Building an AI writing and content creation pipeline from scratch: a nanny-level tutorial

🛒 For newcomers to content operations and self-media, a complete operation guide from registration tool to the completion of the first article.

Tutorial objectives and applicable readers

This tutorial takes you step by step to build an "AI writing and content creation pipeline" from scratch: registering tools, establishing brand corpus, writing standard prompt words, generating articles, quality inspection and publishing. The entire process does not rely on programming skills. Any content operator, market editor, or self-media manager can pass the first page by following these steps. The full text is expected to take 60-90 minutes to complete.

1. Preparation Checklist

Please confirm each item before starting to avoid getting stuck midway:

  • [ ] Have registered an available large model account (choose one of ChatGPT / Claude / DeepSeek, this tutorial uses DeepSeek as the default example, free and Chinese-friendly)
  • [ ] An online search AI (Kimi or Secret Tower AI search) has been registered for material sorting
  • [ ] Prepare 3-5 high-quality articles on team history (as style corpus)
  • [ ] Prepare a list of brand tone keywords (eg: professional, restrained, data-driven)
  • [ ] Create a unified material directory (local folder or Notion page)
  • [ ] Clarify the first topic to be produced (it is recommended to choose the keywords of "medium and high frequency, moderate competition")

2. Account and environment preparation

  1. Open the DeepSeek official website, complete registration and log in, and enter the dialogue interface.
  2. If possible, turn on the "Internet search" switch in the settings (different versions have different locations, if you can't find it, just say "Please search online" in the conversation).
  3. Open Kimi (kimi.moonshot.cn) to register and log in to organize materials in batches.
  4. Create a "Content Production" folder, which is divided into four subdirectories: 01-Topic Selection, 02-Draft, 03-Quality Inspection, 04-Published. Each link in the pipeline corresponds to a directory for easy traceability.

3. Establish a brand corpus

No matter how powerful the AI is, it does not know your brand tone and will only produce "general text". The corpus is the foundation of the pipeline.

  1. Paste 3-5 high-quality historical articles into the Kimi conversation and enter the following command:
Please analyze the common style characteristics of the following articles and output:
1. Sentence pattern preference (proportion of long and short sentences, questions about whether you like to use them)
2. Characteristics of word usage (density of professional terms, degree of colloquialism)
3. Tone and stance (authoritative/friend/neutral)
4. Structural habits (total scores, cases first, etc.)
5. 5 reusable high-frequency expressions
  1. Organize the analysis results into a "Brand Tonal Baseline" of less than 200 words, and save it to Tonal Baseline.md under the 01-Topic Selection directory. All subsequent prompt words will reference this baseline.

4. Design standardized prompt word templates

A reusable "article generation prompt word" looks like this (copy directly and replace the content in brackets):

# role
You are the senior content editor of {brand name}, and your writing style strictly follows the following tonal baseline:
{Paste tonal baseline}

# Task
Write an article for {target readers} around the topic "{Topic}".

# Requirements
1. First output 5 title candidates (including 1 numeric type and 1 question type)
2. Output the article outline (including H2/H3 levels)
3. After confirming the outline, expand the text section by section.
4. Label each data/reference point [to be verified]
5. The full text should be 1200-1800 words, avoid empty words and clichés

#material
{Paste the reference materials compiled by Kimi}

Four elements of prompt words: role (putting the model into the state), task (clear goals), requirements (acceptable constraints), and material (feed the context). If one is missing, the output quality will decline significantly.

5. Use large models to generate articles

  1. Enter the topic into Kimi, let it sort out more than 10 reference materials online, output a summary of key points, and save it to 02-Draft/Material.md.
  2. Fill in the tonal baseline and material into the prompt word template above and paste it into the DeepSeek conversation.
  3. First let it produce only "Title Candidates + Outline", manually select a group and then let it expand section by section.
  4. If you are not satisfied after each section is expanded, use additional instructions to fine-tune it, for example:
The example in the third section is too dry. Change it to a real business scenario and add a specific number.
  1. After generating the full text, let the model "review" it by itself:
Please review the full text as a critical editor: pick out 3 logical breaks, 3 empty words, and 2 points that can add data, and give suggestions for modification.
  1. Save the satisfactory version as 02-draft/first draft.md. The first draft usually only takes 15-30 minutes at this point.

