claude-code-config

agent-harness-design

Designing agent harnesses and tool systems — risk taxonomy for tools, permission decisions, draft/commit pattern, structured tool results, agent budgets (10 types), context trust labels against prompt injection, plan-artifact, approval records, observability and traces, evals (13 categories), event model, streaming buffering, 3rd-party skill install checklist, agentic RAG, self-improving SOP loops, model policy, reasoning effort, and Programmatic Tool Calling adoption gates. Use when building a new Agent SDK app, custom orchestrator, MCP server, Cloudflare Worker with tool calls, agentic RAG pipeline, model router, or model-tier policy; when designing tools and permissions; when writing an agent loop; or when you need trust labels for external content. Do NOT use for improving or auditing an already-built harness (use harness-audit / harness-design instead), nor for ordinary Claude Code sessions where the harness is already given.

diffusion-engineering

Практическая инженерия диффузионных моделей: архитектуры, обучение, инференс, оптимизация памяти. Использовать при любых задачах с диффузионными моделями: проектирование или модификация архитектуры (UNet/DiT/Flow/Flux), выбор и настройка schedulers/samplers, дообучение (LoRA/DreamBooth/full fine-tune), оптимизация памяти (AMP/checkpointing/ZeRO/FSDP/quantization), замена или fusion текст-энкодеров (CLIP/Qwen), работа с Diffusers, отладка диффузионных пайплайнов, оценка качества (FID/CLIPScore/LPIPS), latent diffusion, VAE, guidance/CFG, rectified flow, Stable Diffusion, SDXL, Flux. Также применять при вопросах про GPU-память при обучении генеративных моделей, text-to-image пайплайны, ControlNet, multi-encoder fusion, WebDataset. Do NOT use for writing FLUX.2 Klein generation prompts (use flux2-klein-prompting) или для FLUX.2 Klein / Qwen-Edit LoRA-тренинга и edit-LoRA (use flux2-lora-training); это общая диффузионная инженерия, не специфика FLUX.2 Klein.

flux2-klein-prompting

Expert prompt engineering for FLUX.2 [klein] image generation and editing model. Use this skill whenever the user wants to create prompts for FLUX.2 [klein], generate images, edit photos with the klein model, work with multi-reference image editing, or needs templates for T2I/I2I tasks. Trigger for any mention of: FLUX.2, flux klein, BFL API, image editing prompts, text-to-image prompts for FLUX, product mockups, poster generation, UI mockups, sticker packs, character design, seamless textures, or any request to write/improve/translate prompts for FLUX-family models. Also trigger when user asks about guidance_scale, inference steps, distilled vs base modes, or multi-reference workflows. Do NOT use for training a FLUX.2 Klein / Qwen-Edit LoRA (use flux2-lora-training), nor for reconstructing a prompt FROM an existing source image (use forensic-prompt-compiler); this skill is for authoring generation/edit prompts only.

flux2-lora-training

Comprehensive reference for training LoRAs on FLUX.2 Klein 9B and Qwen Image Edit 2511 models. Use this skill whenever the user asks about: training LoRAs for flux2/flux 2 klein/qwen-image-edit, before/after edit LoRAs (head swap, face swap, image editing), inpainting LoRAs, training at larger resolutions, latent space expansion, VAE fine-tuning, multi-reference training (2 input images → 1 output), dataset preparation for edit models, zero_cond_t, ai-toolkit/SimpleTuner/DiffSynth configs, BFS head swap LoRA methodology, Qwen Edit architecture, consistency mode, dual encoding, FuseAnyPart, ACE++, maximum training resolution, или любые вопросы об обучении диффузионных моделей. ВСЕГДА используй этот скилл. Do NOT use for writing FLUX.2 Klein generation/edit prompts at inference time (use flux2-klein-prompting), nor for general non-training diffusion architecture/inference/memory work (use diffusion-engineering); this skill is about LoRA/VAE training, not prompting or serving.

