Chinese AI Pro: Build a Reviewed Content Pipeline with API and n8n

5 viewsAI content creationChinese AIworkflow automationn8nAI API

For creators who already publish consistently: use the DeepSeek API, Volcengine Ark, Jimeng, Kling or Hailuo, Jianying, and n8n to turn repetitive topic, script, shot, and cost work into a measurable, recoverable, traceable content system with human review gates.

Chinese AI Pro Edition cover

Chinese AI Pro Edition cover

For people who can already publish consistently and want to reduce copy-pasting, standardize quality across a content series, and support team collaboration. The "automation" discussed in this article always includes human review — it does not mean letting the system publish without anyone checking.

Introduction: When Is the Paid Route Actually Worth It?

Free tools can help you finish content, but as your publishing frequency rises, new problems appear: the same set of requirements has to be typed in every day; titles, scripts, and shot lists are scattered across different chat sessions; when several people collaborate, nobody knows which version is final; when a generation step fails, you have to start over; and at the end of the month you only know "we spent quite a bit," with no way to tell which piece of content the cost belongs to.

At that point, the value of paying is not just getting a stronger model — it is turning your work into a system that is measurable, recoverable, and traceable.

A common misconception is that subscribing to more tools will automatically make the content better. In reality, the professional route requires solving three foundational problems first:

  1. Standardized input. Describe every topic with the same fields to reduce prompt drift.
  2. Traceable process. Save sources, model outputs, human edits, and the published version.
  3. A responsible owner at every critical node. Facts, copyright, brand voice, and the publishing decision must never lack an owner.

This article builds a concrete case: every morning, screen "AI office productivity" topics from a pool of candidate materials, automatically generate 10 candidate titles, 3 script versions, and 1 shot list, then stop at the review queue, where an editor decides whether to continue into production.

Chinese AI professional workflow

Chinese AI professional workflow
graph LR
    A[Scheduled trigger] --> B[Collect candidate materials]
    B --> C[Dedupe and classify]
    C --> D[Model scoring]
    D --> E[Generate content package]
    E --> F{Human review}
    F -->|Approve| G[Generate assets]
    F -->|Send back| H[Record reason]
    H --> D
    G --> I[Edit and publish]
    I --> J[Archive data]

1. First, Distinguish Subscriptions, APIs, and Automation Platforms

The professional route usually involves three kinds of spending. They solve different problems and should not be lumped together.

TypeBest-suited tasksStrengthsCommon risks
Personal subscriptionManual conversation, drafting, research, and revisionQuick to start, full-featured interfaceHard to call in bulk, versions get scattered
Pay-as-you-go APIProgrammatic generation, batch processing, structured outputCosts are calculable, embeddable in pipelinesRequires key management, retries, and budget controls
Automation platformConnecting spreadsheets, models, storage, notifications, and other stepsVisual orchestration, less copy-pastingOverly long workflows become hard to maintain

Comparison of subscriptions, APIs, and automation platforms

Comparison of subscriptions, APIs, and automation platforms

The official DeepSeek API documentation covers models, invocation methods, billing, and error handling; Volcengine Ark provides documentation for model services, APIs, and pricing; the official n8n docs define a workflow as an automated process formed by connected nodes and support integration with many AI services. Specific model names, prices, and quotas may change, so your system design should avoid hard-coding any particular model or unit price into the pipeline.

The first principle of the professional route: write the process down clearly before buying tools. If you do not yet know which actions you repeat every day, adding subscriptions will only add decision overhead.

2. The Minimal Professional Tool Stack

Division of labor in the Chinese AI professional stack

Division of labor in the Chinese AI professional stack

Below is a fairly general-purpose combination — not the only correct answer.

