YunchengLiu
- 11 skills
- 0 followers
- 23 hours ago last updated
- ▌ Pytorch · yunchengliu bundleGuide modern, reliable PyTorch mechanics for single-machine workloads: modules, datasets, dataloaders, train/eval/inference loops, devices, dtypes, checkpoints, AMP, torch.compile, and basic single-node DDP. Use for PyTorch implementation choices, not general experiment planning or cluster design.
- ▌ Ai4science · yunchengliu bundleGuide stage-aware Python AI and AI4Science experiment work: clarify the experiment stage, evaluation protocol, representation, reproducibility, artifacts, and framework reuse while keeping the current workflow sufficient, clear, maintainable, and repo-context driven.
- ▌ Modern Cpp · yunchengliu bundleChoose and apply modern C++ language and standard-library facilities for clearer interfaces, ownership, lifetime, generic code, and everyday implementation. Use for C++ implementation, review, focused modernization, or target-standard decisions where these choices affect the result.
- ▌ Modern Cmake · yunchengliu bundleChoose and apply modern CMake facilities for target-based builds, usage requirements, source and header ownership, presets, dependencies, testing, and packages. Use when implementing, reviewing, initializing, or modernizing CMake code and build structure.
- ▌ Task Handoff · yunchengliu bundleKeep compact current task state for continuation across conversations, pauses, context compaction, or agent handoff. Use during long or changing tasks when key decisions, progress, and the next action would be costly to reconstruct.
- ▌ Use Subagents · yunchengliu bundleDecide when subagents materially improve a task, dispatch concise neutral assignments, preserve independent review, evaluate outputs, and synthesize evidence without outsourcing judgment. Use for independent review, design or algorithm critique, debugging hypotheses, research analysis, verification, or bounded parallel work.
- ▌ Spec Driven Dev · yunchengliu bundleDiscuss, write, refine, review, and implement repo-aware development specs. Use after a rough requirement or plan exists to settle important design choices, produce a concise execution-ready spec, or implement a supplied spec in gated copilot or explicitly authorized autonomous mode.
- ▌ First Principles · yunchengliu bundleReason from required outcomes, facts, and constraints, and test each part through ablation. Use throughout analysis, planning, design, implementation, problem solving, verification, and review; scale depth to the decision.
- ▌ Planning Clarification · yunchengliu bundleClarify unsettled requests, probe assumptions and design choices, and converge on evidence-backed execution briefs or fresh-context prompts. Use when goals, constraints, or material decisions need working through before execution.
- ▌ Cpp Project Engineering · yunchengliu bundleApply detailed personal C++ project conventions when implementing, refactoring, or reviewing source, comments, tests, errors, logging, CMake, and delivery. Covers how to organize, explain, integrate, and verify a complete change.
- ▌ Personal Markdown Note Writer · yunchengliu bundleTurn provided discussions, notes, excerpts, drafts, and command fragments into long-lived Chinese Markdown personal notes. Use for learning notes and tool/setup notes that preserve settled conclusions, stay independent of chat history, and write to a user-specified output path or chat output.