, resolved confusion through /interview or /dogpile, and can show a transparent understanding report before creating an orchestration plan.

Benchmark and compare small GPTs for task-specific inference. Tests base, fine-tuned, and prompted models against shared eval datasets. Finds minimum viable model, compares fine-tuned vs prompted, and generates reports.

Ground architecture recommendations in source evidence before advising. Use when a user asks for architecture recommendations, refactor options, system design critique, migration plans, cross-module tradeoffs, asks whether enough files have been read to understand the relevant architecture, or invokes /grounded. This skill prevents premature recommendations by requiring file-backed evidence, explicit architectural boundaries, confidence labels, and unknowns before final advice.

Containerized hardening validation and authorized security auditing tools. All operations run in isolated Docker containers for safety.

Handoff skill to assess project state and provide context for a new agent. Uses /assess logic to identify doc-code gaps, broken features, and next steps.

Persona humming pipeline. Downloads songs, separates vocals, converts to persona voice via RVC, and caches for playback during idle conversation. Orchestrates: /ingest-youtube, /create-stems, /learn-artist, /create-music.

Survey a codebase's animation and motion code as a senior motion advisor, then produce a prioritized audit and self-contained implementation plans for other agents (or cheaper models) to execute. Read-only on source code — it plans improvements, it does not apply them. Use when the user asks to "improve the animations", "audit the motion", "make this app feel better", or wants a roadmap of animation fixes rather than a review of a single diff.

Process audiobooks from the inbox: transcribe to text and organize audio for voice training. Part of Horus's learning pipeline.

Manage book library via Readarr, search for books, and monitor downloads. Handles service health checks and API interactions. Part of Horus's learning pipeline.

Ingest codebases into /memory for knowledge extraction and CWE scanning. Phase 1 extracts functional knowledge (module docstrings, function signatures, class hierarchies, markdown docs) via Python AST. Phase 2 scans for CWE mappings via /taxonomy. Designed to run nightly via /monitor-codebase.

Single-command compliance document ingestion pipeline. Chains extractor → cui-marker → doc2qra → taxonomy → memory for defense/compliance PDFs. Handles NIST, CMMC, DISA STIG, and ITAR documents end-to-end.

Single-command document ingestion pipeline. Chains extractor → cui-marker → doc2qra → taxonomy → memory for PDFs and documents. Handles compliance, research, and general documents end-to-end.

Ingest Kindle books and highlights into the library pipeline. Parses My Clippings.txt, converts .epub/.mobi/.azw3 to text, and outputs to ~/clawd/library/books/ for downstream QRA extraction.

Locate NZBGeek releases with subtitle-rich movie files, auto-tag rage/anger/humor cues, and emit PersonaPlex-ready JSON for Horus Theory-of-Mind benchmarks.

Thin wrapper around the SPARTA pipeline project (79K LOC Python). Runs pipeline steps, checks status, runs tests, launches Explorer UX. The real code lives at ~/workspace/experiments/sparta/src/sparta/pipeline/ — this skill is an interface, not a reimplementation.

Training-only datalake corpus acquisition and coverage balancer. Assesses sector/file-type coverage, plans gap-filling URL manifests, and runs fetcher downloads into approved training corpus roots.

Ingest an entire website into /memory as a RAG resource. Chains /fetcher (crawl + render JS pages) → /extractor (structured extraction) → /doc2qra (extract QRAs) → /taxonomy (bridge tags) → /memory (store). Supports image download and local file output.

Extract transcripts from YouTube videos. Use when user says "get transcript", "transcribe this video", "what does this YouTube video say", "extract captions", "youtube transcript", or provides a YouTube URL and wants the text content.

Ingest YouTube and YouTube Music watch history from Google Takeout exports. Builds music taste profile for Horus persona with /memory integration.

Intent classification and mapping for natural language commands. Uses unsloth/transformers for ML-based intent detection with ArangoDB storage.

Structured human-agent Q&A via HTML or TUI forms. Wizard-style interface with tabbed navigation, numbered options with descriptions, automatic "Other" option, and image support. Mirrors Claude Code's AskUserQuestion UX.

Retrieval-augmented Lean4 proof generation with self-improving lab. Queries 94k+ exemplars from DeepSeek-Prover V1+V2, uses hybrid search (BM25 + semantic + graph), generates via Claude, compiles via lean-interact (20 parallel silos), retries on failure. Lab mode: HF dataset ingestion, parallel compilation, English↔Lean4 autoformalization, GRPO training with compiler rewards, convergence loop, and adversarial benchmarking.

Train RVC models from artist names for vocals and instruments. Full pipeline: YouTube search, download, stem separation (via create-stems), preprocessing, training, and model indexing. Builds a library organized by category.

User-facing continuous datalake learning orchestrator. Watches a directory, runs review-pdf quality loops for PDFs, and ingests non-PDF assets into graph memory for multi-hop traversal.

Analyze movies to extract cinematographic techniques and store them in memory.

General-purpose timeout estimation skill. Trains dual models (duration regression + risk classification) from corpus data and observation feedback. Returns calibrated timeout predictions with confidence intervals for any task type.