forensic-prompt-compiler

Forensic image-to-prompt compiler for image generation models. Use this skill whenever the user wants to: convert/describe an existing image into a generation prompt, reconstruct a scene as a prompt, generate prompts from reference images for AI image tools (Midjourney, FLUX, Stable Diffusion, DALL-E, or any diffusion model), write prompts that preserve exact visual properties of a source image, or needs precise control over identity-safe subject description, geometry lock, lighting reconstruction, color anchoring, or handler-based special cases (floating scenes, collages, close-ups, jewelry, garments, surreal elements). Also trigger for requests involving: image editing prompts, reference-driven generation, pose description, camera angle locking, fabric/material description, or any "turn this image into a prompt" task. Do NOT use for composing prompts from scratch with no source image (use flux2-klein-prompting), nor for training a LoRA on edit pairs (use flux2-lora-training); this skill reconstructs a promp

ml-research-lab

Machine-learning research loop for dataset curation, fine-tuning, evaluation, inference deployment, experiment tracking, and model explainability. Use when working on ML experiments, training data, model benchmarks, RunPod/GPU runs, classifier quality, vLLM/GGUF serving, SHAP-style model explanations, or research-to-code iterations. Do not use for a simple code edit that has no ML dataset, metric, model, or experiment artifact.

notebooklm-grounded-research

Use when: NotebookLM, notebooklm MCP, large documentation sets, courses, books, papers, or citation-backed research are mentioned. Retrieves a small grounded answer from a stable corpus, preserves citations, and verifies claims against primary documentation, repository code, and tests. Do not use when: the answer is already in a small local file, the source is rapidly changing, or a live runtime/test is the authority.

vlm-segmentation-engineering

Экспертный скилл по прикладной инженерии VLM, сегментационных моделей и диффузионных архитектур для GPU-деплоя. Используй ВСЕГДА когда речь идёт о: SAM2, SAM3, Florence-2, LLaVA, Grounding DINO, OWLv2, YOLO-World, EdgeTAM — выбор модели, интеграция, pipeline, код; диффузионных моделях — UNet/DiT/Flow/Flux, schedulers, LoRA, AMP, ZeRO/FSDP, text encoders (CLIP/Qwen), VAE, CFG; GPU-деплое — MIG, MPS, torch.compile, TorchAO, Triton, memory optimization, два инстанса на H100; open-vocab сегментации и phrase grounding; part-level labeling и instance masks из текстового промпта; замене/fusion текст-энкодеров; fine-tune/LoRA/DreamBooth диффузионных моделей. Триггеры: SAM, Florence, LLaVA, Grounding DINO, YOLO-World, diffusion, UNet, DiT, Flux, LoRA, scheduler, guidance_scale, VAE, CLIP embeddings, Qwen embedder, MIG, MPS, TorchAO, Triton inference, сегментация по тексту, instance masks, open-vocab detection, text-conditioned segmentation. Do NOT use for pure diffusion-only work without a VLM/segmentation component —

feature-new

Scaffold a new feature narrative document in an existing layer following the ULTRAPACK-style template extended for feature-layer architecture (principle 28). Creates docs/layers/<layer>/features/feat-NNN-<slug>.md with Design / Plan / Verify / Conclusion sections, populates layer README features table, and adds entry to feature_list.json if present. Use when: "create a new feature", "start work on feature", "scaffold feature doc", "/feature-new", "new feature in <layer>", "begin feature narrative". Auto-allocates next F-NNN ID. Do NOT use to create the layer itself or its bounded-concern KB scaffold; use /layer-new for that (a feature lives inside an already-existing layer).

harness-design

Design and build multi-agent harness architectures for long-running AI application development. GAN-inspired Generator-Evaluator pattern, Sprint Contract negotiation, context management, quality criteria calibration. Based on Anthropic Engineering patterns. Use when: "build a harness", "multi-agent architecture", "agent orchestration", "generator-evaluator", "long-running app", "harness design", "agent pipeline", "quality evaluation loop", "sprint contract", "build app with agents", "Claude Agent SDK architecture", or when building complex full-stack apps that need planning → generation → evaluation cycles. Also use when discussing context degradation, self-evaluation bias, or assumption testing in AI workflows. Do NOT use to stress-test or critique an already-written plan document; use plan-swarm-review for that (this skill designs the harness, it does not review plans).