LayerExample toolsMain responsibilities
Orchestrationn8nScheduled triggers, conditional logic, retries, notifications, logging
Text modelDeepSeek API, text models on Volcengine ArkTopic scoring, scripts, rewriting, structured output
ImageJimeng or Ark image modelsCover drafts, shot illustrations, infographic assets
VideoKling, Hailuo, or whichever video model your account supportsKey shots, transition footage, image-to-video
Final cutJianying (known internationally as CapCut)Subtitles, voiceover, track editing, final human review and export
DataSpreadsheets, databases, or project management toolsStoring topics, statuses, versions, costs, and performance data

Do not plug in every tool at the start. Begin by connecting n8n to just "the topic sheet + one text API + a notification channel," and only add image and video steps once that runs reliably. Text is cheap and its errors are easy to check, which makes it the ideal first automation chain.

3. Replace Casual Chat with Structured Content Cards

When creating by hand, an author can fill in context from memory; an automated system cannot. Every topic should first be converted into a structured content card.

{
  "topic_id": "20260720-001",
  "working_title": "What AI most often misses when organizing meeting minutes",
  "audience": "Office workers who handle meeting notes",
  "platform": "vertical-short-video",
  "duration_seconds": 75,
  "core_question": "How to make AI-generated meeting minutes more reliable",
  "source_items": [
    {
      "source_id": "S1",
      "title": "Internal anonymized demo notes",
      "type": "author-test",
      "verified": true
    }
  ],
  "must_include": [
    "AI only produces a first draft",
    "Verify owners and deadlines",
    "Sensitive material follows organizational rules"
  ],
  "must_avoid": [
    "Accuracy guarantees",
    "Unverified productivity multipliers",
    "Real customer information"
  ],
  "status": "research_ready"
}

You do not need many fields, but the fields must be stable. must_include and must_avoid are the two most important — they reduce stylistic drift each time the model generates.

Structured content card

Structured content card

You can also build a template for each content series. For example, a "method demo" format can be fixed as: problem, wrong approach, three-step method, limitation reminder, action suggestion. The model only fills in the content and never changes the structure.

Content state machine

Content state machine

4. When Designing a Workflow, Draw the Failure Paths First

Many automation tutorials only draw the happy path: trigger, generate, save, publish. Real systems more often run into empty sources, output that is not JSON, generation timeouts, exhausted quotas, duplicate titles, or scripts that violate series requirements.

In n8n, every critical node must answer four questions:

  • What happens when the input does not exist?
  • How many retries when the model returns a malformed response?
  • Who gets notified after consecutive failures?
  • Do already-successful steps need to run again?

Failure path design

Failure path design

A robust text workflow can be designed like this:

Schedule Trigger
  ↓
Read pending topics
  ↓
Check whether source_items is empty
  ├─ Yes → mark need_sources → notify the editor
  └─ No
       ↓
    Call the model to generate JSON
       ↓
    Validate against the JSON Schema
       ├─ Fail → append a "fix format only" prompt → retry at most 2 times
       └─ Pass
            ↓
         Fact and banned-word checks
            ├─ Fail → review_required
            └─ Pass → save as draft_v1
                         ↓
                      Send to the review queue

The key concept is idempotency: when the same topic_id runs again, the system should not unconditionally generate a second identical set of content. Use topic_id + stage + version as a unique key and check whether a successful result already exists before running.

JSON validation and retry

JSON validation and retry

5. The First Automation: Generate a Daily Candidate Topic Package

n8n node canvas (illustrative)

n8n node canvas (illustrative)

1. Scheduled trigger

Run at a fixed time every day, but do not chase minute-level frequency from the start. Content production is not a trading system — stability matters more than frequency.

2. Read candidate materials

Sources can be a manually curated link sheet, a list of public materials, questions from comments, or an internal topic library. Do not let the system scrape all trending content without boundaries. The wider the sources, the higher the noise, copyright, and factual risk.

The candidate material sheet should contain at least: title, summary, source name, publish date, original link, whether it has been read, whether it is cleared for use, and notes.

Candidate material sheet

Candidate material sheet

3. Deduplication

Deduplicate with simple rules first, then call the model for semantic merging. Simple rules include normalizing titles, stripping tracking parameters, and comparing source links and publish dates. The model's job is to spot cases where "different titles are actually about the same thing."

