git:20260918.2429d96 to git:20260920.b0ed958

+ Zynkr platform CRM and writes a concise, sales-context zh-TW brief back onto

+ that record as a CRM note (備註) via `mcp__zynkr__create_note`. The research

+ runs on the Claude Code side — WebSearch/WebFetch plus the Taiwan government

+ registry APIs keyed off the 統一編號 — and only the write-back touches the

+ platform, so it spends NO workspace AI quota. The platform's own "AI 背景研究"

+ button and its MCP twin `mcp__zynkr__research_company` do the same job but

+ bill the workspace AI budget; prefer this skill unless Peter asks for that

+ one by name. Trigger EAGERLY whenever Peter says "trigger company

+ process: "Resolve the CRM company record → web research + 統編 registry lookups (what they do, size/HQ, facts, sales-fit angle) → dated zh-TW brief flagged unverified → write back as a CRM note via create_note → optional note on a named contact → report."

+ output: "A dated 🔎 AI 背景研究 note on the company's CRM timeline (crm_activities, kind='note'), plus an optional note on a named contact — additive, never overwriting prior records."

+ Zynkr's own products/services — and writes a concise, dated, zh-TW brief back as a

+ **note on that record's CRM timeline**, flagged as AI-researched and unverified.

+ Trigger it whenever Peter asks to research, look into, or check the background of a

+ company or contact tied to the CRM.

+ Zynkr's own AI platform (`platform.zynkr.ai`) has an "AI 背景研究" button on the

+ company/contact/deal detail page that runs a Claude tool-use loop

+ and saves a zh-TW markdown brief as a CRM note. Since then it has an MCP twin,

+ `mcp__zynkr__research_company`, so the button *can* now be pressed from here.

+ **Prefer this skill anyway.** Both platform-side routes bill the workspace AI

+ quota and cap out at ~8 generic web searches. This skill runs the research on the

+ Claude Code side — where the model and the search budget are already paid for —

+ and touches the platform only for the write-back, which is a plain

+ `create_note` CRUD call. That buys two things: it costs no CRM AI quota, and it

+ can reach sources the platform loop can't, notably the Taiwan government registry

+ APIs keyed off the 統一編號 (see step 2). Use `research_company` only if Peter

+ asks for the platform's own pipeline by name.

+ Because this is a rebuild rather than the audited in-app pipeline, **always label

+ the output as AI-researched and unverified** — the same posture the platform's own

+ feature takes (an aid to the salesperson, not a source of truth).

+

+ real structured fields you found (name, domain, industry, address, tax_id,

+ employee_count), and land the brief itself as a note in step 4. Do **not**

+ stuff the brief into `description` — see step 4 for why. Preview (`confirm`

+ omitted/false), show Peter what would be created, then re-call with

+ `confirm: true`.

+ Two tracks. Run the registry track **first** whenever the company is Taiwanese and

+ you have (or can find) a 統一編號 — it is fast, free, authoritative, and it anchors

+ whatever the web track turns up.

+ #### Track A · Government registry, keyed off the 統編

+

+ Take the 統編 from `get_company` → `tax_id`, or find it by company name. Use `curl`,

+ not `WebFetch` — WebFetch summarises through a small model, and a value that

+ overturns a CRM record should come from the raw bytes.

+

+ - **經濟部商業司 公司登記** — the authoritative name (including the 有限 / 股份有限

+ type), 公司狀況, 設立 + 最後核准變更 dates, 資本總額 and 實收資本額, 代表人,

+ registered address:

+ ```bash

+ curl -sS -g 'https://data.gcis.nat.gov.tw/od/data/api/5F64D864-61CB-4D0D-8AD9-492047CC1EA6?$format=json&$filter=Business_Accounting_NO%20eq%20<統編>&$skip=0&$top=1'

+ ```

+ - **董監事 · 所營事業 · 財政部稅籍** — the g0v mirror returns all three in one call,

+ including the shareholding split and the registered business-scope codes:

+ ```bash

+ curl -sS 'https://company.g0v.ronny.tw/api/show/<統編>'

+ ```

+ **所營事業 is the highest-signal field in the whole brief.** It says what the

+ company registered to *do*, which often contradicts how it presents itself — a

+ headhunting firm carrying `I301010 資訊軟體服務業` and `I501010 產品設計業` is

+ building software whatever its website says. 董監事 出資額 ÷ 已發行股份總數 gives

+ you the ownership split, which tells you how short the decision chain is.

+ - **Industry licence registers**, where one applies. Recruitment agencies, for

+ instance, appear in 勞動部's 私立就業服務機構 open data with licence number,

+ expiry and **員工人數** — often the only public headcount that exists:

+ ```bash

+ curl -sSL 'https://apiservice.mol.gov.tw/OdService/download/A17000000J-020013-IFV' # XML

+ ```

+

+ Treat third-party directories (twincn, datagovtw, findcompany, iyp) as leads only.

+ They lag the registry and routinely show superseded company types and addresses —

+ believing one over the registry has already produced a false "登記名稱存疑" flag.

+

+ #### Track B · Web

+

+ A handful of targeted searches, not an open-ended crawl:

+

+ 1. `WebSearch` the company name (+ domain if known) for what they do, recent

+ facts (funding, news, leadership), and anything the registry can't show.