layer-new

Scaffold a new layer in a project's docs/layers/ tree following the feature-layer architecture (principle 28). A layer is a bounded concern (security, data, ui, infrastructure, domain) with its own invariants, decisions, gotchas, patterns, and feature narratives. Use when: "create a new layer", "add security layer", "scaffold layer", "start tracking <concern> separately", "/layer-new", "add bounded concern". Operates on the kb-skeleton structure; idempotent -- will not overwrite existing layers. Do NOT use to scaffold an individual feature narrative inside an existing layer; use /feature-new for that (a layer is the container, not the per-feature doc).

AI & Automation Featured

plan-swarm-review

Iterative plan review using multisampling + focused decomposition. Launches parallel independent agents to find issues that single-pass review misses. 4 escalating rounds: broad -> multisample -> focused -> focused+multisample. Use when: "swarm review", "review plan thoroughly", "multisample review", "deep plan review", "plan swarming", "stress test the plan", or before implementing any plan >500 lines or with >3 interacting components. Also use proactively when a large plan is about to be implemented — catch issues before code, not after. Do NOT use to design a multi-agent harness or Generator-Evaluator architecture from scratch; use harness-design for that. Do NOT use to review already-written code/diffs; use deep-review for that (this reviews plans, not implementations).

Show all 36 bundled skills

pixel-art-storyboard

Convert short scene descriptions, book/album cover briefs, or 2-paragraph synopses into seamless-loop animated pixel-art covers rendered as self-contained HTML+canvas. Companion skill to pixel-art-studio. Use when the user asks to "make a cover for", "animated book cover", "looped pixel scene", "convert this story description to pixel art", "create ambient pixel animation", "обложка для книги в пиксель-арте", "анимированная обложка", "封面像素画", "픽셀 아트 표지", or provides a short narrative/synopsis and wants a visual result. Covers: 5-element scene framework (Subject + Setting + Lighting + Palette + Motion), iconographic shorthand for symbolic accents, seamless loop techniques (phase-based parametric, sub-pixel breathing, LCM-clean parallax, deterministic particle systems), loop period selection by mood, three prompt registers (LLM agent / human artist / SDXL LoRA), single-HTML-file deliverable in dark-atmospheric style with parametrized canvas rendering. Generates the same engine pattern as the user's `Grass Fiel

pixel-art-studio

Create production-quality pixel art and animations programmatically. Use when the user asks to "create pixel art", "draw a sprite", "make pixel animation", "generate sprite sheet", "convert image to pixel art", "pixelate this image", "make a pixel character", "пиксель арт", "пиксельная графика", "спрайт", "像素画", "像素艺术", "도트 그래픽", "픽셀 아트", "8-bit/16-bit/hi-bit style", "retro game art", "Aseprite-like output", "indie game sprite". Covers single-frame sprites, frame-by-frame animations, walk cycles, idle/attack/death animations, sprite sheets, GIF/APNG export, image-to-pixel-art preprocessing (downsample + quantize + dither), 30+ bundled palettes (NES, GameBoy, PICO-8, Endesga 32/64, DawnBringer 16/32, Sweetie 16, Resurrect 64, Korean 오방색/단청, Chinese 故宫/青花/五行, Russian Stoneshard-inspired), 5 dithering algorithms (Bayer 2/4/8, Floyd-Steinberg, Atkinson, Ordered, Blue Noise), automated quality scoring (orphan pixels, doublies, banding, pillow-shading, AI-slop detection), and Generator-Evaluator review via the pixe

deep-review

Parallel competency-based code review. Launches independent Agent reviewers per competency (security, performance, architecture, database, concurrency, error-handling, frontend, testing), each with a focused checklist and isolated context. Synthesizes findings into unified report with FIX/DEFER/ACCEPT triage. Use when: "deep review", "thorough review", "parallel review", "review by competency", "full code review", or for large diffs (200+ lines) where /review may be too shallow. Complements /review (pre-landing) — this is for deep dives. Do NOT use just to orient in an unfamiliar codebase or get a structural symbol overview; use repo-map for that (this audits a concrete diff for defects, it is not a navigation map).