You are a topic editor. Perform semantic deduplication on the candidate materials.

Rules:
- Merge reposts of the same event into one topic
- Keep the source closest to the original publisher
- Do not treat items as different events just because the titles differ
- Do not add facts beyond the list
- Output JSON: topic_cluster, source_ids, difference, keep_reason

4. Scoring

Do not let the model output a single "virality score." The scoring criteria must map to your account's goals.

Score the topic on the following dimensions, 0-5 points each, with a one-sentence reason:

1. Fit with the account's positioning
2. Whether the audience's problem is specific
3. Whether reliable, verifiable sources exist
4. Whether it can be explained clearly within 90 seconds
5. Whether it can produce distinct visuals
6. Whether there is high risk or overpromising

The total is not a simple sum: if "reliable sources" is below 3, mark it as not proceeding to the drafting stage.
Output strict JSON with no additional explanatory text.

Topic scoring card

Topic scoring card

5. Human selection

Scoring is a ranking tool, not the final judgment. The review card the editor sees should include sources, the core question, risk items, and the recommendation rationale — not just a number.

6. The Second Automation: Generate a "Content Package," Not an Isolated Script

A production-ready content package contains at least:

{
  "title_candidates": [],
  "hook_candidates": [],
  "script": "",
  "subtitle_script": [],
  "shot_list": [],
  "cover_copy": [],
  "risk_flags": [],
  "source_map": [],
  "editor_questions": []
}

Split the generation prompt into two steps. In the first step the model only plans; only in the second does it write the actual copy. This makes it much easier to intercept a wrong direction at the planning stage.

Content package structure

Content package structure

Planning prompt

You are the managing editor of a knowledge-focused short-video channel. Design a content plan based on the content card. Do not write the full script yet.

Must follow:
- Only use information that can be confirmed from source_items
- Do not write income or results guarantees
- There can be only one core conclusion
- Link every key fact to a source_id
- Put anything unconfirmable into editor_questions

Output JSON:
angle, audience_problem, one_sentence_conclusion, structure, source_map, risk_flags, editor_questions

Drafting prompt

Generate the content package based on the approved plan.

Requirements:
- A 75-second vertical knowledge video
- Conversational tone; keep each sentence under roughly 22 Chinese characters (as an English adaptation, roughly 12-15 words per sentence)
- Open by going straight into the problem
- Provide three concrete actions in the middle
- Include at least one limitation reminder
- End with a low-friction action the viewer can try
- Output must conform to the specified JSON Schema
- Do not add facts that are not in source_map

Have the model output structured JSON so that downstream nodes can write titles into the sheet, feed the shot list into the visual pipeline, and send risk items to the reviewer.

7. Add a Human Review Gate That Actually Works

A human review gate is not "send a message and ask the editor to take a look." The review interface should let a person decide quickly.

Recommended display:

  • A one-sentence description of the topic
  • The source list with publish dates
  • A script readable within 60 seconds
  • Risks the model has proactively flagged
  • Similarity against the last 30 pieces of content
  • Estimated text, image, and video costs
  • Three buttons: approve, send back, abandon

When sending back, a reason must be selected, for example "insufficient sources," "the argument does not hold," "the title overpromises," "the visuals are not feasible," or "duplicates recent content." These reasons feed into a rule library and are automatically used as negative examples in the next generation round.

Human review queue (illustrative)

Human review queue (illustrative)
graph TD
    A[Draft enters review] --> B{Editor decides}
    B -->|Approve| C[Lock the script version]
    B -->|Send back| D[Select a send-back reason]
    B -->|Abandon| E[Archive and stop spending]
    D --> F[Generate a revised version]
    F --> A

8. Visual Automation Should Come After Text Automation

The per-unit generation cost, wait time, and failure rate of images and video are usually higher than text. If you start generating 12 shots before the script has passed review, you are spending budget on a plan that may be abandoned.

