This is worth building, and the build path matters more than the build itself. I and the Agent team surveyed 16 open-source projects and mapped them into 8 modules of our own. Phase one automates everything before the publish button and keeps a human pressing the final key: the efficiency gain lands immediately, while legal and platform risk stays at the level of ordinary manual operation. 這件事值得做，而做法比做本身更關鍵。我和 Agent 團隊盤點了 16 個開源項目，規劃蒸餾成 8 個公司自有模組。第一階段把發文前的全流程自動化，發布由人工按最後一鍵：效率收益立刻拿得到，法律與平台風險維持在一般人工營運的水準。
The core selling point is ownership. We are not stitching third-party tools together; we read how they work, keep only what our license review allows, and rebuild each capability as a skill tailored to the company, plugged into the design system, publishing scheduler and API client we already run. 核心賣點是自有化。我們沒有把第三方工具拼裝上線，我們讀懂它們的邏輯，只保留授權審查放行的部分，然後把每一項能力重寫成為公司量身打造的技能，接上我們已在運轉的設計系統、排程工具與 API 客戶端。
We are building one content production line for Xiaohongshu, run end to end by our own skills. Open-source projects contribute ideas only; the code and the workflow are ours. Most of the line lands on assets that already exist in house: our Xiaohongshu design system and three post templates, a working API client that already talks to the top MCP project in this space (MCP is a public standard plug that lets an AI assistant operate external tools directly; the technical terms in this report are explained in plain words in the cheat sheet at the end), and the Postiz scheduler. 我們要建的是一條由公司自有技能串起來的小紅書內容產線。開源項目只提供思路，程式碼與流程都是我們自己的。產線大半接在公司既有資產上：小紅書專屬設計系統與三套圖文版型、一支已經在對接該領域最大 MCP 項目的 API 客戶端（MCP 是一種公開的通用接頭規格，讓 AI 助理可以直接操作外部工具；報告裡的技術名詞，文末有白話解釋的術語小抄），以及 Postiz 排程工具。
Every automated action in this line writes a full audit log, and no step ever publishes without the human sign-off gate. That is a design decision, not a temporary limitation; section 06 explains why. 產線上每個自動化動作都留完整日誌，任何步驟都不會跳過人工終審自行發布。這是設計決定，也是長期原則，理由見第 06 節。
Stars, last update and license were pulled live from the GitHub API on 2026-08-13, not quoted from memory or third parties. A missing license means the author keeps all rights by default: such code may inform our design but must never be copied into commercial use. Nothing in the flagged rows is treated as usable for commercial code reuse. Star 數、最後更新與授權皆於 2026-08-13 用 GitHub API 即時查證，非憑印象或第三方轉述。沒有授權聲明代表作者預設保留全部權利：這類程式碼只能參考設計思路，絕對不可複製進商用。凡掛標記的項目，一律不視為可以商用引用程式碼。
| # | Project項目 | Category分類 | StarsStar | Last update最後更新 | License授權 | What we take from it我們拿它做什麼 |
|---|---|---|---|---|---|---|
| 1 | xiaohongshu-mcpxpzouying | MCP | 15,232 | 2026-08-10 | Apache-2.0 | Our primary reference. MCP protocol layer and REST endpoint design are reusable under this license; our own code already talks to its API, so this is an extension, no cold start.主力參考。MCP 協議層與 REST 端點設計在此授權下可引用；我們已有程式碼在對接它的 API，這裡是擴充，沒有冷啟動。 |
| 2 | social-auto-uploaddreammis | Distribution分發 | 14,193 | 2026-07-31 | no license, all rights reserved無授權檔，作者保留全部權利 | Multi-platform upload abstraction as a design idea only. No code reuse. Contacting the author for a written license is worth trying and has not been done yet.多平台上傳抽象層只取設計思路，不碰程式碼。可嘗試聯絡作者取得書面授權，此事尚未進行。 |
| 3 | Wechatsyncwechatsync | Distribution分發 | 6,157 | 2026-05-27 | GPL-3.0 copyleft, legal review before any code reuseCopyleft 授權，引用程式碼前需法務複核 | One-source-to-many-platform bridging architecture. GPL-3.0 can force us to open-source derivatives if we distribute; until legal signs off we take architecture ideas only.單一內容源對多平台的橋接架構。GPL-3.0 在對外散布時可能觸發衍生作品開源義務，法務放行前只取架構思路。 |
