lecutu/DeepSeek-PPT-skill

README 摘要

ppt-reflex · Correct PowerPoint without a vision model ppt-reflex is an engine that lets LLMs without vision produce correct PowerPoint decks — plus a DeepSeek Harness (DSH) plugin that adds live preview and a feedback loop. Blind LLMs — DeepSeek first, but any text-only model — cannot see the .pptx they generate. The usual workaround makes the model guess harder: it writes coordinates, crosses its fingers, and hopes. ppt-reflex inverts that pipeline. The AI never writes a coordinate and never guesses how anything looks. It declares layout intent — an archetype, parameters, a recipe, a decoration skin — and a deterministic constraint-solving engine computes every coordinate, measures every glyph, and reports back in text. Three things replace vision: 1. Real font measurement — PIL glyph-level FreeType metrics (CJK-aware, East Asian Width-aware) compute how much space text actually needs before render. 2. Structured diagnostics — every build returns machine-readable issues with phase , kind , severity , message , and options the AI can act on. 3. Three-tier ASCII feedback — an L0 structure map, an L1 element map, and an L2 numeric text table give the model a "picture" it can actuall…

分类