Use a two-stage visual pipeline:

  1. Low-cost preview. First generate the shot list, composition descriptions, and a few keyframes.
  2. Production after locking. Only after the copy and visual direction are approved, generate high-quality images and short clips.

Two-stage visual pipeline

Two-stage visual pipeline

Image prompts can be assembled from structured fields:

Subject: {subject}
Action: {action}
Scene: {scene}
Composition: {composition}
Lighting: {lighting}
Palette: {palette}
Style: {style}
Aspect ratio: 9:16
Constraints: no generated text, no brand logos, no unrelated people

Image prompt builder

Image prompt builder

Video prompts should separate "what the frame shows" from "how it moves":

Reference frame: a modern office desk, with a computer displaying a structured meeting table.
Motion: the camera slowly pushes in; the table highlight moves from "Conclusions" to "Action items"; objects on the desk stay stable.
Duration: 5 seconds.
Constraints: do not change the on-screen text structure, do not add people, no severe distortion.

Video capabilities in Kling, Hailuo, Jimeng, and Volcengine Ark will keep evolving. Your workflow should save "vendor, model, version, parameters, generation time, cost, output file" as metadata, making it easy to swap tools and compare results.

Generation metadata card

Generation metadata card

9. Cost Control: Set the Ceiling Before You Start Generating

Do not wait for the end-of-month bill to calculate costs. Every content package should have a budget before it runs.

A simple formula:

Estimated cost per piece
= text call cost
+ image attempts x cost per image
+ video seconds x cost per video second
+ voiceover characters x cost per audio unit
+ storage and automation execution cost

In practice, prices follow each platform's latest pages. The system should only store a "price table version" and a "budget cap" — do not scatter unit prices across individual nodes.

Set three layers of limits:

  • Per-step limit: how many times a single shot may be retried.
  • Per-content limit: pause when the total cost of one piece hits the cap.
  • Daily limit: stop new tasks when the day's total budget hits the cap.

Three-layer budget gates

Three-layer budget gates

Changes when upgrading to the professional route

Changes when upgrading to the professional route

Low-budget teams can use "text automation + manual visual selection"; medium budgets can use "low-resolution previews + high-quality generation for key shots only"; large-scale video generation is only appropriate once a content series is validated and its revenue or business value is clear.

10. Keys, Privacy, and Asset Permissions

The professional route touches API keys, internal materials, and team accounts, so security rules must be written into the process.

  • Store API keys in n8n credential management or secure environment variables — never in workflow notes, spreadsheets, or chat logs.
  • Use different keys and budgets for test and production environments.
  • Do not log sensitive raw text in full; when needed, store only hashes, IDs, or anonymized summaries.
  • Be explicit about which materials may be sent to third-party models and which must stay internal.
  • Record the origin and license status of images, voices, music, and reference materials.
  • When removing a team member's account, revoke their access to the automation platform, model platforms, and storage at the same time.

For voiceovers and human likenesses in particular, do not assume that "it can be generated" means "it can be used commercially in public." Check the specific terms of service, license scope, and platform policies before use.

Credential and data boundaries

Credential and data boundaries

11. Publishing Should Not Be the Default Final Node

Many people treat "auto-publish" as the endpoint of the workflow. For knowledge and brand content, a more reasonable endpoint is "generate a publish-ready package and notify the owner." The publishing action should meet four conditions:

  1. The script version is locked.
  2. Facts and sources have passed review.
  3. Asset permissions are clear.
  4. The owner has actively approved.

If the target platform offers an official publishing API, follow the API and platform rules; if there is no stable, compliant interface, do not rely on fragile schemes like simulated clicks. The value of automation comes mainly from organizing, generating, archiving, and notifying — do not sacrifice reliability just to make the last step automatic.