| 4 | XiaohongshuSkillswhite0dew | Publishing發佈 | 3,315 | 2026-08-12 | MIT | CDP browser automation steps for the web upload flow. We exclude its auto-comment feature by policy.網頁版上傳流程的 CDP 瀏覽器操作步驟。其自動評論功能依我們的規範排除不取。 |
| 5 | xiaohongshu-ops-skillXiangyu-CAS | Content內容產製 | 2,246 | 2026-03-19 | no license, all rights reserved無授權檔，作者保留全部權利 | Topic selection and post-mortem methodology as ideas only; its skill format also targets a different agent stack, so a rewrite was needed regardless.選題與復盤方法論只取思路；它的技能格式本來就設計給另一套 agent 架構，無論如何都要重寫。 |
| 6 | xhs_ai_publisherBetaStreetOmnis | Content內容產製 | 2,057 | 2026-06-28 | Apache-2.0 | Copy template and cover template data structures, plus its session-reuse pattern for staying logged in. Its desktop UI is irrelevant to us.文案模板與封面模板的資料結構，以及登入態複用的實作模式。它的桌面 UI 我們用不到。 |
| 7 | creatorhub3441293738 | Accounts帳號管理 | 975 | 2026-08-12 | no license, all rights reserved無授權檔，作者保留全部權利 | Dashboard data model (accounts, posts, comments, notifications) as an architecture reference only. Its monitoring approach needs its own compliance check before anything is borrowed.面板資料模型（帳號、作品、留言、通知）僅作架構參考。其監控方式在借鑒前需另做合規查證。 |
| 8 | x-mcpxpzouying | Publishing發佈 | 427 | 2026-08-11 | no license, all rights reserved無授權檔，作者保留全部權利 | Browser-extension-to-MCP bridging idea only; no code reuse.瀏覽器擴充與 MCP 橋接的架構思路，不碰程式碼。 |
| 9 | xiaohongshuLuckyTime1025 · archived | Publishing發佈 | 242 | 2023-04-26 | BSD-3-Clause | Historical reference only. License is permissive but the project is archived and almost certainly broken against today's platform.僅作歷史參考。授權寬鬆但項目已封存，對上今天的平台大機率已失效。 |
| 10 | openclaw-xhszhjiang22 | MCP | 117 | 2026-03-08 | NOASSERTION, terms need manual readingNOASSERTION，需人工開檔核對條款 | Hot-topic tracking design idea. Its license file exists but cannot be classified automatically; until someone reads the actual terms we treat it as unusable for code.熱點跟蹤的設計思路。它有授權檔但無法被自動歸類，實際讀過條款之前，程式碼一律當作不可用。 |
| 11 | xiaohongshu-spider-visualizerKaitoHH | Data數據 | 33 | 2018-08-01 | no license, all rights reserved無授權檔，作者保留全部權利 | Not adopted. Eight years unmaintained; listed to show this category exists but is dead.不採用。八年未維護，列出只為說明這類項目存在但已死。 |
| 12 | xhs_mcp_agentproerror77 | MCP | 18 | 2025-04-25 | no license, all rights reserved無授權檔，作者保留全部權利 | Not adopted. Low maturity and no license; listed for completeness.不採用。成熟度低且無授權，為完整性列入。 |
| 13 | xhs_one_spidermashukui | Data數據 | 12 | 2026-08-11 | MIT use case needs compliance review使用場景需合規複核 | Keyword search and comment parsing logic. The license permits commercial use, but the author markets it for traffic-diversion purposes, so every borrowed piece goes through our compliance review first.關鍵字搜尋與留言結構化解析邏輯。授權允許商用，但作者以引流獲客為賣點，凡借鑒的部分一律先過我們的合規複核。 |
| 14 | mcp-xiaohongshutimecyber | MCP | 12 | 2025-06-04 | MIT | Cross-check reference for publish parameter design (tags, privacy settings) against project #1. Not a primary source.發布參數設計（標籤、隱私設定）與第 1 項交叉驗證用，不當主力。 |
| 15 | rednote-analyzer-mcpShellyDeng08 | Data數據 | 8 | 2026-03-19 | no license, all rights reserved無授權檔，作者保留全部權利 | The product idea of decomposing viral posts into structured metrics. Idea only; low maturity, no code reuse.把爆款筆記拆成結構化指標的產品思路。只取思路，成熟度低，不碰程式碼。 |
| 16 | BatchShuBatchShu | Distribution分發 | 8 | 2025-08-01 | no license, all rights reserved無授權檔，作者保留全部權利 | Not adopted. Listed as proof the batch-publishing need has been attempted before.不採用。列出僅證明批量發佈這個需求已有人做過。 |
