Save Mem
Preserve useful project knowledge through StrictDoc, Magic Context, and available memory tools, including overlapping storage where useful. Maintain readable current norms in StrictDoc; update them when a genuine decision change is clear and briefly inform the user, without routinely churning the canon. Ordinary memories and explanatory docs can evolve more freely. Favor concise phrases, lists, t…
Accel Skill Template
深度学习平台初始化完成后的落盘规则与模板：把验证过的平台知识固化进固定通用名的 accel 三件套与宿主机层。Use when an accelerator platform setup just finished and needs to be distilled, when specializing or updating the accel-* trio's content for the local platform, or when asked to 落盘/沉淀/固化/特化/回流. 三件套正文随平台改写、名称永不变；宿主机层骨架模板在这里。
Load Mem
Understand a project through its existing memory when starting or resuming work, or when past decisions matter. Use StrictDoc docs/, Magic Context, claude-mem, and project evidence as complementary, overlapping sources. Identify current authoritative norms separately from history, recover relevant context, and surface meaningful contradictions. Choose the depth and retrieval path that suit the ta…
Personal Ui Taste
个人前端审美、功能驱动的信息组织与持续演化的 UI 设计指南。Use when designing, building, polishing or reviewing this user's frontend UI, dashboards, charts, forms or admin panels; choosing task-fit interactions such as joint rankings, rollups and drill-down; following 我的品味/个人审美; or recording approved feedback. Applies scoped preferences and interaction patterns across projects. Not for backend-only work.
Writing Skill
把反复出现的经验、对话结论或踩坑固化成可移植的 skill，或在既有 skill 上增删改查。Use when asked to 写个skill/固化成skill/沉淀成skill/把这个流程做成skill，when a lesson keeps being retaught across sessions or machines，or when reviewing or trimming an existing skill. 判定"该进 skill 还是记忆"是第一步；算力平台四层落盘走 accel-skill-template，中文行文风格走 readable-docs，本机事实去情境化后回流前跑 platform-environment-skill-audit。
Plan Brief
Turn a complex, thought-through requirement into a detailed execution plan document plus a self-contained launcher prompt for a fresh agent session, and later review that session's report. Use when the user wants a plan before execution, wants execution handed to a separate session, or mentions 执行计划 / 计划文档 / 新会话执行 / 交接 prompt / 审查 report; also when reviewing a plan's execution report or dispatchi…
Exp Discuss
Talk a research idea through with the researcher before anything is designed or run. Use when a researcher brings a fuzzy idea, hypothesis, or direction ("我在想……", "你觉得这个方向……") and wants to think together rather than execute. Explore it through evidence-informed rounds of frontier questions, surface overlooked details, and sharpen what the researcher actually wants to observe; stop when both sides…
Platform Environment Skill Audit
Audit accelerator environment skills and platform/cluster docs for personal repository, user, Conda, credential, and validation-artifact leakage; use when republishing CUDA, MUSA, PPU, ROCm, or other server environment guidance for multiple users.
Accel Pytorch Python
Python/wheel 层：为加速平台的 PyTorch 栈创建或修复 Python 环境——厂商源 vs 公共源纪律、版本匹配矩阵、环境变量、安装流程、验证与常见失败，适用任意厂商平台。Use when creating or repairing a Python environment for accelerator-platform PyTorch (vendor torch backend, torchvision/torchaudio, platform triton, vendor comm libs) on any platform. 正文为当前已验证实例（现 MUSA/摩尔线程源）；非该平台首次初始化时先按头部「如何特化」改写正文。触发词：装 torch、wheel 源、版本匹配、pip 源、依赖冲突。
Exp Batch
Execute one bounded research experiment branch in the current session with ponytail-for-experiments taste. Use when explicitly requested, or for a single coherent batch of runs with known configuration (one method × dataset × budget, or one ablation branch of a campaign). First check the branch was not already pruned or answered, grill only unresolved config decisions, preflight environment and r…
Accel Pytorch Code Porting
代码层：在本机加速平台上编写/移植/调试/运行 PyTorch 应用代码——CUDA→平台 API 映射、设备探测抽象、AMP 与数值稳定、分布式、第三方包兼容、迁移审计清单，适用任意厂商平台。Use ONLY when writing, porting, reviewing, debugging, or running application code for the local accelerator platform. 正文为当前已验证实例（现 MUSA/CUDA→MUSA 映射表）；非该平台首次初始化时先按头部「如何特化」改写正文。Python 环境与包装机用 accel-pytorch-python，系统层安装用 accel-platform-install。触发词：移植、torch.cuda 改写、设备探测、is_available、分布式后端、算子兼容。
Editable Vector Slides
将自包含的 SVG 科研流程图、架构图转换为保留真实文字的矢量 PDF，以及原生可编辑的 PowerPoint PPTX。Use when asked to convert SVG/PDF figures to editable PDF/PPTX, preserve editable text and vector shapes, or avoid a slide that is only a screenshot. Includes a tested SVG→PDF→text-template→native DrawingML workflow; the PPTX helper targets flattened path-based SVGs, not arbitrary SVG artwork.
