Claude Skills by hiyenwong

Live-Evo - Online Evolution of Agentic Memory

SE-Search - Self-Evolving Search Agent

Self-Improving LLM Agents at Test-Time

SERP - Self-Evolutionary RePlanning

Repository synchronization workflow for AI research skills. Ensures consistent naming between skill directories, SKILL.md frontmatter, and INDEX.md references. Includes verification steps to prevent sync gaps.

Artificial Intelligence applications in complex network science - network analysis, topology learning, dynamics prediction, and emergent behavior detection. Comprehensive survey covering AI potential, methodology, and applications. Use when analyzing complex networks, network topology learning, dynamics prediction, emergent behavior, social networks, biological networks, or transportation networks. Keywords: complex networks, network science, AI networks, topology dynamics, emergent behavior,...

Methodology for mapping real-world AI-enabled cyber attacks onto MITRE ATT&CK framework with AI Risk Enablement Score (ARiES) — identifying patterns in how threat actors weaponize AI for cyber operations.

Methodology from Anthropic research (Jun 3, 2026) mapping a year's worth of AI-enabled cyber threats using MITRE ATT&CK framework — threat categorization, attack pattern analysis, and security implications.

Statistical-causal reframing of AI interpretability: treating explanations as parameters of statistical models inferred from computational traces, with uncertainty quantification and testing against alternative computational hypotheses. Inspired by the famous 'dead salmon fMRI' study. Activation: dead salmon AI, interpretability statistics, causal interpretability, explanation uncertainty, statistical AI explanation, false discovery interpretability.

AI interviewer methodology for operationalizing Descriptive Experience Sampling (DES) into an explicit, inspectable reasoning architecture. Grounded in the established phenomenological method with co-development by DES originator Russell T. Hurlburt. Use when implementing AI systems for studying inner experience at scale, conducting qualitative interviews with temporal grounding, or developing LLM-based research platforms for subjective experience sampling.

AI-assisted mathematical discovery methodology. Use when: (1) collaborating with LLMs to generate mathematical conjectures, inequalities, bounds, or proofs; (2) verifying AI-generated mathematical results; (3) structuring human-AI mathematical research workflows; (4) exploring AI's role in mathematical research; (5) analyzing mathematical inequality patterns (Gaussian perimeter, moment comparison, autoconvolution, Sidon sets, Szarek's inequality). Trigger words: AI math discovery, Grokability...

Design and implement memory-augmented AI agents using modular architecture (extraction, update, retrieval, response). Inspired by MemFactory (arxiv:2603.29493) - unified training/inference framework for agent memory with RL-driven policy optimization (GRPO). Use when building long-term AI agents, memory management systems, or implementing Memory-R1/RMM/MemAgent paradigms. Keywords: agent memory, memory-augmented LLM, MemFactory, Memory-R1, memory lifecycle, GRPO, memory extraction, memory ret...

Methodology for coordinating multiple AI agents in autonomous research workflows. Covers parallel agent orchestration with diverse initialization, shared communication forums, independent experimentation with shared knowledge, cross-domain generalization testing, reward hacking detection, and the taste-vs-volume tradeoff. Use when: designing multi-agent research systems, orchestrating parallel AI experimentation, building autonomous discovery pipelines, or evaluating automated research qualit...

High-fidelity spatio-temporal mathematical models of Alzheimer's disease progression using 3D brain geometries and network-based connectome models, validated against PET-SUVR imaging data. Activation triggers: Alzheimer's disease, brain network modeling, protein propagation, tau pathology, amyloid-beta, PET-SUVR, computational neurodegeneration.

Two-stage interpretation of attention as in-context empirical Bayes inference via particle dynamics with posterior mean recovery guarantees

注意力残差（AttnRes）方法论。改进 Transformer 注意力机制的残差连接。 提升模型性能和训练稳定性。 触发词：注意力残差、AttnRes、注意力机制、残差连接、Transformer优化、 attention residuals, attention mechanism, residual connection。

Mechanistic explanation of attention sink phenomenon. Variance discrepancy amplified by FFN super neurons. Head-wise RMSNorm fixes it. Based on arXiv 2605.06611.

Gradient descent-based physics-constrained Jacobian Fuzzy Cognitive Map (FCM) with attractor dynamics, residual memory, and BPTT. Uses Newton's method for fixed point attractor finding with adaptive landscape manipulation. Triggers: attractor FCM, fuzzy cognitive map gradient, FCM attractor dynamics, physics-constrained FCM, Jacobian FCM.

Attractor Models for language and reasoning — backbone proposes output embeddings, attractor module refines them by solving for fixed point via implicit differentiation. Constant memory for effective depth, adaptive iteration count, equilibrium internalization phenomenon. Outperforms standard and looped Transformers across language modeling and challenging reasoning tasks (Sudoku-Extreme 91.4%, Maze-Hard 93.1% with 27M params). Use when designing recurrent/iterative refinement architectures, ...