Coverage caveat: this survey was time-boxed, and projects hosted outside GitHub or distributed privately were not exhaustively searched. not exhaustive 涵蓋範圍說明：本次盤點有時間上限，非 GitHub 平台或私下流通的同類項目未做窮盡式搜索。非窮盡盤點
Every module below is redesigned and rebuilt, never copied. The internalisation method, at outline level, is four steps: read the source logic without lifting code; filter by license so only Apache, MIT or BSD material may ever be quoted; rewrite each capability as a company skill using our own skill-building toolchain and wire it into existing assets; then wrap it with a human review gate and full logging. The detailed design docs live outside this report. 下面每個模組都是重新設計與實作，不搬運程式碼。內化方法講大綱層級就好，共四步：讀懂對方邏輯但不搬程式碼；用授權過濾，只有 Apache、MIT、BSD 類授權的內容才可能被引用；用公司自己的技能建置工具把能力重寫成公司技能並接上既有資產；最後包上人工審核閘門與完整日誌。細部設計文件另出，不佔這份報告篇幅。
We are not starting from zero. The company already owns a Xiaohongshu design system with three post templates, a white-theme stylesheet, image generation and social content skills, the Postiz scheduler, and an API client that already calls the REST endpoints of project #1 above. a few of these not individually re-tested yet All of these are on the company's tool inventory and in use; a few were not individually switched on and re-tested in this pass, and each will be verified working before build starts. 我們沒有從零開始。公司已有小紅書專屬設計系統與三套圖文版型、白底主題樣式、生圖與社群內容技能、Postiz 排程工具，以及一支已在呼叫上表第 1 項 REST 端點的 API 客戶端。其中幾項本次未逐一實測 這些資產都在公司的工具清單上且有在使用，其中幾項這次沒有逐一開起來實測，動工前會先逐一驗過確認能跑。
| Our module我們的模組 | Capability distilled from蒸餾自哪些能力 | What we rebuild it into我們改造成什麼 | Effort工時 |
|---|---|---|---|
| 1 · Topic and trend radar1 · 選題與熱點雷達 | Topic and retro methodology (#5), hot-topic tracking design (#10), viral-pattern metrics (#15). Ideas only; none of the three permits code reuse.選題與復盤方法論（#5）、熱點跟蹤設計（#10）、爆款指標框架（#15）。全部只取思路，三者都不允許引用程式碼。 | A topic pipeline on our own research and digest tools, producing a weekly topic list.接上我們自己的研究與彙整工具，重寫成每週產出選題清單的產線。 | 2 to 3 wks rough estimate2 至 3 週 粗估 |
| 2 · Content generation2 · 圖文內容生成 | Copy and cover template structures from #6 (Apache-2.0, quotable).#6 的文案與封面模板結構（Apache-2.0，可引用）。 | Plugged straight into our existing Xiaohongshu design system and three templates; most material already exists.直接接上公司既有的小紅書設計系統與三套版型，多數素材已現成。 | 1 to 2 wks rough estimate1 至 2 週 粗估 |
| 3 · Pre-publish packaging and handoff layer3 · 發佈前打包與人工按鍵銜接層 | MCP architecture and REST endpoints from #1 (Apache-2.0), parameter design cross-checked against #14.#1 的 MCP 架構與 REST 端點（Apache-2.0），參數設計以 #14 交叉驗證。 | Extend our existing read-only API client to assemble each approved post into a paste-ready package and queue it; a person presses the publish key. The publish action itself is not automated in phase one, so this module touches no unknown channel.把我們既有的唯讀 API 客戶端擴充成銜接層：把過審的貼文組成可一鍵貼上的成品包並排進佇列，發布鍵由人來按。發布動作本身第一階段不自動化，所以這個模組不依賴任何未知通道。 | 2 wks rough estimate2 週 粗估 |
| 4 · Browser upload automation4 · 瀏覽器自動化上傳 | CDP operation patterns from #4 (MIT, quotable). Auto-comment excluded.#4 的 CDP 操作模式（MIT，可引用）。自動評論功能排除。 | Rewritten as our own Playwright scripts, upload only, connected to the Postiz scheduling layer. Evaluated only after official channels are ruled out.重寫成公司自己的 Playwright 腳本，只做上傳，接上 Postiz 排程層。官方通道查清確定沒有之後才評估啟用。 | 3 wks rough estimate3 週 粗估 |
| 5 · Data and competitor monitoring5 · 數據抓取與競品監測 | Keyword and comment parsing logic from #13 (MIT), trend structuring idea from #15.#13 的關鍵字與留言解析邏輯（MIT）、#15 的趨勢結構化思路。 | Rebuilt to read public data only, at human pace, for internal decisions only, with no anti-control evasion of any kind.重寫成只讀公開數據、維持人類節奏、僅供內部決策的抓取器，不含任何規避風控的行為。 | 2 to 3 wks rough estimate2 至 3 週 粗估 |
