Top Agent Skills

Yuanbao (元宝) groups: @mention users, query info/members.

Pitfalls, examples, and edge cases for Hermes Kanban workers. The lifecycle itself is auto-injected into every worker's system prompt as KANBAN_GUIDANCE (from agent/prompt_builder.py); this skill is what you load when you want deeper detail on specific scenarios.

Design one-off HTML artifacts (landing, deck, prototype).

Decomposition playbook + specialist-roster conventions + anti-temptation rules for an orchestrator profile routing work through Kanban. The "don't do the work yourself" rule and the basic lifecycle are auto-injected into every kanban worker's system prompt; this skill is the deeper playbook when you're specifically playing the orchestrator role.

Throwaway experiments to validate an idea before build.

54 production-quality design systems extracted from real websites. Load a template to generate HTML/CSS that matches the visual identity of sites like Stripe, Linear, Vercel, Notion, Airbnb, and more. Each template includes colors, typography, components, layout rules, and ready-to-use CSS values.

Refactor Vue 3 giant files (>500 lines) using Blueprint-First Mode. Extract composables and UI sub-components while preserving 100% business logic.

Debug Python: pdb REPL + debugpy remote (DAP).

Knowledge comic creator supporting multiple art styles and tones. Creates original educational comics with detailed panel layouts and sequential image generation. Use when user asks to create "知识漫画", "教育漫画", "biography comic", "tutorial comic", or "Logicomix-style comic".

Use Alpaca CryptoDataStream websocket for real-time crypto bars in live trading (OHLC only, volume excluded)

Example skill demonstrating the registry structure and format

Patterns for integrating OpenAI-compatible LLM chat into Shiny apps. Covers UF Navigator, reasoning models, streaming, fallback logic, and credential management.

Stitched Image Quality Control

Fix PyPI package name collisions when local package name exists on PyPI

KINTSUGI quantification plots: Always save QC heatmaps and profiles as PDF

Batch Signal Isolation with Recipe-Driven Processing

Python package reorganization with pyproject.toml inside package directory

Fix PPO training early-stop issues. Trigger when: (1) impossible drawdown values (>100%), (2) training stops too early, (3) need adaptive recovery instead of hard stop.

GPU-only SLURM scheduling for KINTSUGI: never use CPU fallback — measured speedups: stitch 25x, decon 5x, EDF 15x (~13x full cycle)

Fix CuPy NVRTC compilation errors on Windows. Trigger: NVRTC_ERROR_BUILTIN_OPERATION_FAILURE, nvrtc-builtins64 not found

Colab notebook setup pattern. Trigger when: (1) Creating new Colab notebooks, (2) API key file not found, (3) file path errors in Colab, (4) repository extraction fails, (5) 'yfinance fallback' despite keys existing.

CRITICAL: Alpaca API is MANDATORY for all OHLCV data. yfinance is NOT a valid fallback - EVER. Trigger when: (1) any code attempts yfinance for price/volume data, (2) crypto volume filter fails, (3) zero-volume bars detected, (4) API key configuration issues, (5) fallback behavior proposed.

Workflow for private repos + Colab training. Trigger when: (1) training on Colab with private repo, (2) separating dev from production, (3) avoiding untested code in production.

v4.1.0 reward function redesign to fix overtrading and DSR dominance

Multi-format channel name parsing for KINTSUGI CHANNELNAMES.txt files

KINTSUGI NOTEBOOK processing principles: Never sacrifice quality for speed, always use GPU when available. Trigger: performance optimization, CPU/GPU choice, fast mode, quality vs speed in NOTEBOOKS (not SLURM).

Spatial region-aware cell matching for CODEX/scRNAseq integration

Agent validation overhaul: reward weight overrides, fitness decline gate, pinned data, staged experiments

Agent validation v4.3.0 — Make agents act effectively by disabling harmful actions, lowering gates, and injecting cross-run learning

VaR Unit Mismatch Fix

Patterns and pitfalls for modularizing large Shiny apps with plotly, sf, and cross-module reactives

Python dataclass best practices: slots, frozen, validation. Trigger when optimizing dataclasses or creating config classes.