distill-feedback

Turn captured user-correction signals into durable rules (learn-from-corrections loop). Use when - /distill-feedback, "process feedback queue", "what corrections did I give you", "encode lessons from my corrections", session-feedback-capture queued sessions, "обнови правила по моим поправкам", "разбери очередь обратной связи". Reads ~/.claude/feedback/queue.jsonl, LLM-semantically detects durable corrections, proposes atomic rules, applies human-gated via delta-merge. Do NOT use to act on a single in-session correction (just apply the fix directly) or to hand-edit settings.json behaviors; this only mines the queued feedback backlog into durable rules.

proof-verify

Plan-based verification - freeze acceptance criteria before building, then verify after with an independent fresh-context agent (the builder must not verify their own work). Use when - "verify against plan", "proof check", "independent review", "check the implementation", or confirming a feature built from a plan meets spec. Do NOT use for quick one-off checks with no plan, or for letting the builder self-verify.

repo-map

Ranked symbol map of a codebase within a token budget — a compact "what matters in this repo" before reading files. Use when starting work in an unfamiliar/large codebase, before a refactor or deep-review fan-out, when you need JIT context instead of dumping whole files, or asked "give me a map of this repo / where are the important functions / what's the structure". Zero-dependency (stdlib only); faithful reimplementation of Aider's repo-map (tree-sitter+PageRank → ranked tags). Do NOT use to find correctness/security defects in a change or to audit a diff; use deep-review for that (this only ranks and lists symbols, it does not evaluate code quality).

workflow-orchestration

Написание и запуск Claude Code dynamic workflows (детерминированный JS-оркестратор субагентов, research preview 2026-05-28). Use when пишешь или запускаешь workflow, видишь keyword workflow в запросе, нужен fan-out на десятки-сотни агентов, codebase-wide аудит/миграция, cross-checked research, competency-review, batch-обработка списка элементов через стадии. Покрывает: примитивы phase/agent/parallel/pipeline/workflow, pipeline vs parallel, schema, budget, resume, quality-паттерны (adversarial verify, judge panel, loop-until-dry), и наши добавки к платформе (retry-обёртка, error policy, .runs observability, eval-harness, billing-дисциплина). Триггеры: workflow, воркфлоу, оркестратор, fan-out, ultracode, deep-research, 1000 агентов, детерминированный скрипт агентов. Do NOT use to design the agent/Generator-Evaluator architecture itself (use harness-design) or for a single one-shot subagent/review where no deterministic multi-stage script is needed; this writes the JS orchestrator, it is not for ad-hoc one-off a

Web & Frontend Solid

frontend-design

Создание высококачественных, визуально выдающихся фронтенд-интерфейсов. Используй ВСЕГДА когда пользователь просит создать веб-страницу, компонент, лендинг, дашборд, UI-кит, форму, карточки, навигацию, анимации, или любой другой веб-интерфейс. Скилл покрывает: HTML/CSS/JS компоненты, React/Vue/Svelte, Tailwind CSS, адаптивный и мобильный дизайн, визуальные стили (glassmorphism, neomorphism, material, flat, градиенты, тёмная тема), интерактивность (drag-and-drop, анимации, hover-эффекты, transitions), верстку (Flexbox, Grid, Container Queries), производительность, доступность (WCAG/ARIA), дизайн-системы и токены. Если пользователь хочет что-то "красивое", "современное", "стильное" в вебе — обязательно используй этот скилл. НЕ используй для нативных мобильных приложений (iOS/Swift/SwiftUI, Android) — для iOS бери скилл ios-development; этот скилл только про веб-UI (браузер).