| 6 · Account and asset dashboard6 · 帳號與素材管理儀表板 | Four-entity data model from #7. Architecture reference only, no code.#7 的四類實體資料模型。僅參考架構，不引入程式碼。 | Rebuilt on our own stack, connected to Postiz and internal tooling.用公司既有技術棧重寫，串接 Postiz 與內部後台。 | 3 to 4 wks rough estimate3 至 4 週 粗估 |
| 7 · Multi-platform bridge7 · 多平台分發橋接 | Sync bridging architecture from #3 (GPL-3.0, legal review before any code) and batch ideas from #16 (reference only).#3 的同步橋接架構（GPL-3.0，引用程式碼前需法務複核）、#16 的批量思路（僅供參考）。 | Bridge covering Xiaohongshu plus the platforms Postiz already supports; no attempt at full-platform coverage.只做小紅書加上 Postiz 已支援平台的橋接，不追求全平台覆蓋。 | 1 to 2 wks rough estimate1 至 2 週 粗估 |
| 8 · Comment response assistant8 · 互動與留言回應輔助 | Feature definitions from #4 and #5, taken as scoping references only.#4 與 #5 的功能定義，僅作範圍界定參考。 | Classifies comments and drafts suggested replies; a person confirms and sends every one. No automatic commenting, ever.做留言分類加建議回覆草稿，每一則都由人確認後手動發送。永遠不做全自動留言。 | 1 to 2 wks rough estimate1 至 2 週 粗估 |
Honesty about these five points is the value of this report. Each one states where it hurts, how we plan to solve it, and what it costs us if it stays unsolved. Anything without a verified number carries a flag; I will not fill gaps with guesses. 把這五點講老實，就是這份報告的價值。每一點都寫清楚難在哪、我們打算怎麼解、解不了要付什麼代價。凡沒有查證數字的一律掛標記，我不用猜的值補空。
There is no official token channel, so all automation lives on a logged-in browser credential that can die at any moment, with no warning beforehand. Device, network or IP changes can also force re-verification. Some tools in the wild survive this by rotating large numbers of accounts; that exact behaviour was named by a court in a final judgment, and we will not touch it.平台沒有官方的換發 Token 通道，所有自動化都靠一份登入後的瀏覽器憑證活著，它隨時可能失效，事前不會有任何預警。換裝置、換網路、換 IP 也可能觸發重新驗證。市面上有些工具靠大量帳號輪替撐過去，那個做法已被法院在生效判決中點名，我們不碰。
Our plan: one company-owned official account, a single encrypted credential, scheduled low-risk health checks, instant Telegram alerts on failure, and a task queue that pauses instead of dropping work. The goal is turning a silent death into a pause that a person knows about within minutes.我們的解法：只用一個公司自有官方帳號，單一登入態存進加密憑證庫，定時做低風險健康檢查，失效即時打 Telegram 告警，任務進佇列暫停保留，不丟棄。目標是把無聲死亡變成幾分鐘內就有人知道的暫停。
If unsolved: the system degrades from automatic to semi-automatic, one manual QR scan per expiry. If expiry happens several times a day, manual cost eats most of the value. actual credential lifetime unknown, needs 2 to 4 weeks of live testing No number will be promised before our own account measures it.解不了的代價：系統從全自動降為半自動，每次失效要人工掃碼一次。若一天失效數次，人工成本會吃掉大半價值。憑證實際有效天數未知，需自有帳號實測 2 至 4 週 實測完成前不承諾任何天數。
This is the single biggest unknown in the project. Every official open-platform capability we found targets e-commerce merchants and service providers; we found no note-publishing API for ordinary brand accounts. inferred from the public document catalogue, not officially confirmed The official pages render as JavaScript shells to our fetcher, so this conclusion is inferred from the visible document catalogue; confirming it requires an enterprise developer account to view the grantable permission list.這是全案最大的未知數。目前查到的官方開放平台能力全部指向電商商家與服務商場景，沒有找到面向一般品牌帳號的筆記發布 API。依官方文件目錄推斷，尚未取得官方明文確認 官方頁面對我們的抓取只回傳 JS 殼，這個結論屬於依公開文件目錄推斷，要坐實需要企業開發者帳號登入查看可申請權限清單。
