skills

Windows-first Visio diagram reconstruction workflow for flowcharts, architecture diagrams, and paper-style module figures. Reuses ppt-master style analysis and composition discipline on the front half, but outputs editable Visio .vsdx plus exported .svg and .png through a scene.json to Visio pipeline. Use when the user wants a diagram recreated as editable Visio shapes instead of a pasted screenshot or PPT-only result.

Use when the user wants PPT slide images or imagegen-generated presentation pages from a thesis, graduation thesis, paper, article, literature review, tables, research notes, reference PPT/images, or asks for 答辩PPT图片版, 每页PPT讲什么/画什么, PPT内容框架图, 论文生成PPT, image-2, gpt-image-2, or simplaj imagegen slide visuals.

当用户需要在 LPU/NPU/AI 加速器类项目中编写、开发、重构或修改 pymodel 算子时使用，包括确认计算图、规划 tile/buffer 数据流、实现算子指令流、生成 task bin、插入或检查 set_flag/wait_flag/fence、排查 cmodel sequential 与 full replay 不一致、定位 RTL/UVM replay/GM mismatch。适用于 prefill/decode attention、flash streaming、KV cache、GQA、RoPE、RMSNorm 等复杂算子实现和验证。

Use when revising English theses or papers in AI systems, AI chips, compilers, computer architecture, or systems engineering for contribution framing, abstract/introduction logic, related-work positioning, method/evaluation chapter organization, natural academic prose, reference-thesis structure, bilingual alignment, bibliography/format checks, or text that is AI-like, too implementation-heavy, code-like, vague, or not publication-ready.

Use when revising Chinese theses or papers in AI systems, AI chips, compilers, computer architecture, or systems engineering for contribution framing, abstract/introduction logic, related-work positioning, method/evaluation chapter organization, natural academic Chinese, figure/table prose, reference-thesis structure, bilingual abstracts, bibliography/format checks, or text that is AI-like, too low-level, code-like, obscure, or not human-written.

Use when generating or editing images through image2.0/OpenAI-compatible APIs from Codex, including image2-router, text-to-image, image-to-image, 图生图, sakura imagegen, fhl imagegen, monitor imagegen, ccccapi imagegen, tokenx24 imagegen, super imagegen, riclab imagegen, sub2api imagegen, SAKURA_API_KEY, FHL_API_KEY, MONITOR_API_KEY, CCCCAPI_API_KEY, TOKENX24_API_KEY, SUPER_API_KEY, RICLAB_API_KEY, image2.0, or gpt-image-2.

Use when debugging xwave-supported local hardware waveforms, AXI/APB hangs or timeouts, RTL signal behavior, waveform traces, signal values, or simulation failures that require evidence from xwave.