Claude Skills by SungFeng-Huang

Export official BibTeX for all papers a Heptabase/Obsidian card links to — read the anchor card's card-links, resolve each linked paper card's arxiv ID property, and fetch entries from official sources only (Semantic Scholar / arxiv.org/bibtex / ACL Anthology .bib / OpenReview). Never fabricates an entry: unresolved papers become TODO comments. Use when the user asks to 匯出 BibTeX / 產生 .bib / export references / citations for an overview card, a project card, or a set of paper ids.

Rewrite a Heptabase paper card into the "teaching-style" format — written to onboard the reader configured at config profile.reader — assumed NOT an expert in that paper's subfield: a one-line essence + why-you-should-read framing, a 先備知識 glossary that defines the jargon in plain language, teaching prose that explains the WHY (not just the what), bilingual terms (English kept alongside Chinese), a filled-in 快速摘要, preserved figures, and re-applied colorization. Use when the user asks to rewrit...

Mirror research-cards collections to HackMD (the plugin's third note surface, aimed at sharing/publishing): incremental forward sync from the active backend (heptabase/obsidian/both) into HackMD folders with real note-to-note links, plus OPTIONAL level-2 write-back (config hackmd.write_back) — HackMD-side edits on owner-writable notes merge back into the vault paragraph-by-paragraph; shared-writable notes and two-sided edits stay conflicts. Use when the user asks to sync/publish cards to Hack...

(formerly obsidian-sync) The heptabase segment engine of note-sync: block-level level-2 BIDIRECTIONAL sync between Heptabase and the local .md store (vault) — incremental forward mirror, block-level write-back against a ProseMirror snapshot (all-or-nothing per card), three-way property sync, adoption of hand-made .md files, deletion propagation into .trash/, and a conflict ledger (Sync Conflicts.md). Prefer invoking via note-sync (--mode heptabase); run standalone for cron or engine-level deb...

ONE entry point for the note-surface STAR: local (plain .md) is always the hub, every enabled surface (Heptabase, HackMD, …) syncs over its own local↔surface segment per the config backends list (first = canonical). Converges HackMD write-backs onto Heptabase in the same run and aggregates all segments' conflicts into one report. --mode heptabase|hackmd runs a single segment (obsidian is the accepted legacy alias). Use when the user asks to sync notes / sync everything / 同步筆記、全部同步、跑同步、note sy...

Add or refresh a Hung-Yi-Lee-style teaching 導讀 (narrative intro) at the top of a Heptabase comparison/overview card — weaving the papers the card lists into ONE evolution thread with a punchline, grounded ONLY in the card's own content. Idempotent: re-running updates the existing 導讀 in place instead of duplicating. Use when the user asks to add/update/refresh a 導讀 / 教學導讀 / 導言 / teaching intro on an overview card (Tokenizer/Codec, Spoken LM, Reasoning, Duplex S2S, Audio Understanding, …).

Maintain the study/overview knowledge graph in Heptabase: the hierarchy (📚 root index → topic hubs → comparison cards, Level property = tree depth), the lateral edges (相關主題（橫向連結） mention sections, ↔/→ direction semantics), the knowledge-map whiteboard, hub creation, and hung-yi-lee narrative-act card re-splits when a card nears the 100K cap or its taxonomy stops matching the story. Use when the user asks to 新增/調整 hub、拆卡/重分 overview、加/改 graph 邊 or 橫向連結、調 Level/目錄/知識地圖 whiteboard、或稽核 knowledge...

Maintain ALL Heptabase comparison-overview card families with one skill — tokenizer/codec, spoken-LM/speech-LLM (nesting the auditory 聽覺理解・推理 sub-topic: audio understanding/reasoning/audio-LM 驗收), duplex S2S, 語音生成 (TTS family), ASR (non-LLM), and speech frontend/security. Find newly Tasks-tagged papers an overview is missing, route each to the right topic and act card, author a hung-yi-lee-style section + dimension row + paradigm/synthesis updates, re-sort by arxiv ID, refresh the 導讀, and kee...

Generate and refresh the VISUAL views of a research-project card chain as Obsidian JSON Canvas files: the timeline canvas (git-graph style — log cards as a commit column, the entry card as the HEAD pointer riding beside the newest distilled log with the 📎 backlog stacking above it, colored by origin machine by default) and the context mind map (three decomposition modes: logs = per-log-card subtrees hung on the 前情提要 citation graph; chain = the chain body's H2/H3 sections; story = the researc...

Re-scope a project card CHAIN to its handoff role — 最高指導原則 + 實驗現狀紀錄 + 待辦 handoff — against an external authority layer (paper draft, .private/chapter-plan & evidence-map, or the axis card). Unlike project-card-merge (which FOLDS appends into the body), cleanup DISTILLS writing-phase appends, SUPERSEDES history pile-ups, and POINTERIZES content that now lives in the draft/plan/axis card, while preserving every experiment number, method record, figure, and card-link. Use when the user says 清理卡鏈...

From a project-wise session, find THIS project's Heptabase card and log progress as SELF-CONTAINED log cards — weekly reports to the project lead/mentor: each log event becomes its own card (前情提要 recap that needs no back-reading, what was done, results with explained metrics, what it means, decisions pending — no unexplained abbreviations), and the project chain only gains one human-readable timeline line (📎 date [[card]] one-sentence summary) — the chain stays a readable project timeline fo...

Consolidate a Heptabase research-project card on the Mac into a STANDALONE full report for the project lead (前情提要 + evolution story + current state — readable without session memory): fold the cluster's append-only `📥 cluster 補充/進度` blocks (and any `📝 待補成 paper 級參考` brief) into ONE focused, paper-grade, coherent card — update 現狀 to the latest state, supersede stale info, keep all paper-grade detail + figures + file citations + card-links, and remove the dated append shells. Also folds tail-...

Repair broken card links in Heptabase project + project/progress cards — the dead plain-text `[[card:UUID]]` literals left by a `hb` bridge append (remote cluster) or a bare CLI append, which the heptabase CLI never renders into a real, clickable card-mention. Seals structural header/timeline shapes plus safe inline citations on progress/log cards back into live mentions: a log/progress card's `專案：[[card:…]]` back-ref to its project, a project card's `📎 date [[card:…]]` timeline line to a lo...

Scaffold and drive autonomous research-experiment campaigns: a MISSION.md brief that lives IN the target repo (runs/auto_research/), a persistent experiment ladder (queue.json) and results ledger (ledger.jsonl), hard measurement discipline (significance gating, representative eval slices, one change per rerun), and research-cards integration (project card bootstrap + Op5 gap-analysis pre-read + per-job progress append). Use when the user wants to 開/設定一個研究實驗 campaign for a repo, 接續/continue a ...

Read all unread Scholar Inbox emails from Mac Mail.app (Scholar Alert Digests and Trending Papers), fetch each paper's AI overview from alphaXiv (or fall back to the raw paper), translate to the configured output language (default Traditional Chinese), apply color highlights, and create Heptabase cards. Use when the user asks to process Scholar Inbox papers, clip papers from alphaXiv, or update Heptabase with new research. Invoked with NO paper URL, it instead runs a BACKFILL maintenance pass...

Interactive config wizard for research-cards: create, inspect, or adjust ~/.config/research-cards/config.json by interviewing the user against config.example.json's inline docs. Use when the user asks to set up / configure / initialize research-cards, 建立 config、設定 research-cards、調整設定、初始化、幫我設定、check my config, enable a backend (backends list: local/heptabase/hackmd), change the output language, or turn features on/off.