Our plan is three stages. Stage one builds everything before the publish click: topics, copy, images, tags, spec checks, one paste-ready package, and a person presses the final key; this carries no compliance or stability risk. Stage two gets the enterprise developer account and reads the real permission list. Stage three, only if no official channel exists, evaluates browser automation on our own account under human supervision.我們的解法分三階段。第一階段做發文前的一切：選題、文案、圖片、標籤、規格檢查，產出可一鍵貼上的成品包，發布由人按最後一鍵，這段沒有合規與穩定性風險。第二階段完成企業開發者認證，查清真正的權限清單。第三階段在確定沒有官方通道時，才評估自有帳號、有人監看的瀏覽器自動化。
If unsolved: we still capture the content production gain, which is most of the manual workload, but unattended operation stays out of reach and expectations for the word automatic need adjusting. the exact share of workload saved is unmeasured解不了的代價：內容生產的效率收益照樣拿得到，那已是人力的大宗，但無人值守營運做不到，對自動發這三個字的期待需要調整。實際省下的工作量比例未實測
A note can show as published and still get near-zero distribution. The platform's machine-plus-human review and its distribution rules are undocumented, so an automated system cannot tell success from silent suppression. Anyone claiming guaranteed reach is selling marketing talk.一篇筆記可以顯示已發布，實際幾乎沒有曝光。平台的機審加人審與分發規則都不公開，自動化系統分不出真成功與被靜默壓量。任何宣稱保證不限流的說法都是行銷話術。
Our plan: measure outcomes instead of guessing rules. A post-publish tracker re-checks reach and engagement at fixed intervals, marks posts that fall well below our own historical baseline as suspected throttling, and builds an internal case library. Before publishing, a self-check screens content against our own list of publicly documented banned categories and claim patterns.我們的解法：不猜規則，改量結果。發布後追蹤器在固定時間點回查曝光與互動，明顯低於我們自己歷史基準線的標記為疑似限流，累積成內部案例庫。發布前另有一道自檢，比對我們自己整理的公開可查違禁品類與宣稱用語清單。
If unsolved: output rises while reach does not, and the return on the whole project becomes unmeasurable. That is the most dangerous failure mode here, so the tracker gets the same priority as the generator. no official numbers exist for review time, throttling thresholds or daily post limits解不了的代價：產能上去了曝光沒上去，整個專案的投入產出比失真。這是最危險的失敗模式，所以追蹤器的優先級與生成器同級。審核時長、限流閾值、單日發文上限皆無官方數字
A wrong cover ratio, file size or word count voids the whole post, and the specs change without notice. There is no official spec file to sync against, and AI-generated images often miss the platform's preferred ratios.封面比例、檔案大小或字數錯了整篇作廢，而且規格會不預告地改。沒有官方規格文件可以同步，AI 生成圖片又常不符合平台偏好的比例。
Our plan: one central spec config that every generation step reads, a mandatory post-generation validator that resizes or regenerates failures, and covers produced in the platform's mainstream portrait ratio with subjects kept in the safe zone.我們的解法：規格抽成一份中央設定檔，所有生成流程都從它讀取；生成後強制驗證，不合格自動重算或退回重生成；封面固定用平台主流直式比例，主體放在安全區內。
If unsolved: this is the lightest of the five. Worst case we trail spec changes by a week or two with manual checks in between; no systemic failure. first spec baseline must be built from real upload tests, no official document obtained解不了的代價：這是五點裡最輕的。最壞情況是平台改規格後我們慢一到兩週跟上，期間人工檢查，不會造成系統性失敗。第一版規格基準需靠實際上傳測試建立，未取得官方文件
Data is the agent's eyes, and this eye breaks two ways. Technically, page structures change and requests get intercepted unpredictably. Legally, what you collect, how much, and what you do with it decided both court cases in section 06; the grey-market answer of reverse-engineering signatures and rotating IPs is exactly what a court ruled to be circumvention.數據是 Agent 的眼睛，這隻眼睛會壞在兩個地方。技術上，頁面結構會改、請求會被不可預測地攔截。法律上，抓什麼、抓多少、拿來做什麼，正是第 06 節兩個判決的核心爭點；灰色市場那套逆向簽名、輪替 IP 的答案，正是法院認定的規避技術保護措施。
