AI & ML

griddynamics Bundle

To orchestrate parallel coding-agent farms (Claude, Codex, Copilot, Gemini, etc.) on isolated git worktrees.

griddynamics Bundle

To author, adapt, review, and validate prompts (skills, agents, workflows, rules, etc.) with brief, contracts, and a validation pack.

wenjunduan Skill

augment-context-engine MCP 工具的调度说明. PACE plan / design 阶段做项目代码语义检索, 找"是否已经实现过"或"相关代码在哪里". 避免重复造轮子.

wenjunduan Bundle

经项目 MCP 能力服务器只读线上业务数据与流程状态。需要查用户、部门、字典、审批或报表实况时触发；不生成也不修改代码。

wenjunduan Skill

Bugfix 路径的结构化问题流程: report → analyze → fix-note 三件套档案, 落在 sprints/{slug}/. 把"偏轻"的 Bugfix 升级成可复现 / 可追溯 / 可复盘. 借 CodeStable issue 实体, 适配 agent-in-loop.

wenjunduan Skill

Bugfix 路径的结构化问题流程: report → analyze → fix-note 三件套档案, 落在 sprints/{slug}/. 把"偏轻"的 Bugfix 升级成可复现 / 可追溯 / 可复盘. 借 CodeStable issue 实体, 适配 agent-in-loop.

wenjunduan Skill

会话记忆固化 skill (v9.8.0 新). 会话结束/中途, agent 自己总结本会话增量写进 .ai_state (_index.md 当前状态 + sprints/{slug}/session-log.md), 免去用户每次手动描述一堆让它存. 手动 /checkpoint 触发. 与 compact-snapshot hook 互补 (hook 机械兜底, skill 做需推理的总结).

wenjunduan Skill

维护 .ai_state/requirements/ 长效需求档 (软件要素: 为什么要这个能力), Codex. 原始用户故事 + 当时权衡, 独立于会演化的 design.md —— 代码烂掉时弃码重生的逃生通道. 借 CodeStable requirements 实体, 适配 agent-in-loop: 需求落盘可召回, 隐知识不丢.

wenjunduan Skill

维护 .ai_state/requirements/ 长效需求档 (软件要素: 为什么要这个能力). 原始用户故事 + 当时权衡, 独立于会演化的 design.md —— 代码烂掉时弃码重生的逃生通道. 借 CodeStable requirements 实体, 适配 agent-in-loop: 需求落盘可召回, 隐知识不丢.

bejeweledme Bundle

Run, design, review, or debug multi-model councils using independent first answers, peer critique, and synthesis traceable to intermediate outputs.

bejeweledme Bundle

Design, debug, or review LLM and agent evaluations: task datasets, tool/state grading, calibrated judges, traces, candidate selection, repeated trials, and release gates. Use when evaluating model, prompt, tool, or workflow changes; domain owners define the behavior being measured.

bejeweledme Bundle

Use when designing, implementing, reviewing, debugging, evaluating, or operating retrieval-augmented generation systems. Trigger for ingestion, OCR, chunking, metadata and provenance, embeddings, sparse/dense/hybrid retrieval, ANN, reranking, context assembly, citations, abstention, GraphRAG, agentic retrieval, index versioning, or stage-by-stage RAG diagnosis.

bejeweledme Bundle

Design, review and debug predictive/classical ML systems: whether ML is needed, build or buy, metrics, data, validation, baselines, lifecycle pipelines, predictive serving, benchmark claims, placement, A/B versus canary, monitoring, drift, degradation and ownership. Includes ML system design interviews. Modality-specific CV/LLM work and neural execution have dedicated owners.

bejeweledme Bundle

Use when designing, reviewing, or evolving an LLM product as a system. Trigger for prompt versus RAG versus agent versus adaptation decisions, model or provider routing, prompt and context architecture, token/cost/latency budgets, degradation and fallbacks, tenant isolation, and versioned LLM releases. Not for detailed retrieval, agent-loop, model-training, eval-harness, security-test, or GPU-runtime implementation.

bejeweledme Bundle

Build, diagnose, and operate shared AI infrastructure for predictive and GenAI serving (KServe, Triton, vLLM Production Stack), accelerator capacity, replica controllers, model registry releases, training-job placement, and edge fleet rollout. Runtime tuning and training update semantics stay with their execution owners.

bejeweledme Bundle

Adapt a primary informational artifact for technical, nontechnical, executive, or execution readers. Use alongside the skill that owns the source, research, explanation, decision, edit, or deliverable; it is a reader-model companion rather than a primary content operation.

bejeweledme Bundle

Use when threat-modeling, hardening, or authorized testing of LLM, RAG, and agent systems the user controls. Trigger for prompt or indirect injection, RAG ACL or tenant leakage, tool/data-flow misuse, excessive agency, guardrails, sandboxing, least privilege, service identities, self-test/red-team planning, false-refusal evaluation, security evidence, stop conditions, and regression loops.

bejeweledme Bundle

Explain a concept, mechanism, system, or distinction so the reader builds a correct and usable mental model.

bejeweledme Bundle

Diagnose, optimize, scale, or recover neural training runs: OOM, slow steps, mixed precision, gradient accumulation, DDP/FSDP, collectives, and checkpoint consistency. Owns training execution; model/objective choices belong to modeling skills and serving execution to inference skills.

bejeweledme Bundle

Use when designing, reviewing, validating, or debugging computer-vision evaluation. Trigger for product-linked and task metrics, thresholds, calibration, class/scale/source slices, visual error analysis, imbalance, robustness, corruptions and OOD, regression suites, human review, and release gates. Not for training procedure, annotation operations, or generic LLM and agent evals.