Exp Campaign
Design a staged, prunable research experiment campaign for a researcher's idea, hypothesis, or optimization direction, with ponytail-for-experiments taste. Use when explicitly requested, or when a researcher proposes a direction that needs many experiments, ablation branches, or cross-factor sweeps. Observe existing results, grill the researcher to shared understanding, enumerate the maximal expe…
Kb Accel Sdk
公共知识库：各大加速计算平台（CUDA/MUSA/Ascend CANN/ROCm/阿里PPU/沐曦MACA）的 SDK 组成、安装方式、triton 与 tilelang 生态、vllm 与 sglang 推理框架生态速查。Use when working on any accelerator server, choosing SDK/wheel/inference-framework versions, or when a platform-specific package/behavior surprises you. 知识类 skill：可被覆盖刷新，不做本机特化；新事实经 writing-skill 流程回流仓库。触发词：SDK 组成、装 CANN、torch_npu、vllm 昇腾、sglang 国产卡、triton fork、tilelang、PPU。
Accel Init
深度学习算力平台初始化主入口：探测（或由用户直说）加速器平台与现状，搭 Python 双栈底座（miniconda+uv+清华源），按平台调研 wheel 矩阵/fork/生态兼容，出方案、执行、走验证阶梯，最后交棒 accel-skill-template 落盘。Use when setting up a fresh compute server for deep learning（CUDA/MUSA/Ascend/ROCm 等驱动、SDK、PyTorch、triton、transformers 生态），when asked 初始化算力环境/装平台栈/新服务器配置深度学习环境，or when a platform stack needs repair and re-validation. 只做 Python 环境不涉平台时直接按常规流程；落盘模板细节看 accel-skill-temp…
Exp Probe
Run one-off research spot experiments with minimal ceremony. Use when explicitly requested, or for a missing datapoint, a rerun after failure (OOM, crash), a sanity check of one configuration, a minimal bug reproduction, or a cheap measurement that a campaign needs before deciding. At most one focused question; run, record, report in one line. Escalate to exp-batch if the spot check turns into a…
Quick Do
Execute small, well-understood changes directly in the current session with minimal ceremony and ponytail taste. Use when explicitly requested, or automatically for localized fixes, mechanical edits, config tweaks, and short scripts with clear behavior and a known verification path. No plan documents or grilling; do not add tests by default. Prefer steady-do when design, compatibility, or accepta…
Readable Docs
Write and edit human-facing project technical text for scanning, wherever it lives — README, specs, design notes, plans, reports, runbooks, changelogs, and project memory, not only docs/ or StrictDoc. Use whenever creating or revising these records, without waiting for an explicit style request; also for 说人话、去 AI 味、精简文档、短语分点、PPT 式表达. Prefer Chinese phrases, lists, and compact comparison tables wh…
Accel Platform Install
系统层：本机加速平台栈（驱动/SDK/工具链/通信库）的安装、升级、排障与验证阶梯，适用任意厂商平台（CUDA/MUSA/Ascend/ROCm 等）。Use when installing, upgrading, or troubleshooting the local accelerator platform stack regardless of vendor — user-space installs, Python compatibility, dynamic libraries, validation ladder. 正文为当前已验证实例（现 MUSA/Ubuntu 22.04）；非该平台首次初始化时先按头部「如何特化」改写正文。触发词：装驱动、装SDK、工具链、平台栈、用户态安装、验证阶梯。
Steady Do
Develop features in an existing project within the current session with ponytail taste. Use when explicitly requested, or automatically for ordinary feature, plugin, extension, and cross-module work that needs project understanding and clarification before implementation. Survey the real flow, grill only unresolved decisions, then implement and verify here without plan or handoff documents. Prefe…
Migrate Mem
Establish useful memory for a project that already has development history, or improve an incomplete memory setup. Survey the project and its existing evidence, then initialize or enrich human-readable StrictDoc docs/ with current norms and historical context. Use Magic Context and other available memory alongside it where useful. No previous workflow, legacy folder convention, or pre-existing St…
Model Datasets
模型与数据集的统一下载、落盘与加载约定：平台（HF / ModelScope）只是下载手段，本地一律 /data/models/<org>__<name>、/data/datasets/<org>__<name>，只用绝对路径喂给 transformers / datasets。含新机器的 /data 磁盘准备规则（无 /data 时初始化或软链）。Use when downloading or organizing models/datasets, deciding where to store them, using hf / modelscope CLI, or loading with from_pretrained / load_dataset. 触发词：模型放哪、数据集放哪、下载模型、from_pretrained 路径、/data、hf、modelscope、大文件放哪。