6. Manual quality inspection and SEO optimization

  1. Fact Check: Check all data, references and cases marked [to be verified] in the article item by item. If there are errors or cannot be verified, delete or rewrite them directly. This is a gate that cannot be skipped before publishing.
  2. Original check: Use the free version of Grammarly or a duplication check tool to do a spelling/grammar check; important content can be randomly checked for originality to ensure that the manual increment accounts for more than 30%.
  3. SEO Optimization: Write core keywords into the title, first paragraph, an H2 title and the end; the title should be 20-30 words; the abstract should be 120-150 words with keywords.
  4. Image: Use Canva free templates or AI-generated images (please refer to the AI ​​design and illustration tutorial) to complete the header image and illustrations.
  5. Archive to 03-Quality Inspection/Finalization.md and record the verifier.

7. Verification method

After running through, please verify whether the assembly line is qualified according to the following standards:

  • First draft time: ≤30 minutes from topic selection to first draft for a single article
  • Quality inspection pass rate: ≥4 out of 5 consecutive articles pass three levels (facts/tonality/originality) at one time
  • Style consistency: Let 3 colleagues blindly read the new draft and the historical draft to recognize the same brand
  • Data feedback: Comparing the historical average reading/interaction 2 weeks after publication, there is no obvious decline.
  • After reaching the standard, you can batch copy to the next topic to form a daily rhythm (such as 3 articles per week).

8. Frequently Asked Questions and Troubleshooting (FAQ)

  1. The generated titles are always bland? Let the model "give 10 titles at a time, covering digital/question/suspense/pain point types", and then manually select them; or let it imitate the structure of the three popular headlines you gave.
  2. Models always write empty words like "in today's era"? The prompt clearly states that "it is prohibited to use vague beginnings such as 'In today's era, with the development of science and technology, it is well known' and go directly to the topic."
  3. How ​​to modify one draft for multiple platforms? For the same material, the public account version (long text + typesetting), Xiaohongshu version (short sentence + emoji + topic), and Zhihu version (professional + structured) are generated respectively, and processed in batches using rewriting instructions.
  4. Why does AI make up data? Large models will "complete" information that seems reasonable. Countermeasures: The prompt words must be marked [to be verified], and the editor will verify them one by one; the key figures involved shall be subject to official sources.
  5. The free version has a limit on the number of times it can be used? Use the free model to write drafts during the day and focus on manual quality inspection at night; or upgrade to the paid version and include it in your budget.
  6. Is the team afraid that the content will be too "AI-flavored"? It is mandatory to add 2-3 personal experiences/real cases/exclusive data during human review. Treat the AI ​​manuscript as an "intern who is good at writing" and you are the final review editor.

9. Advanced expansion

  • API batch pipeline: solidify the prompt word template into code call, and realize the semi-automatic pipeline of "keyword pool → automatic generation → automatic quality inspection reminder".
  • Knowledge Base Enhancement: Import historical articles into Notion AI or enterprise knowledge base to generate models based on internal data to reduce illusions.
  • Channel Differentiation Template: Create a set of exclusive prompt words for each official account/Xiaohongshu/Zhihu/SEO, and switch the tone with one click.
  • Integration of graphics and text: Access the AI ​​graphics process to realize "one text, one picture" and even automation of information graphics.

10. Practical Example: Completely walk through from topic selection to final draft

Based on the topic "How to maintain efficiency in remote working?" 》For example, string together the entire pipeline:

  1. Topic selection: Determine the target reader is "middle-level managers in hybrid office", and the keyword is "remote office efficiency".
  2. Material sorting: Enter "Practical methods for remote office efficiency, please sort out 10 sources online" in Kimi, get a summary of key points, and save it as material.
  3. Write prompt words: Fill in the tonal baseline, topic selection, and material according to the template in Chapter 4, and paste it into DeepSeek.
  4. Come out titles and outlines: Let the model give you 5 titles first, and select "Is telecommuting inefficient?" Change these 5 habits first” and then expand on the outline.
  5. Section-by-section generation: After confirming the outline, expand the text into sections, and add fine-tuning instructions at the end of each section, such as "Add a specific example of time block arrangement in the second section."
  6. Self-review: Let the model pick out empty words and logical breaks as an editor, and manually adopt the modifications.
  7. Quality inspection: Check the data and case sources in the article, Grammarly check the grammar, and confirm the original increment.
  8. SEO and Publishing: Write "Remote Office Efficiency" into the title and first paragraph, use Canva to match the header image, publish it to the official account and synchronize the rewritten version of Xiaohongshu.

After a complete walk, record the actual time and pitfalls of each link, return to the template and update it, so that it will be smoother next time. This cycle of "run once, record once, change once" is how the assembly line continues to evolve.

User Reviews

  • Loading reviews...