Our plan splits data into three buckets. Our own account data comes from our own dashboard and gets built out fully first. Public industry trends are read at low frequency, at human pace, with no circumvention, for internal decisions only. Anything requiring circumvention to obtain is abandoned outright and written into the design docs as a prohibited item.我們的解法是把數據分三桶。自家帳號數據來自我們自己的後台，最安全，優先做完整。公開行業趨勢採低頻率、人類節奏、不繞任何技術保護措施的方式讀取，僅供內部決策。凡需要繞過技術保護措施才拿得到的數據，直接放棄，寫進設計文件明列為禁止項。
If unsolved: our competitor intelligence stays shallower and our market reaction slower than players who take the grey route. That is a deliberate price paid for the company's legal safety, and I state it here as a choice, not a shortfall.解不了的代價：競品情報做得比走灰色路線的人淺，市場反應慢半拍。這是為公司法律安全刻意付的代價，我在這裡講明這是主動選擇，不是能力缺口。
Case one: a data-tool vendor scraped Xiaohongshu at scale and sold the data, using rapid user-ID and IP rotation to bypass protections. Final judgment by the Hangzhou Intermediate People's Court, RMB 4.9m in damages, reported April 2025 (Tencent News, Southern Metropolis Daily). defendant name and case number not disclosed in public reports Reports identify the defendant only as a Xiamen network company; naming it would require the original judgment. Case two: Xingyin Information Technology (Shanghai) Co., Ltd., Xiaohongshu's operator, versus the maker of the two Guqiao tools. Fujian High People's Court, second instance, case (2022) Min Min Zhong No. 1871, upheld the first-instance ruling: batch-downloading watermarked images did not constitute unfair competition, while altering video MD5 fingerprints to evade detection did. The case entered the 50 exemplary IP cases of Chinese courts for 2023, published by the Supreme People's Court in April 2024 (counsel's case note, Sina Finance). The boundary is precise: reading public pages was not the offence; evading protection and reselling data was. 案例一：數據工具商大規模爬取小紅書並打包販售，以頻繁更換使用者 ID 與 IP 繞過保護措施，杭州市中級人民法院終審判賠人民幣 490 萬元，2025 年 4 月經媒體報導（騰訊新聞、南方都市報）。被告全名與案號公開報導未披露 報導僅稱被告為「廈門某網路公司」，要點名公司需調閱原始判決書。案例二：小紅書運營方行吟信息科技（上海）有限公司，對上開發「固喬」兩款軟體的公司。福建省高級人民法院二審，案號（2022）閩民終 1871 號，維持一審：批量下載帶浮水印圖片被認定不構成不正當競爭，修改影片 MD5 特徵值規避檢測被認定構成。該案入選最高人民法院 2024 年 4 月發布的「2023 年中國法院 50 件典型知識產權案例」（代理律所案例說明、新浪財經）。界線畫得很準：單純讀取公開頁面不是違法行為，規避保護措施與轉售數據才是。
| Standard把關標準 | What it means in practice具體含義 |
|---|---|
| Company-owned official accounts only只用公司自有的官方帳號 | One formal account per business line; clear ownership, traceable responsibility一個業務一個正式帳號，主體清楚、責任可追 |
| Public content only, at human pace只讀公開內容，維持人類節奏 | No high-frequency collection, no load imposed on the platform不做高頻抓取，不製造平台負擔 |
| Full logs on every automated action所有自動化行為留完整日誌 | Who did what, when, on which account; fully replayable誰在什麼時間用哪個帳號做了什麼，可完整回溯 |
| Human final review before publishing發布前人工終審 | A person signs off every post in phase one; responsibility never rests on a machine第一階段每篇出街前有人簽字，責任不落在機器上 |
| Data for internal decisions only數據僅供內部決策 | Never packaged, sold, or passed to any third party不打包、不販售、不轉手給任何第三方 |
| Official channels first優先走官方通道 | We use what the platform opens; what it does not open gets evaluated before anything else happens官方開放什麼我們用什麼，官方沒開放的先評估再說 |
legal review pending This section is our operating standard, not a legal opinion. Before the project formally starts I will have it reviewed by legal counsel, particularly the frequency boundary for reading public content, where case law gives no explicit number. 待法務複核 本節是我們的營運標準，不是法律意見。專案正式啟動前我會把這一節送法務或外部律師複核，特別是公開內容讀取頻率的界線，司法實務沒有給出明確數字。