AVSD (Adaptive-View Self-Distillation) methodology from arXiv:2605.20643 (May 2026). Multi-view self-distillation separating consensus from view-specific residual signals for token-level supervision in LLM self-training. Use when: multi-view distillation, self-distillation with privileged information, token-level supervision, consensus-residual decomposition, LLM reasoning improvement.

Empirical scaling laws in balanced networks with conductance-based synapses. Shows that conductance-based synapses + spike time correlations together produce realistic membrane potential variability — neither alone suffices. Activation: balanced network scaling, conductance synapse, membrane variability, spike time correlation, current-based synapse.

Bayes-consistent orchestration patterns for multi-agent AI systems. Use when designing, implementing, or analyzing agentic AI systems that compose multiple agents, tools, or reasoning steps.

Agentic framework for online system operations with Flexible Skill Arrangement. Use when designing LLM-based agents for system O&M, IT operations, DevOps automation, alert management, root cause analysis, or self-evolving agent systems. Activation triggers: system operations, O&M automation, agentic operations, skill arrangement, alert root cause analysis, self-evolving agents, release interception, proactive inspection, KuaiShou operations.

Compute-matched scaling analysis of byte-level modeling revealing context fragility disparity between MDM and AR paradigms, with structural bias recommendations for modality-agnostic designs.

Byzantine-resilient consensus via active reputation learning methodology. Core idea: embed active reputation learning into the consensus loop, where agents evaluate neighbor behaviors using outlier-robust loss functions and historical information, constructing reputation vectors on a probability simplex. This creates a learning-control co-design dual objective: improved consensus enhances Byzantine identifiability, while refined reputations improve consensus. Applicable to distributed systems...

Two-site cavity method for analyzing large nonlinear recurrent neural networks. Derives linear equivalence of nonlinear RNNs, computes full covariance matrices for specific quenched realizations, and separates Gaussian from non-Gaussian contributions in recurrent network dynamics. Use when analyzing: (1) high-dimensional RNN covariance structure, (2) nonlinear-to-linear network equivalence, (3) cavity method applications to neural dynamics, (4) quenched disorder in recurrent networks.

Extended Sompolinsky-Crisanti-Sommers (SCS) theory for two-population Excitatory-Inhibitory networks with target-specific inhibition. DMFT derivation of phase diagrams showing quiescence, asynchronous chaos, persistent activity, structured chaos, and coherent oscillations. Shows target-specific inhibition determines which collective instability dominates. Activation: SCS E/I theory, DMFT neural networks, chaos-synchrony transition, E/I balance, neural phase diagram, 兴奋抑制网络混沌.

Link microscopic chaos in recurrent neural networks to macroscopic geometry of neural representations using kernel methods and dynamical mean-field theory. Chaotic dynamics act as intrinsic regularizer enhancing generalization while preserving expressivity.

CLANE - 在神经形态硬件（Intel Loihi 2）上从事件相机实现动作的持续学习。首个端到端部署的神经形态持续学习系统，结合脉冲 2D CNN 和 CLP-SNN 学习头，实现 100x 能量降低和 16x 延迟降低。

Token optimization methodology for CLI coding agents (Claude Code, OpenCode, Gemini CLI, Codex). Reduces overhead 70%→35%, effectively 2-3x capacity increase. Covers CLAUDE.md bloat, session history, plugin hooks, cache expiry, skill/MCP overload, extended thinking, and wasted output.

Coding agents in social sciences research methodology — using AI coding agents to automate data analysis, simulation, and empirical research in economics, political science, and sociology. Covers reproducibility, agent reliability, and domain-specific challenges.

Complex-Valued Kuramoto Networks control framework - unified control-theoretic approach for synchronization in coupled oscillator networks via complex state space embedding. Activation: Kuramoto, coupled oscillators, synchronization control, phase dynamics, complex-valued control.

Cross-species computational modeling of the auditory periphery using 1-D nonlinear cochlear transmission-line models adapted across human, mouse, and gerbil. Covers species-specific anatomical/physiological parameterization, BM mechanics, OHC deficits, and cochlear synaptopathy simulation.

Causal framework for studying multilingual brain-model alignment using targeted "computational lesions" in multilingual LLMs. Zero out parameters to separate shared vs language-specific brain processing. Use when: multilingual LLM analysis, brain-model alignment, fMRI encoding studies, computational lesions, cross-lingual neuroscience, language processing in brain. Trigger: computational lesion, multilingual brain alignment, language-specific processing, fMRI encoding models, shared backbone,...

多语言语言模型计算性损伤方法论。通过因果干预分离共享和语言特异性脑对齐模式，揭示语言理解的神经基础。适用于多语言神经科学、语言模型解释性、跨语言脑对齐。触发词：计算损伤、多语言、脑对齐、因果干预、共享表征。

Concept-Reasoning Expansion (CoRE) methodology for continual learning in medical imaging. Maps image tokens to structured concept libraries, simulating clinical reasoning to guide interpretable expert routing and demand-based model growth. Prevents catastrophic forgetting while avoiding redundant parameter expansion. Use when implementing continual learning for medical image analysis, sequential task adaptation, or scenarios requiring model growth without forgetting.