Web & Frontend Solid

ios-development

Comprehensive iOS app development skill. Use this skill for ANY iOS-related task: writing Swift/SwiftUI/UIKit code, architecting apps, debugging crashes, setting up navigation, networking, data persistence, animations, performance optimization, App Store submission, Xcode configuration. Trigger when user mentions: iOS, Swift, SwiftUI, UIKit, Xcode, iPhone/iPad app, Combine, CoreData, SwiftData, MVVM, TCA, URLSession, async/await, @State/@Binding/@ObservableObject, NavigationStack, XCTest, TestFlight, provisioning profiles, or any Apple platform development. Always use this skill before writing iOS code or architecture. Do NOT use for web frontends (HTML/CSS/JS, React/Vue, browser UI) — even a WebView's page content is web work; use frontend-design for that. This skill is native Apple-platform code only.

massed-compute-ops

Operate Massed Compute GPU virtual machines through its MCP server. Use when the user mentions Massed Compute, massedcompute, MassedCompute, asks what GPU is available, wants to choose or launch training/inference VMs, inspect running instances, configure SSH access, audit hourly spend or billing runway, restart a VM, or terminate rented compute safely.

cross-harness-continuation

Use when work moves between Claude Code and Codex, a session resumes from a handoff, or an agent may rewrite an existing implementation. Loads the shared CONTINUITY.json contract, preserves accepted decisions, checks Git baseline and scope, and requires explicit replan mode for intentional redesign.

desktop-sessions-discovery

Discover, search, and selectively restore Claude desktop app sessions hidden across multiple accountIds. Use when user mentions "missing sessions after account switch", "lost desktop sessions", "where do my old sessions live", or runs multiple Claude accounts on the same machine. Do NOT use for Claude Code CLI session history or resuming work within one session — this only recovers the native desktop app's per-accountId session files; for in-session continuity use handoffs.

gemini-delegate

Делегирование задач в Gemini CLI (несколько OAuth-аккаунтов, свитчер, квоты, передача контекста). Use when - спроси/делегируй gemini, second opinion от другого вендора, bulk-курация картинок/данных, нужен 1M-контекст на чтение, квота gemini выгорела (TerminalQuotaError), переключи gemini аккаунт, прогнать N задач через gemini пока Claude занят. НЕ используй для делегирования в OpenAI Codex (другой вендор/CLI — это скилл codex) и не передавай секреты во внешний LLM.

harness-audit

Score a project's agent harness across 5 subsystems (Instructions / State / Verification / Scope / Lifecycle), identify the bottleneck, and produce a prioritized improvement plan. Use when assessing if a project is ready to graduate to [LONG-RUN] status, when an agent keeps failing despite good models, or when adopting our stack on a new codebase. Do NOT use to design or build a new harness from scratch — this only scores an existing one; for greenfield harness/agent architecture use harness-design (or agent-harness-design).

observability-monitoring

Design, audit, and troubleshoot production monitoring and observability using user-impact checks, layered telemetry, USE/RED, SLI/SLO/SLA, error budgets, cardinality controls, actionable alerting, burn-rate response, and postmortems. Use when asked about monitoring, наблюдаемость, алерты, Prometheus, Grafana, OpenTelemetry, logs, traces, profiles, service health, or incident evidence. Do not use for generic dashboard styling, frontend-only UI work, or unrelated code review.

plan-to-tickets

Use when a large plan, PRD, feature, refactor, research plan, or multi-step coding task must be split into small ready-for-agent tickets with acceptance criteria, verification commands, blockers, and vertical tracer-bullet slices. Do not use for small tasks that should be implemented directly, single-bug fixes, or chat-only summaries.

product-meaning-extractor

Deep product analysis before creating videos, presentations, or ads. Use when: 'analyze product', 'extract value', 'product brief', 'what makes this product special', 'prepare brief', 'understand the product', 'video brief'. Takes a URL or product description and outputs a structured brief with core insight, enemy, transformation, proof, mechanism, and emotional hooks. Based on JTBD, StoryBrand, Obviously Awesome (April Dunford), and Value Proposition Canvas frameworks. Do NOT use for writing the script or scene timing (use video-narrative-arc), scoring an existing script (use script-evaluator), or rendering video (use remotion-production-guide); this is the upstream brief-only step before any script is written.