Train RVC voice models from artist names. Full pipeline: YouTube search, download, stem separation, preprocessing, training, and model indexing. Builds a library of singing voices organized by category (voice/instrument).

Deterministic 6-layer process verification for AI self-improvement loops. Detects reward hacking, evaluator monkey-patching, and criteria drift via hash-chain audit trails, sealed grading (SHA256 + AST), Lean4 formal proofs, process conformance (embed intent vs action), taxonomy graph sparsity, and SetFit classification. The audited agent does not choose its own auditors.

Local-first, always-on interview and meeting copilot that transcribes speech with RealtimeSTT/faster-whisper and surfaces source-bound evidence from Graph Memory, indexed code, and current repositories. Use when the user says "live interview evidence", "listen during my interview", "surface repo proof while I talk", "meeting copilot", or "find supporting code while someone is speaking".

Evaluate and compare LLM models for task-specific accuracy, cost, and latency. Find minimum viable model, side-by-side comparison, Agent-as-Judge verdicts.

Capture and package a local HTML page or localhost URL into one WebGPT-ready zip with screenshots, source/assets, and reports. Use when the user asks to bundle, zip, screenshot, capture, or upload a local page for WebGPT analysis.

Run one bounded executable harness for a scoped artifact: explorer -> coder -> checks -> code-reviewer -> repair until PASS, BLOCKED, or max attempts.

Mine a mailbox into the typed contact graph under an enforced export-control and PII redaction contract, and gate outbound career mail behind a mandatory /ask roundtable. Mailbox access is delegated entirely to /gmail; this skill adds only what /gmail does not own — what may enter a searchable knowledge graph, and what must happen before a message is sent. Use when an agent needs warm contacts from email, referral paths, reply outcomes for tracked outreach, or validation of an outbound draft ...

Thin CLI wrapper for memory service /match/* endpoints. All domain logic lives in the memory project.

MEMORY FIRST - Query memory BEFORE scanning any codebase. Use when encountering ANY problem, error, or task. Call "recall" FIRST, then scan codebase only if nothing found. Triggers: "check memory", "recall", "have we seen", "remember how".

Mine real human CLI conversation transcripts into labeled training data for bridge/classifier improvement.

Iterative UI mockup generation via Gemini (multimodal) and Google Stitch. Gemini generates mockups from reference screenshots + text specs. Stitch SDK pulls/manages designs from Stitch web UI (text-only, no image input). /review-design scores quality, /interview collects human decisions.

Continuous Claude Code process health monitor. Detects memory-leaked sessions, zombie headless agents, and runaway CPU usage. Fires Discord alerts and optionally auto-kills processes exceeding hard limits. Designed to run on a scheduler.

Continuous codebase health monitoring for all registered projects. Composes project-state, cleanup, ingest-code, quality-checks, skills-ci, dogpile research, and orchestrated fix verification.

Always-on contact freshness monitoring service. Periodically re-researches stale contacts via /dogpile, detects job changes and company news, and alerts via Discord when significant changes are found.

CUSUM and Page-Hinkley statistical drift detection on sensor data streams

Monitor episodic archiver health and run nightly analysis pipeline. Health dashboard, failure patterns, session aging alerts. Nightly: archive + high-fidelity taxonomy + user profiling + lessons.

Monitor Herdr-visible Codex/Claude agent panes for stalled, blocked, unknown, or confused state, nudge genuinely stalled agents back to work with their repo, skill, and open tickets already resolved, and sweep stale disposable workspaces. Never re-asks an agent that stated a blocker and never re-nudges a pane that has not changed since the last prompt.

Nightly orchestrator that verifies the full memory pipeline is healthy. Composes existing skills (ops-arango, taxonomy, persona-journal, monitor-personas, assess, create-walkthrough) plus read-only code-projection checks across 8 tiers. Delegates lifecycle work and never activates projections itself.

Nightly analysis of misuse_events collection to detect new patterns, propose corrections, and track skill health. Self-improving misuse guards.

Continuous Ollama health monitor with auto-restart. Detects hung runner processes (1000%+ CPU spin where /api/tags responds but /api/chat hangs). Runs the watchdog check and restarts Ollama if needed. Designed to run on a scheduler every 2 minutes.

Nightly opportunity monitor that researches a bounded set of target employers and clients, ranks only eligible opportunities, compiles claim-bound resume variants, and emits one interactive morning report covering human-transmitted outreach and separately gated ATS application state. Use when the user asks for today's opportunities, the morning opportunity report, a targeted resume, what needs human action, or to run or schedule the nightly opportunity sweep.

Dashboard and monitoring tool for PDF harvesting and classifier training data preparation. Tracks downloads (Arxiv, MIC, Industry), batch classification, and image extraction progress.

Self-contained persona learning pipeline. Monitors ALL sources (YouTube, RSS, arXiv, Books, Movies, Music, Code), extracts QRAs, classifies into Intent/Persona streams, learns to memory with Federated Taxonomy, archives episodically, verifies edges, reflects on gaps, and trains persona models.