Use when tuning predator-prey dynamics, regime detection coefficients, or cooldown mechanisms

Extended hours trading support for pre-market and after-hours sessions via Alpaca

Debugging constant Markov regime features in RL observations - when HMM probabilities show uniform values instead of dynamic regime estimates

SUPERSEDED by persistent-cache-gap-filling (v2.8.0). Cache data during symbol selection for instant repeat runs.

Globus CLI workflow for staging HuBMAP CODEX datasets from remote endpoints to HiPerGator

Batch signal isolation for Phenocycler spleen channels using matched blank pairs and kintsugi.signal

Windows Console Unicode Compatibility

Diagnosing and fixing bimodal matching score distributions in MaxFuse

End-to-end review of agent validation system for live trader readiness. Audits gating thresholds, model health wiring, live feedback loop, notebook consistency.

GPU-accelerated correlation matrix computation with persistent SQLite caching to eliminate bottleneck at correlation calculation during symbol selection

Use when implementing unified position sizing, Kelly criterion, or combining GARCH+Capital+Portfolio risk systems

Integrate Claude agents into training and live trading (v3.0). Trigger when: (1) setting up multi-agent training, (2) adding agent consultation to live trading, (3) configuring orchestrator, (4) understanding agent roles and safety mechanisms.

Mandatory training archive with model gating (APPROVED/REVIEW/DROP). Trigger when: (1) training run completes, (2) need to decide which models to deploy, (3) want historical training reference, (4) need checkpoint recommendations for overfitting.

Training Data Lifecycle Management (v5.4.2)

Convert single-channel TIF directories to pyramidal OME-TIFF with SubIFDs

GPU-safe parallel processing patterns for KINTSUGI to prevent OOM crashes and ensure Jupyter-compatible progress output

Detect processed data produced with old/buggy KINTSUGI code via timestamp forensics

Per-channel skip-existing checks in Snakemake wrapper scripts for resuming interrupted SLURM jobs. Trigger: Snakemake interrupted job, skip existing channels, resume incomplete cycle, sentinel missing but outputs exist, per-channel completeness check, wrapper script resume logic.

Efficient development testing workflow for HPC environments with VS Code tunneling. Trigger: testing code changes on HPC, development workflow with external data directories, VS Code Remote SSH development

Alpaca broker limitations: crypto shorts blocked (broker doesn't support), stock shorts allowed. Trigger when: (1) shorting gate blocks wrong assets, (2) SELL signals blocked, (3) order type confusion.

Integration fixes between backtest engine, FastAPI dashboard, Vue.js frontend, and live trader

MCP (Model Context Protocol) server configuration for Claude Code integration with trading platform. Trigger when setting up MCP servers, fixing alpaca-mcp-server issues, or adding new integrations.

DEPRECATED in v5.6.0 — see joint-multi-tf-v560 skill. Documents the v5.2.0 dual-model approach (train separate 15Min/1Hour models, combine via weighted voting). Still relevant for: (1) loading legacy v5.5.0 dual models, (2) understanding the historical aggregation layer, (3) resampling pattern via origin='start'.

Risk managers should SIZE trades, not BLOCK them. Trigger when: (1) capital manager blocks trades, (2) risk manager rejects orders, (3) trades fail due to 'exceeds limit', (4) components don't work in harmony.

Fix MCP server 'Executable not found' by resolving absolute paths in .mcp.json

Multi-project signal isolation with cascading recipe resolution

Sentinel Validation & Project Promotion

VALIS registration for CODEX: rigid + non-rigid with tuned smoothing. Earlier 'rigid-only' conclusion was caused by a parameter passing bug — all non-rigid tests ran with unsmoothed OpticalFlowWarper.