Give an AI tool a product sheet and ask for “viral copy,” and it will probably return something polished, enthusiastic, and interchangeable with hundreds of other posts. It may also fill the gaps with a personal reaction, a performance number, or a customer quote that nobody supplied.
AI is useful here, but its job needs to stay narrow. It can organize material, suggest a structure, and revise language. It cannot use a product on the author's behalf or prove a marketing claim. A longer prompt does not transfer that responsibility to the model.
The workflow below preserves the seven-stage sequence in the source material: define the category, submit product details, choose a tone, choose a structure, draft titles, create the post, and revise it. I added a product evidence pack before generation and a publication review after drafting. The goal is not to promise reach. It is to make every claim traceable.
Build a product evidence pack before prompting
A product evidence pack is not a page of sales language. It is the set of facts the model is allowed to use:
- Product name, specification, material, ingredients, or relevant technology.
- Intended users, usage conditions, and known limitations.
- Instructions, price, sales channel, and the date those details were checked.
- The author's actual usage period and any original notes or photographs.
- Sources and publication rights for tests, comparisons, and customer feedback.
Write “unknown” when the information is missing. Do not ask the model to complete it.
“The texture felt light during my recorded test” can be a first-hand observation. “Visible improvement in three days” and “customers love it” are claims that need evidence. Without test conditions, source material, or a reproducible record, they do not belong in a publishable draft.
This also prevents a common failure: turning brand copy into a first-person review. If the author has not used the product, the post cannot say, “I tested it for a month.” Write a factual product explainer or wait for real usage notes.
Step 1: Define the category and reader
“Beauty products for women aged 20–35” leaves most of the task undefined. The model will fill that space with familiar platform clichés.
Describe a concrete problem instead: who the reader is, where the product is used, what constraint matters, and what they have already tried. “Office workers who take short trips with carry-on luggage and want to reduce leaks from repacked toiletries” gives the draft a job to do. It does not invent a personal story.
Record the reader and context. Do not draft the post yet.
Category: travel toiletries.
Reader: office workers who take short trips with carry-on luggage.
Problem: limited capacity, leaking containers, and repeated repacking.
Keep missing information unknown. Do not infer it.
Step 2: Submit the product description
The source checklist contains ten useful fields: name and brand, features, instructions, usage experience, intended users, packaging, price and channel, customer feedback, competitor comparison, and before-and-after results.
The rule is simple: include a field only when there is evidence for it. Add a source label such as packaging, manual, official product page, purchase record, author test, licensed customer feedback, or independent test report.
Experience, feedback, comparisons, and outcome claims deserve more scrutiny than dimensions or packaging. Preserve warnings and test conditions with the favorable details. If two sources disagree, record the conflict and pause the draft.
Step 3: Choose a tone
The original method offers five tones: humorous, educational, empathetic, practical, and personal. Tone changes presentation, not the strength of a claim.
Humor should not target a group of people. Educational copy needs sources and cannot turn an ingredient name into a guaranteed result. Emotional writing cannot fabricate a family or health story. A practical tone works well for instructions. A personal tone requires personal use.
“Make it relatable” is too vague to control the output. A better instruction is: “Use short sentences, limit exclamation marks, avoid platform slang, and remove subjective outcomes that lack evidence.”
Step 4: Choose a structure that matches the evidence
One evidence pack can support several formats. The available material should decide which one is honest.
- First-person review requires actual usage notes and a time period.
- Professional testing requires a repeatable method, consistent conditions, and recorded results.
- A scenario can be hypothetical, but it must not masquerade as the author's experience.
- A data-led post needs a source, date, and scope for every number.
- A tutorial needs stable steps, misuse warnings, and acceptance checks.
Choosing a test format without test data invites fabricated precision. Choosing a personal review without personal use turns product claims into a fictional memory. Better prose will only make those defects harder to notice.
Step 5: Generate and screen title options
Asking for five titles is a useful way to explore angles. Do not select one on curiosity alone.
- Does every promised result appear in the evidence pack?
- Can words such as “best,” “must-buy,” or “instant” be supported?
- Does the title manufacture anxiety or an unsupported risk?
- Do emojis improve scanning, or are they disguising a problematic term?
Pinyin, homophones, and emojis do not make a restricted or misleading claim compliant. Remove the claim, reduce its certainty, or provide valid evidence.
Step 6: Generate the draft
Pages of role labels and capability claims do not teach the model more about the product. OpenAI's public prompting guidance recommends clear instructions, relevant context, and iterative refinement. I prefer a compact prompt that can be audited line by line.
Task: Draft a Xiaohongshu product recommendation post of about 500 Chinese characters.
Reader and context: Use the confirmed information below.
Facts: Use only the product evidence pack.
Tone: Practical, natural, short sentences, few exclamation marks.
Structure: Problem context, product facts, instructions, suitability, review reminder.
Do not invent: usage experience, reviews, test data, price, performance, sources, or purchase outcomes.
Output: Five title options, then the post. List missing information at the end instead of filling it in.
The same input pattern works across ChatGPT, Kimi, Doubao, and other conversational tools. Model names, subscription terms, and button locations change. The method does not depend on a particular interface.
Step 7: Revise with testable feedback
“Write another version” and “make it more viral” are not useful review instructions. They tend to add excitement, slang, and punctuation without fixing the evidence.
Point to a defect in the text:
- Remove performance statements that have no source.
- Cut the opening to 80 Chinese characters while preserving the travel context.
- Replace “suitable for everyone” with the documented usage conditions.
- Check each instruction against the product manual.
- Mark every sentence that still needs author confirmation.
Change one type of issue per pass when comparison matters. A longer draft is not necessarily a better one. Removing an unsupported sentence can be the most valuable edit.
Run four checks before publishing
A finished draft is not automatically ready to post.
Assess whether the content is advertising based on the actual commercial relationship. China's Measures for the Administration of Internet Advertising require online ads to be recognizable. Product promotion presented as knowledge, experience sharing, or consumer testing and accompanied by a purchase method must carry a prominent advertising label. Advertisers are responsible for the truthfulness of the content. A personal-looking post is not automatically outside advertising rules.
Keep AI declarations and platform labels intact. China's Measures for Labeling AI-Generated and Synthetic Content took effect on September 1, 2025. Users publishing generated or synthetic content are required to declare it and use the labeling function provided by the distribution service. Existing labels must not be maliciously removed, altered, hidden, or forged.
Trace every outcome, number, and quotation back to a record. If the source cannot be produced, remove the statement or rewrite it as neutral product information without a promised result.
Check privacy and permissions. Usernames, order numbers, addresses, private messages, other people's photographs, third-party screenshots, and unlicensed reviews should not be copied into a post casually. Medical products, pharmaceuticals, health foods, cosmetics, and other regulated categories may require additional review beyond this general checklist.
Let AI organize; keep human sign-off
Reliable recommendation copy does not come from a secret prompt. It starts with real product information, uses a structure that the evidence can support, and ends with a publication decision made by a person.
AI can make that process faster. The author still decides whether a fact is true, whether a source may be published, whether the post is advertising, and whether the final sentence should go live.
For a review of an AI-assisted publishing workflow, provide the product evidence pack, target reader, usage context, publishing platform, and applicable industry rules. Clear inputs leave less room for the model to guess.