Every number in this section is an unvalidated single-person estimate; none has been through real scheduling or engineering review. I give them so the scale is visible, not as a commitment. all effort figures are rough estimates, not commitments 本節每個數字都是未經驗證的單人粗估，沒有經過實際排期或工程複核。給數字是為了讓量級看得見，不是承諾。所有工時皆為粗估，非承諾
| Module模組 | Precondition前置條件 | Effort (rough)工時（粗估） |
|---|---|---|
| 1 · Topic and trend radar1 · 選題與熱點雷達 | None; existing research tools無，用既有研究工具 | 2 to 3 wks2 至 3 週 |
| 2 · Content generation2 · 圖文內容生成 | None; design system exists無，設計系統已現成 | 1 to 2 wks1 至 2 週 |
| 3 · Pre-publish packaging and handoff layer3 · 發佈前打包與人工按鍵銜接層 | Existing API client; includes testing buffer for platform changes既有 API 客戶端，含平台改版的測試緩衝 | 2 wks2 週 |
| 4 · Browser upload automation4 · 瀏覽器自動化上傳 | Only after official channels are confirmed absent確認沒有官方通道後才啟動 | 3 wks3 週 |
| 5 · Data and competitor monitoring5 · 數據抓取與競品監測 | Includes compliance review time含合規複核時間 | 2 to 3 wks2 至 3 週 |
| 6 · Account and asset dashboard6 · 帳號與素材管理儀表板 | UI plus backendUI 加後端 | 3 to 4 wks3 至 4 週 |
| 7 · Multi-platform bridge7 · 多平台分發橋接 | Legal sign-off on the GPL question firstGPL 問題先過法務 | 1 to 2 wks1 至 2 週 |
| 8 · Comment response assistant8 · 互動與留言回應輔助 | None無 | 1 to 2 wks1 至 2 週 |
| Total, single person, sequential合計，單人全職依序 | rough estimate, not reviewed by engineering粗估，未經工程複核 | 15 to 21 wks15 至 21 週 |
Two facts keep the real calendar shorter than the raw total: modules 1, 2, 3 and 8 have no preconditions and can start immediately, and I run an Agent team, so several modules can proceed in parallel rather than strictly one after another. A precise cost figure has to wait for the answer to Q1 in section 09, because the browser-automation route carries noticeably higher build effort and long-term maintenance cost than the API route. 兩件事會讓實際日曆時間短於帳面合計：模組 1、2、3、8 沒有前置條件可立即動工，而且我手上有 Agent 團隊，幾個模組可以平行推進，未必嚴格排隊。精確的成本數字要等第 09 節 Q1 的答案，因為走瀏覽器自動化比走 API，工時與長期維護成本都會明顯增加。
On money, plainly: the person in these estimates is me, and the weeks are absorbed by me plus the Agent team within my current role, so this asks for no new headcount, no outside contractors, and no software spend beyond tools we already run. The real cost is opportunity cost: while this is being built, the time I can put into other company work shrinks by the same amount, and the company decides where I rank this against everything else on my plate. 錢的部分講明白：這些工時估的人就是我，由我加上 Agent 團隊在現職範圍內消化，所以不新增人事、不外包、不買既有工具以外的軟體。真正的成本是機會成本：做這件事的期間，我能投入公司其他工作的時間會等量減少，這件事排在我手上其他事情的前面還是後面，由公司決定。
Everything we found in the official open platform targets e-commerce and service-provider scenarios; no note-publishing API for ordinary brand accounts appeared anywhere. The publishing API documents circulating online are third-party reverse-engineered material, and using them crosses the red line in section 06.目前查到的官方開放平台能力全部集中在電商與服務商場景，沒有找到面向一般品牌帳號的筆記發布 API。網路上流傳的發布 API 文件多是第三方逆向整理，用了就踩到第 06 節的紅線。
final answer requires enterprise developer login The complete answer sits behind the enterprise developer console. Once the company provides registration documents, verification plus checking the permission list is a rough 3 to 5 working days.完整答案需企業開發者帳號登入後台 完整版答案在企業開發者後台裡。公司提供企業資質後，認證加查證粗估 3 至 5 個工作天。
API openness of the five commercial platforms unverified The same login answers a second critical unknown: whether the platform's five commercial tools (influencer marketplace, ad buying, shop tooling, content boosting, data insights) expose any API. That decides whether ad placement and influencer collaboration can ever be automated; no official API documentation for any of the five was found in this pass.五個商業平台的 API 開放程度未查證 同一個後台還能回答第二個關鍵未知數：平台的五個商業工具（蒲公英：達人合作媒合、聚光：廣告投放、千帆：商家店鋪經營、薯條：貼文加熱推廣、靈犀：數據洞察）是否對外開放 API。這決定了投放與達人合作這一段能不能自動化，本次未查到其中任何一個的官方 API 文件。