12. Build a Data Feedback Loop, Not Just Content Output

Professional-route retrospective loop

Professional-route retrospective loop

After each piece is published, write the performance data and human evaluation back to the same topic_id. Record at least:

{
  "topic_id": "20260720-001",
  "published_at": "2026-07-20T10:00:00+08:00",
  "platform": "primary-platform",
  "title_used": "What AI most often misses when organizing minutes",
  "hook_version": "H2",
  "duration_seconds": 78,
  "views": null,
  "average_watch_seconds": null,
  "completion_rate": null,
  "saves": null,
  "comments": null,
  "editor_score": 4,
  "next_change": "Shorten the opening; show the checklist earlier"
}

Platform data tells you "what happened"; the editorial retrospective should explain "why it might have happened." Do not let the model draw definitive conclusions from a small amount of data. It can generate hypotheses, but the next round should test only one change.

13. Six Reusable Professional Prompts

1. JSON fixer

Only fix the JSON formatting of the content below so it conforms to the given Schema. Do not rewrite the meaning of any field and do not add new facts. If it cannot be fixed, return error with the reason.

2. Source mapping

Link every factual sentence in the script to a source_id. Mark sentences that cannot be linked as unsupported; do not search for new sources.

3. Duplication check

Compare the new script against the last 30 scripts and point out overlaps in topic, structure, opening, and examples. Give minimal differentiation suggestions.

4. Shot planning within budget

The total video generation budget is {budget}, with at most {max_clips} dynamic shots. Prioritize the frames that most need motion; use screen recordings, subtitle cards, or static images for the rest.

5. Learning from send-back reasons

Based on the last 20 human send-back reasons, summarize five actionable rules. The rules must be checkable before generation; avoid vague statements.

6. Publish package check

Check whether the publish package contains: the locked script, subtitles, cover, asset license records, source mapping, risk notes, and owner approval. If any item is missing, return blocked.

14. Common Failures and How to Fix Them

The workflow keeps growing and nobody dares change it

Split the process into five sub-workflows — research, writing, visuals, final cut, and data — connected by explicit data structures. Each sub-workflow can be tested and rerun independently.

The same topic gets billed repeatedly

Give tasks a unique key; query the status before running; reuse the results of successful steps by default, and only produce a new version when a human explicitly chooses "regenerate."

The model keeps refusing to output JSON

Reduce the number of fields per output; validate with a JSON Schema; on the first failure, only ask it to fix the format rather than re-create; hand off to a human after consecutive failures.

Auto-generated titles keep getting more exaggerated

Put banned words, title-delivery rules, and recent negative examples into the system prompt; also record the title the editor ultimately chose, not just the model's candidates.

Video costs spiral out of control

Lock the script first, preview first, cap retries, prefer screen recordings and static images, and use video generation only for shots that genuinely need motion.

The team does not know who owns what

Assign exactly one owner per status. For example, research_ready belongs to the research editor, script_review to the managing editor, and asset_review to the visual editor.

15. FAQ

1. Does an individual creator need n8n?

Not if you only produce one or two pieces a week and copy-pasting is not a burden. Introduce it once you face repeated input, version chaos, and batch-processing needs.

2. Which should I buy first, a subscription or an API?

Use a subscription first when you need lots of manual discussion and revision; use an API when you need structured batch calls. Many teams use both, but for different responsibilities.

3. Should I connect image and video APIs from the start?

Not recommended. Get the text pipeline, review process, and data storage running reliably first, then add the more expensive media steps.

4. Can automation guarantee consistent content quality?

It can guarantee consistent formats, processes, and checklists — it cannot guarantee that every topic is worth publishing. Editorial judgment still requires a human.

5. How do I choose a text model?

Test with 20 of your own real tasks: accuracy, structured output, conversational Chinese, latency, cost, and stability. Do not rely on public leaderboards alone.

6. Should I call multiple models at the same time?

High-risk fact-checking or important commercial content can use multi-model cross-checks; ordinary rewriting tasks do not need unconditional multi-model calls, which only add cost and complexity.

7. Can the system auto-reply to comments?

It can generate candidate replies, but complaints, privacy, transactions, disputes, and brand-position issues should be handled by humans. Do not feed personal data directly into the model.

8. Does self-hosting n8n mean my data is absolutely safe?

No. Self-hosting only changes where things run; you still need to manage updates, permissions, backups, logs, keys, and network security.