Critical analysis of current agentic memory systems showing they implement lookup, not true memory. Argues treating lookup as memory is a category error with consequences for agent capability, long-term learning, and security. Distinguishes retrieval-by-similarity from weight-based memory's generalization-by-composition. Activation: agentic memory critique, true memory vs lookup, weight-based memory, retrieval generalization, agent memory design.

Convex Hybrid Modeling methodology using operator theory for process control and systems engineering. Formulates convex learning problems that combine model interpretability with system identification efficiency. Covers three settings: (1) regularization around a reference model, (2) restriction on interpretable subspaces, (3) kernel-based mixture models on interpretable manifolds. Use when: building interpretable control models, combining physics-based and data-driven modeling, designing hyb...

CORNN凸优化递归神经网络方法论。将RNN训练转化为凸优化问题，训练速度比传统方法快100倍，支持百万参数RNN在标准计算机上亚分钟级训练。适用于大规模神经记录实时建模、神经动力学推断、吸引子结构恢复。触发词：RNN训练、凸优化、神经动力学、实时建模、数据约束RNN、convex optimization、recurrent neural network、neural dynamics inference。

Simulation-based reverse engineering methodology for analyzing how cortical microcircuits optimize information flux (mutual information between successive network states). Use when: (1) studying information-theoretic properties of recurrent neural circuits, (2) analyzing the role of embedding networks in cortical microcolumns, (3) investigating recurrence resonance and entropy-driven dynamics, (4) designing reservoir computing systems with optimal information processing.

Simulation-based reverse engineering study of cortical microcircuit information flux. Analyzes whether cortical microcircuits are optimized for information flux in recurrent networks. Use when: studying cortical circuit optimization, information theory in neural networks, reverse engineering brain circuits, analyzing mutual information between network states, comparing biological vs artificial neural circuit architectures.

Cortico-cerebellar modularity as an architectural inductive bias for efficient temporal learning — CB-RNN architecture showing cerebellar-inspired feedforward modules drive learning efficiency while cortical recurrent cores act as fixed reservoirs.

Cortico-cerebellar modular RNN architecture methodology. Augments RNNs with cerebellar-inspired feedforward modules for efficient temporal learning. The cortical RNN acts as a fixed reservoir while the cerebellar module drives learning efficiency. Applicable to temporal sequence learning, neural network architecture design, and brain-inspired AI systems. Activation: cortico-cerebellar, cerebellar RNN, CB-RNN, cortical-cerebellar, modular RNN, temporal learning architecture, brain-inspired RNN...

曲率感知的非凸优化方法论。核心思想：将二阶几何信息（曲率/海森矩阵特征值）显式纳入优化目标或约束，帮助逃逸鞍点、加速收敛。触发词：非凸优化、鞍点逃逸、曲率感知、CRGD、信赖域、二阶方法、Hessian-aware、escape saddle point、non-convex optimization。

Methodology for auditing and distilling deep reinforcement learning hedging policies into interpretable symbolic formulas. Includes framework for analyzing delta corrections relative to Black-Scholes, symbolic regression distillation, and regime fragility stress-testing. Use when analyzing neural hedging strategies, quantitative risk management, options hedging with RL, or making black-box financial AI auditable.

DeferMem methodology from arXiv:2605.22411 (May 2026). Long-term memory QA framework using query-time evidence distillation via RL (DistillPO): high-recall candidate retrieval + query-conditioned evidence rewriting with decomposed-and-gated reward. Use when: long-term memory QA, RAG with long conversations, RL-based evidence distillation, memory systems for LLM agents.

DAOEF framework for scaling multi-agent edge systems beyond 100 agents without synergistic collapse. Three co-designed mechanisms: differential neural caching (delta-aware activation reuse), criticality-based action space pruning (O(n log n) coordination), and learned hardware affinity matching (GPU/CPU/NPU task routing). Activation: multi-agent edge orchestration, synergistic collapse, MADDPG scaling, differential caching, action space pruning, hardware affinity, edge computing latency, visi...

Stochastic Density-Driven Optimal Control (D²OC) for multi-agent systems. Decentralized non-uniform area coverage using Wasserstein distance minimization with convergence guarantees for stochastic LTI dynamics. Activation: density-driven control, multi-agent coverage, Wasserstein distance, optimal transport, swarm robotics.

Stochastic Density-Driven Optimal Control (D²OC) for multi-agent systems. A rigorous Lagrangian framework for decentralized non-uniform area coverage using Wasserstein distance minimization. Use for: multi-agent coverage control, swarm robotics, distributed optimization, stochastic MPC, area coverage missions with spatial priority. Activation: D2OC, density-driven control, multi-agent coverage, swarm control, Wasserstein distance control.

Digital twin-based consensus control for multi-agent cyber-physical systems under noisy perception and input failures. Combines digital twin modeling with lag consensus protocols for robust distributed coordination. Use when: (1) Designing multi-agent CPS coordination protocols, (2) Analyzing consensus under noisy digital twin perception, (3) Building fault-tolerant distributed control systems, (4) Studying second-order lag consensus in stochastic networks, (5) Modeling physical-digital twin ...