bejeweledme Bundle

Diagnose and optimize self-hosted LLM runtime latency, goodput, cost, and capacity through prefill/decode profiling, KV cache, batching, admission, quantization, and inference parallelism. Use for execution bottlenecks and runtime scaling signals; replica provisioning and fleet placement belong to ai-platform-llmops.

bejeweledme Bundle

Use when choosing or adapting a text model. Trigger for rules, TF-IDF or BM25 baselines, encoder or reranker selection, embeddings, BERT-style discriminative models, seq2seq or decoder LLMs, tokenizer changes, continued pretraining, SFT, PEFT/LoRA/QLoRA, preference optimization, distillation, or training-data and adaptation evaluation. Not for LLM product topology, detailed RAG, agent loops, eval harnesses, or serving bottlenecks.

bejeweledme Bundle

Design, implement, or remediate general application security controls: code-linked threat models, authentication and recovery, sessions, authorization and tenant isolation, OAuth/JWT, browser/API/uploads, secrets and data protection, secure failure, dependency policy, or ASVS implementation. Independent assessments belong to security-review; specialist LLM/RAG/agent threat modeling, security design/review, hardening and authorized testing belong to genai-security-testing.

bejeweledme Bundle

Use when designing, collecting, labeling, validating, versioning, or improving computer-vision datasets. Trigger for annotation schemas and QA, source/camera/scene coverage, class and object balance, duplicates, grouped splits, temporal or spatial leakage, augmentation, active learning, auto-labeling, synthetic data, provenance, and production data feedback. Not for choosing model architectures or declaring model quality.

bejeweledme Bundle

Use when choosing, training, adapting, or debugging computer-vision models for image classification, metric learning and retrieval, detection, segmentation, OCR and documents, multimodal vision, or video and tracking. Covers classical feature-based baselines, deep and foundation models, transfer learning, fine-tuning, PEFT, self-supervised learning, objectives, and reproducible experiments. Not for dataset governance, release metrics, or serving optimization.

bejeweledme Bundle

Use when profiling, deploying, or optimizing computer-vision inference pipelines for latency, throughput, memory, cost, FPS, or device constraints. Trigger for decode/preprocess/model/postprocess/index stages, tiling, batching, export and runtime compatibility, precision and quantization, streaming video, frame timing, and capacity planning. Not for model training or generic service and SRE ownership.

ashermahonin Bundle

Inspect an existing repository, prepare an empty project, or map a migration before implementation begins. Use when an agent first enters an unfamiliar project or when a new project needs its stack, platform, test commands, documentation entrypoint, project memory, and next route established. Preserve existing agent configuration and do not scaffold a stack without confirmed intent.

ashermahonin Bundle

Break approved product scope and architecture artifacts into epics, user stories, implementation tasks, dependency graph, critical path, ownership map, validation packs, merge order, and parallelization plan. Use after requirements and architecture are approved and before assigning implementation agents, opening parallel work, or writing artifacts `24-epics.md` through `29-agent-role-matrix.md`.

ashermahonin Bundle

Clarify and scope incoming product or engineering requests before research, architecture, or coding starts. Use when the request is broad, emotionally urgent, ambiguous, risky, multi-phase, missing success criteria, missing constraints, missing non-goals, or likely to create rework if the agent starts implementation too early.

ashermahonin Bundle

Review an LLM, RAG, embedding, fine-tuning, or model-routed feature against the current OWASP LLM risks. Use when untrusted content enters prompts or retrieval, model output affects another system, sensitive data is exposed to a provider, or cost and agency can grow. Produce a data-flow threat model, abuse-case evidence, controls, residual risk, and release recommendation.

ashermahonin Bundle

Create or update a reusable Agent Skill in the open SKILL.md format. Use for project, team, domain, or tool workflows that need a precise trigger and repeatable procedure. Add references, scripts, assets, client metadata, or routing only when the skill needs them; do not create a new skill for a one-off instruction or a narrow variant already owned elsewhere.

ashermahonin Bundle

Check project dependencies, base images, operating-system packages, infrastructure providers, plugins, and agent tools against current vulnerability and active-exploitation sources. Use before release, after dependency changes, during incident review, and on a schedule for long-lived services. Record source, timestamp, reachability, exploitability, mitigation, owner, and release impact.

ashermahonin Bundle

Convert a design choice, architecture option, market assumption, or feature bet into a falsifiable hypothesis with a measurable kill criterion, the cheapest useful test to settle it, and a clear record of what was supported, disproved, or deferred. Use whenever the agent is tempted to commit to a stack, a pattern, a vendor, a model, a UX flow, or a scope without evidence, and as the core supporting skill for ai-pdlc and architecture-review.

ashermahonin Bundle

Repair a measured failure in agent routing, instructions, tool contracts, context, memory, or evaluation, then verify the repair and resume the original task. Use after a reproducible failed check, route mismatch, tool misuse, regression, or specific user correction. Do not activate from vague dissatisfaction or use it to increase permissions or autonomy silently.

ashermahonin Bundle

Review a tool-using or multi-agent system against the current OWASP Agentic Application risks. Use when an LLM can read, write, execute tools, retain memory, act under an identity, or delegate work. Produce risk status, autonomy and permission boundaries, tool inventory, abuse-case evidence, recovery controls, and a release recommendation.

ashermahonin Bundle

Agent harness architect