+ 2. If an official site/About page turns up, `WebFetch` it. Many Taiwanese

+ corporate sites sit behind Cloudflare and return 403 with

+ `cf-mitigated: challenge` to both `curl` and `WebFetch` — when that happens,

+ say the claim is unverified rather than repeating it as fact. Check with

+ `curl -sS -D - -o /dev/null <url> | grep -i cf-mitigated` so the report can

+ state *why* it failed.

+ and training), why might *this* company plausibly be interested? Ground it in

+ what the research actually surfaced. The registry is especially good for this:

+ a tiny headcount next to a self-built software product is a structural

+ capability gap, and that gap is the opening.

+ 🔎 AI 背景研究

+

+ **<主題> — <YYYY-MM-DD>**｜統編 <統編>

+

+ **一句話：** 一句話講完最會改變接觸方式的那件事

+

+ **公司登記（<來源>）** 查得到出處的硬事實：名稱 · 狀態 · 資本 · 地址 · 董監事

+ **所營事業** 登記碼 + 中文名，特別是與自我定位不符的那幾項

+ **其他官方登記** 稅籍 · 產業許可（含員工人數、許可效期）

+ **銷售脈絡** 推論，每一點都扣回上面的事實

+ **本次未能查證** 明說查不到什麼、為什麼查不到

+

+ Two rules about this shape:

+ - **The first line must be exactly `🔎 AI 背景研究`** — that is the marker the

+ platform's own research writes, so the note reads as one of the family.

+ - **Keep verified fact and inference in separate sections, and name the source

+ on the fact side.** A brief that blends "the registry says 430 萬實收" with

+ "they probably need help productising" is a brief nobody can safely quote.

+ The 未能查證 section is not an admission of failure — it is what stops the

+ next reader repeating an unchecked claim as established.

+ Keep it to a short paragraph or two per section — someone will skim this on a

+ timeline, not read a report.

+ ### 4 · Write it back as a CRM note — NOT into `description`

+ **The write target is a note.** This step has gone wrong before, so be blunt

+ about why:

+ - On `platform.zynkr.ai`, **備註 is the note tab** — `crm_activities` rows with

+ `kind='note'`. That is what Peter sees on a record.

+ - `crm_companies.description` is a different thing. The company detail page

+ selects it and then never renders it; the only place it surfaces in the whole

+ UI is the *create company* modal, labelled 說明. After save it is invisible.

+ A brief written there is a brief nobody will ever read.

+ - The platform's own AI 背景研究 writes `crm_activities` too, never `description`.

+

+ So: `mcp__zynkr__create_note(company_id=<id>, subject="AI 背景研究 · <主題> <YYYY-MM-DD>",

+ body=<the brief from step 3>)` — preview first, then `confirm: true`.

+

+ Three mechanics that bite:

+

+ - **`subject` is not rendered for `kind='note'`.** The card shows `body` only.

+ Put everything you want seen in `body` and open it with the literal

+ `🔎 AI 背景研究` line. Still set `subject` — it shows in search and elsewhere.

+ - **Never put the string `⏳ 研究中` in a note body.** The detail page

+ substring-matches it and renders your finished note as a stuck "researching…"

+ spinner with a 停止 button instead of your content.

+ - Notes are additive — no read-merge-write, nothing to clobber. If the record

+ already carries an older brief, write a new note saying what changed and what

+ it corrects rather than rewriting history.

+

+ **Structured fields are a separate and welcome write.** When the research yields

+ a hard value for a column the CRM actually has — `employee_count`, `industry`,

+ `tax_id`, `address`, `domain` — set it with `update_company`, passing only the

+ fields you mean to change. That is how a fact becomes filterable instead of prose.

+

+ `create_note` takes a `contact_id` too — use it. Do **not** file a task: a task

+ says something needs doing, and this is a record.

+ - `body`: the contact portion of the brief, opening with `🔎 AI 背景研究`, plus a

+ one-line pointer back to the company record's fuller brief

+ Only create a task if the research surfaced something that genuinely needs a

+ human action, and say so in the report.

+

+ 研究完成，已寫入公司 timeline 的 note：

+ **天群企管顧問**（公司 id ...）

+ - 登記事實：實收 430 萬 · 員工 4 人 · 所營事業含資訊軟體服務業與產品設計業

+ - 銷售脈絡：4 人編制卻自建 AI 產品 → 缺的是工程與導入量能，賣顧問與訓練而非工具

+ - 已建立 note（timeline 可見）；employee_count 一併補進欄位

+ - 未能查證：官網擋自動存取（Cloudflare challenge），產品功能描述沿用前次研究

+ Say plainly what did not land. A failed CRM write, an unresolved company_id, or

+ a central claim that stayed unverified all belong in the report rather than

+ under a success line that papers over them.

+ - **Write to notes, never to `description`.** The UI never renders that column,

+ so a brief filed there is silently lost — that is exactly how a 09-18 brief

+ and its 09-20 correction both went unseen. Step 4 has the detail.

+ - **Don't spend CRM AI quota unasked.** `mcp__zynkr__research_company` and the

+ in-app button both bill the workspace AI budget. This skill exists to do the

+ same job from the Claude Code side, where it costs nothing.

+ - **The registry outranks the directories.** When twincn/datagovtw/findcompany

+ disagree with 商業司 about a company's name, type or address, 商業司 wins —

+ and name the source you used in the brief.

+ - **Preview-then-confirm on every CRM write** (`create_note`, `create_company`,

+ created, since that's a bigger commitment than adding a note.