Web & Frontend Solid

remotion-production-guide

Remotion (React video framework) production guide with Apple-style design rules. Use when: 'create video with remotion', 'remotion project', 'render video', 'product demo video', 'animated video', 'video from code'. Covers project setup, animation library, spring presets, typography rules, color palettes, pacing tables, scene templates, 3D integration, and export settings for all platforms. Do NOT use to write the script or narrative beats (use video-narrative-arc), to critique a script's impact (use script-evaluator), or for FFmpeg audio mastering/captions/platform color-grade on a finished render (use video-post-production); this stage builds and renders scenes in React/code.

script-evaluator

Evaluate video scripts and presentations for flatness, tension, and emotional impact. Use when: 'is this script good', 'review script', 'evaluate video', 'why is this boring', 'flatness check', 'script review', 'improve script', 'rate this video'. Scores 6 dimensions (tension, specificity, emotional arc, hook, customer voice, visual variety), identifies specific problems, and suggests concrete fixes with examples. Do NOT use to generate a new script or scene structure from scratch (use video-narrative-arc), to build the product brief (use product-meaning-extractor), or to render/finish the video (use remotion-production-guide or video-post-production); this only critiques an existing script/scene.

video-narrative-arc

Proven narrative arc templates for product videos and ads. Use when: 'write script', 'video structure', 'narrative arc', 'scene plan', 'storyboard', '15 second video', '30 second video', '60 second video', 'how to structure the video', 'video script template'. Provides beat-by-beat templates with timing, emotional arc mapping, hook formulas, and pacing rules. Covers 10s-90s formats for social, product demos, launches, and pitches. Do NOT use to judge or score an already-written script (use script-evaluator), to derive the product insight/brief from scratch (use product-meaning-extractor), or to render the video in code (use remotion-production-guide); this stage only produces the structured script/beats.

video-post-production

Video post-production rules: audio mastering, color, captions, platform export. Use when: 'add music', 'add voiceover', 'export for tiktok', 'add captions', 'color grade', 'audio levels', 'master audio', 'export settings', 'platform requirements'. Covers FFmpeg patterns, audio chain, subtitle standards, and platform-specific export configs. Do NOT use to generate or animate the video frames/scenes in code (use remotion-production-guide), to write the script (use video-narrative-arc), or to evaluate script quality (use script-evaluator); this is the finishing stage applied to an already-rendered video.

article-structure-review

Структурный self-review технической статьи перед публикацией. Покрывает три дыры, которые не ловятся точечными скиллами типа humanize/infostyle: thesis/proof balance, жанровая чистота, обязательный блок ограничений. Применяется ПОСЛЕ написания первого черновика, ПЕРЕД humanize + infostyle. Основано на фидбеке реальных читателей на опубликованные статьи - классический паттерн "много тезисов / мало доказательств" и отсутствие честного блока про то, что не решено. Use AFTER first draft is done, BEFORE word-level audits. Do NOT use for word-level / sentence-level style fixes or removing AI-generation markers (use humanize-english or humanize-russian for that); this skill only reviews macro-structure (thesis/proof balance, genre purity, limitations block), not phrasing.

humanize-english

Make AI-generated English text sound natural and human. Use when: writing blog posts, articles, marketing copy, any English content that must not read as AI-generated. Covers: burstiness, perplexity, banned words, sentence patterns, transitions, tone. Based on: Liang et al. (arxiv 2406.07016, 15M+ abstracts), GPTZero/Originality research. Use BEFORE publishing any AI-generated English text. Do NOT use for Russian text (use humanize-russian) or for macro-level article structure such as thesis/proof balance or a limitations block (use article-structure-review); this skill operates at the word/sentence level on English only.

humanize-russian

Натурализация русскоязычного текста - убрать маркеры ИИ-генерации. Использовать когда: Хабр-статьи, блоги, маркетинг, любой русский текст который не должен читаться как ИИ. Покрывает: слова-маркеры, калькирование, канцелярит, ритм, отглагольные существительные. Источники: Liang et al. (arxiv 2406.07016), gramota.ru, Хабр 918226, Sber GigaCheck. Use BEFORE publishing any AI-generated Russian text. Do NOT use for English text (use humanize-english) or for macro-level article structure such as thesis/proof balance or a limitations block (use article-structure-review); this skill operates at the word/sentence level on Russian only.