9. How do I judge whether automation is worth it?

Track the weekly human time spent on repeated steps, the error rate, and the waiting time. The value saved by automation should exceed the cost of subscriptions, API calls, maintenance, and review.

10. What are the most important monitoring metrics?

Beyond execution success rate, watch the human send-back rate, cost per piece, average number of revisions, missing-source rate, and the time from topic selection to a locked script.

Closing: The Goal of a Professional System Is Not "No Humans Involved," but "Humans Only Make the Key Judgments"

A mature content factory does not make editors move text around every day, nor does it let the model decide on its own what deserves to be published. The system handles triggering, organizing, generating candidates, saving versions, calculating budgets, and sending reminders; humans handle choosing topics, confirming facts, forming viewpoints, managing risks, and approving publication.

Once this pipeline runs reliably, tools can be swapped and models can be upgraded while the content series keeps running. The real asset is not the chat history inside some account — it is your content structure, review rules, source library, visual standards, and retrospective data.

Illustration Production Checklist (20 items)

All 20 illustrations in the table below are embedded in the body as SVG diagrams (interface-style images are process mockups, not real product screenshots). For formal publication, the "screenshot-style" mockups can be replaced with real screenshots, which should hide accounts, keys, private materials, and unauthorized content.

No.VisualFormatSuggested placementProduction notes
01Article coverExisting SVGOpeningEmphasize "stable workflow"
02End-to-end workflowExisting SVGAfter the introductionTrigger, generate, review, archive
03Tool stack layersExisting SVGTools sectionOrchestration, text, visuals, video, and final cut
04Subscription/API/automation comparisonThree-column infographic, embedded as 02-cost-types.svgSection 1Explain the responsibilities of the three cost types
05Structured content cardJSON visualization, embedded as 02-content-card.svgSection 3Render fields as a card rather than a large code block
06Content state machineFlowchart, embedded as 02-state-machine.svgSection 3State transitions from research_ready to published
07Failure pathsBranching flowchart, embedded as 02-failure-paths.svgSection 4Empty sources, malformed output, insufficient budget
08n8n node canvasScreenshot-style, embedded as 02-n8n-canvas.svgSection 5Show only the essential nodes and hide credentials
09Candidate material sheetTable screenshot, embedded as 02-source-pool.svgSection 5Sources, dates, licenses, and read status
10Topic scoring cardRadar or scorecard, embedded as 02-scoring-card.svgSection 5Fit, evidence, duration, and risk
11Review queueInterface mockup, embedded as 02-review-queue.svgSection 7Three actions: approve, send back, abandon
12Content package structureTree diagram, embedded as 02-content-package.svgSection 6Titles, script, subtitles, shots, and risks
13JSON validation and retryFlowchart, embedded as 02-json-retry.svgSection 4At most two format-fix attempts
14Two-stage visualsSide-by-side comparison, embedded as 02-visual-two-stage.svgSection 8Low-cost preview vs. production after locking
15Prompt builderField cards, embedded as 02-prompt-builder.svgSection 8Subject, action, scene, lighting, and aspect ratio
16Generation metadata cardInfo card, embedded as 02-metadata-card.svgSection 8Model, parameters, version, cost, and files
17Three-layer budget gatesDashboard mockup, embedded as 02-budget-gates.svgSection 9Per-step, per-content, and daily budgets
18Credential and data boundariesSecurity diagram, embedded as 02-security-boundary.svgSection 10Keys, logs, and anonymized data
19Upgrade changesExisting SVGCost sectionFrom manual shuffling to structured collaboration
20Data feedback loopExisting SVGBefore the closingPerformance, human retrospectives, and rule updates

Source Verification Note

The product capabilities mentioned in this article were verified against the following official materials: DeepSeek API Docs, Volcengine Ark model and API documentation, the official Jimeng AI platform, the official Jianying website, and the official n8n documentation. Model names, prices, quotas, and interfaces may change; when implementing, always defer to the latest official pages and terms of service.