The READ half of Peter's support loop (formerly support-reply-drafter). Parses Peter's Gmail support + inbound-sales queues, detects unreplied inquiries, identifies each inquirer's intent, and matches it against the **Zynkr platform 知識庫** (platform.zynkr.ai/kb — the single source of truth since 2026-07-15, accessed via the `zynkr` MCP server's search_kb / get_kb_article) to draft a safe reply on the thread. Drafts match the inbound language (zh-TW or EN). When the KB has no confident answer it does NOT guess — it leaves a holding reply with an inline `[[NEEDS PETER: ...]]` block flagging exactly what's missing and notifies Peter. Once Peter supplies the real answer, this skill hands the resolved thread off to /zynkr-kms, which writes it into the platform KB so the next identical question resolves automatically. Trigger whenever Peter says '/zynkr-support', 'draft support replies', 'process support inbox', '回覆 support 信箱', '清 support queue', '幫我回 support 信', '回覆 inbound sales', '清 sales inbox', '處理詢價信', or oth
Streamline and re-sequence a business process into a clean 'ideal flow' BEFORE it gets re-architected or automated. Runs a lean optimization pass: validate Input→Process→Output on every step, map dependencies, eliminate friction (duplicative / looping / missing steps), then apply disciplined re-sequencing only where it buys less rework, lower risk, or higher throughput. Use it whenever a process feels bloated, out of order, or 不順暢 — when someone says '幫我把流程理順', '這流程可以更精簡嗎', 'streamline this workflow', 'is my process in the right order', or right after a SIPOC / as-is map and before FE/BE/DB architecture or Lucid. Enforces 'eliminate before you automate.' Does NOT do automation-feasibility scoring or FE/BE/DB layer assignment — hand off to operations-transformation for that.
Three-stage operations transformation pipeline — process discovery (SIPOC), digital transformation assessment, then automation-ready process redesign — to turn pain points into builder-ready blueprints.
The knowledge-curator (WRITE) half of Peter's support loop. After a support / inbound-sales ticket is resolved, zynkr-kms reads the thread, learns Peter's actual answer, and writes it as a clean, LLM-friendly card into the **Zynkr platform 知識庫** (platform.zynkr.ai/kb, via the `zynkr` MCP server) — the single source of truth since the 2026-07-15 cutover. Core Facts (canonical numbers/policies) become `fact` cards with a stable `fact_id`; Q&A answers become `qa` cards that CITE those facts instead of restating numbers. It reasons like a KMS specialist (intent taxonomy → which KB section → how to structure for retrieval), proposes entries, and only writes after Peter approves. /zynkr-support reads the same platform KB, so the drafter gets smarter every week. Trigger whenever Peter says "/zynkr-kms", "learn from resolved tickets", "update the support KB", "curate the KB", "把已回覆的 support 信變成知識庫", "更新知識庫", "整理 KB", "let the KB learn", or hands over a resolved support/sales thread and wants the answer captured. This
Runs AI background research on a company (and optionally a contact) in the Zynkr platform CRM and writes a concise, sales-context zh-TW brief into the company's 備註/description field. This is a Claude Code stand-in for the platform's own "AI 背景研究" button (app/lib/ai/research.ts) — that feature only exists in the platform.zynkr.ai web UI and is NOT exposed through the connected `zynkr` MCP server, which only offers CRUD tools (create_company, create_contact, update_company, etc.), so there is no MCP call to trigger it directly. This skill closes that gap by doing the same job — targeted web research (what the company does, industry, size/HQ, notable facts, and why they might need Zynkr) — then writing the result back via update_company / create_company. Trigger EAGERLY whenever Peter says "trigger company research", "research this company", "研究這間公司", "查一下這間公司背景", "幫我查一下[公司名]", "look into [company]", "跑一下背景研究", or asks to research/enrich/look up a company or contact that is or is about to be in the Zynkr CRM — e
Turn an audio/video file (.m4a/.mp3/.wav/.mp4…) into a timestamped .srt using local Whisper. Two modes: transcribe (fresh speech-to-text with audio-accurate timestamps) or retime (keep an existing transcript's exact wording and re-derive every timestamp from the audio via forced alignment). Use when the user says 'turn this m4a into an srt', 'transcribe this audio to subtitles', '把這段錄音轉成字幕', '幫我做 srt', 'the srt wording is right but the timing is wrong, realign it', or hands over an audio file wanting subtitles. This is the FIRST step before training-srt-optimizer — it produces the raw-STT .srt; the optimizer then cleans the text. It does NOT edit/clean subtitle wording (that is training-srt-optimizer) and does NOT summarize into a recap (that is training-lecture-recap).
Turn a finished sales DEMO / discovery call into a converting follow-up — the post-meeting counterpart to sales-outbound. Give it the demo transcript (a Google Doc URL or pasted notes) and it does two things in one pass: (1) drafts a pre-sales follow-up reply as a threaded Gmail DRAFT in Peter's inbox (never sends), in the prospect's language and Peter's voice — opening by mirroring the exact pain points they raised, mapping each to the matching Zynkr AI 平台 capability (zynkr.ai/ai-platform), and moving them toward adopting the solution, with a "sales kit attached manually" line; (2) syncs the EXISTING CRM deal — finds the deal / contact / company, logs the demo as a note activity, refreshes the deal summary, nudges the stage if warranted, closes the now-stale scheduling task, and creates the agreed follow-up task. Trigger on /sales-follow-up or when Peter hands over a demo / meeting transcript after a call and says "draft a follow-up", "draft the pre-sales follow-up", "follow up on this demo", "寫跟進信", "擬一封 de
The third leg of the slide-assistant three-leg relay (picks up after slide-page-splitter, then hands off to the slide-pptx skill for rendering): once it has the SLIDE_PACKET ▸ Pages, it decides page by page "what is the most effective visual treatment for this page?" (chart / bullets / diagram / image / comparison / big statement…), and converges each page into a render-ready per-page visual spec SLIDE_PACKET ▸ Visuals, handing it off to the installed slide-pptx skill to render into a .pptx. Trigger proactively when the user hands over a ▸ Pages handoff packet, or says 「挑視覺呈現」「選版式」「決定每頁要用什麼圖」「把分頁變成可算繪的視覺規格」「準備餵 slide-pptx」. Scope boundary: it only does the visual mapping and layout of 'content → layout archetype → pptxgenjs primitive'; it does not rewrite the storyline (that is slide-storyline-designer), does not re-paginate (that is slide-page-splitter), and does not actually generate the .pptx file (that is the slide-pptx skill).