no defensible probability figure exists The platform does not publish its risk-control rules, so any ban-probability number from anyone is unfounded, and I will not invent one.封號機率無任何有依據的數字 平台不公開風控規則，任何人給的封號機率數字都沒有依據，我也不會編一個。
What I can state is the risk posture. Phase one does no automated publishing, so ban risk sits at ordinary manual-operation level. Any later automation runs only on our own official account, at human pace, during supervised hours. And all content and publish records live in our own system, so even if the account is restricted, the content assets survive and switch to manual publishing or another platform immediately.能講清楚的是風險姿態。第一階段不做自動發布，被封風險就是一般人工營運的水準。之後若導入自動化，僅限自有官方帳號、人類節奏、有人監看的時段。而且所有內容與發布紀錄都存在我們自己的系統裡，帳號就算受限，內容資產不會消失，可立即轉人工發布或轉往其他平台。
The ones who do it that way include two companies that have already lost in court: final damages of RMB 4.9 million in one case, and a ruling of unfair competition that became a Supreme People's Court exemplary case in the other. The tools still running mostly sit in a not-yet-sued state, which is different from a lawful state. We are a registered company with real assets and reputation; our downside in a lawsuit is a different order of magnitude from a nameless studio's.做得到的那批人裡，已經有兩家輸過官司：一家終審定讞判賠人民幣 490 萬元，另一家被定性為不正當競爭並成為最高人民法院典型案例。市面上還在跑的工具多數處於還沒被告的狀態，這跟合法狀態是兩回事。我們是有登記、有資產、有商譽的公司，敗訴的損失量級跟一個無名工作室完全不同。
Our edge is depth on the lawful side: topic selection, generation, spec checking, performance tracking and internal data accumulation, built to a depth the grey-route players never invest in.我們的優勢在合法那一側做深：選題、生成、規格檢查、成效追蹤、內部數據沉澱，做到走灰色路線的人從來不投資的深度。
The known damages scale is millions of RMB, the cause of action is unfair competition, and the defendants were the companies that built and operated the tools. Our planned approach, own accounts, public content, no circumvention, no external data sales, sits outside every illegal element those judgments identified.已知判決的賠償量級是人民幣數百萬元，案由是不正當競爭，被告是開發並營運工具的公司主體。我們規劃的做法，自有帳號、公開內容、不繞保護措施、不對外賣數據，落在那些判決認定的每一項違法情節之外。
legal review still required One thing I cannot do is substitute for legal counsel. The frequency boundary for reading public content has no explicit number in case law, and section 06 goes to legal before the project formally starts.仍需法務複核 有一件事我不能代替：法務判斷。公開內容讀取頻率的界線在司法實務上沒有明確數字，第 06 節會在專案正式啟動前送法務。
precise schedule and cost blocked on Q1 The effort scale is already on the table in section 07: 15 to 21 weeks for one person building all 8 modules sequentially, a rough estimate. What cannot be quoted yet is a precise schedule and money figure, because that depends on Q1: the browser-automation route carries noticeably higher build effort and long-term maintenance cost than the API route. Pinning an exact total today would be a made-up number, and this report does not do made-up numbers.精確排期與金額卡在 Q1 工時量級第 07 節已經給了：單人依序做完 8 個模組是 15 至 21 週，屬粗估。現在還報不出來的是精確排期與金額，因為那取決於 Q1：走瀏覽器自動化比走 API，工時與長期維護成本都會明顯增加。此刻把精確總數報死就是編數字，這份報告不編數字。
What can be committed today is phase one: the full pre-publish pipeline with a human pressing the final key. It depends on no unknowns, its scope is defined, and modules 1, 2, 3 and 8 start the moment it is approved. precise effort quoted after scope confirmation今天就能承諾的是第一階段：發文前全流程，人按最後一鍵。它不依賴任何未知數，範圍明確，核准當下模組 1、2、3、8 就能動工。精確工時待範圍確認後再報