seacen

Builds and maintains a domain knowledge pack for the data-analysis Skill the way a human analyst would — studying the raw material page by page, synthesising one Understanding Handbook, then deriving the full set of pack files (triggers, README, framework, glossary, computation-reference, data-spec, context-notes, insight-extensions, business-rules, plus an optional design-system slot) from that handbook plus the data profile. It is a research workflow, not an extraction pipeline — the agent reads decks and data visually and thoroughly, writes a handbook, then derives. Use when the user wants to create a domain pack for a new analysis area, refresh an existing pack after new material, new data, or an SME correction, or turn training decks and a sample analysis and a data file into pack files. Triggers include "build/refresh a domain pack", "把这些材料研读成 domain pack", "先写理解手册再派生 pack", and "刷新一下 domain pack".

Runs structured AI-native data analysis on tabular datasets (Excel / CSV / Parquet / DataFrame). Pluggable domain architecture — Step 0 reads `references/resolver.md` (a domain-neutral scoring mechanism) plus each registered domain's own `domains/<slug>/triggers.md` to detect which analytical domain applies; all domain knowledge and trigger words live in the domain packs, never in this Skill or the resolver. Drives every analysis through a deterministic 0+9-step workflow (Domain Detect → Profile → Signal Scan → Clarify → Frame → Hypothesize → Verify → Decompose → Synthesize → Follow-up). Hard rules against hallucination — LLM never computes numbers, every claim cites a verdict_id, state persists to JSON. Auto-produces default insights without asking the user what to analyze. Use whenever the user uploads structured data, asks for "analysis" / "insights" / "default insights" / "what does this say" / "默认洞察" / "分析这份数据", asks "why is X declining / growing / changing", or any structured-data analytical query that

Generate a complete HTML-slide Design System markdown file from a reference URL. Studies the page's color palette, typography, logo, and aesthetic mood, then writes a Design System that bakes in proven engineering rules (1280×720 fit contract, three-layer overflow safety net, single-absorber rule, mobile inline-flex trap catch-all, flip-card mobile fix, logo-as-SVG-symbol-with-currentColor pattern, 12px readability floor) while letting the brand identity drive the visuals. Use this whenever the user gives you a URL and asks to "build a design system from this site," "extract a design system from", "make slides like this brand," "skin slides to match this site's visual language," or wants to author HTML slides in the visual language of any specific website. Also use when the user names a brand site as a slide-deck reference, even if they don't say "design system" explicitly.

Bring humanity's best thinking to bear BEFORE improvising an answer. When the user seeks counsel or takes on a substantive task that rewards real thinking, planning, or design — rather than a quick fact or a mechanical edit — this skill first reaches for the two or three most fitting proven frameworks or mental models and the best books on them, distills their essence, recommends them for deeper reading, and only THEN solves the problem, visibly weaving those lenses together. Use proactively whenever someone is deciding, planning, designing, persuading, writing something that matters, resolving a tension, leading or building, or feeling stuck and unsure how to proceed — in work or in life — even when they never say the words "framework" or "book." The harder, higher-stakes, or more open-ended the task, the more this applies. Skip only genuinely trivial one-step tasks (quick fact lookups, simple mechanical edits, single commands).

Strip the "AI flavor" from writing — the telltale patterns of machine-generated prose (evaluative inflation, mechanical antithesis like "it's not just X, it's Y", rule-of-three padding, hollow hedging, narrator scaffolding, jargon and corporate filler, translationese) — so the text reads like a real person wrote it, without touching terms, quotes, code, numbers, or an author's deliberate voice. Multilingual; Chinese and English are first-class. Apply this proactively as the finishing pass on ANY task that produces prose a person will read or publish: articles, blog posts, marketing and social copy (Xiaohongshu, WeChat public-account posts), slide decks and PPT, emails, reports, documentation, announcements, product and landing copy, newsletters, scripts. Run it before delivering generated writing even when the user never mentions "AI flavor" — a reader should never be handed AI-sounding text, and cleaning it only after they complain is already too late. Also use when explicitly asked to humanize, de-slop, rem