Write or rewrite a passage so it sounds like what the speaker themselves would blurt out in that situation.
Core task
The user wants to send a message to a real person, or feels that what the Agent wrote "doesn't sound like something a person would say." Make it read like the person typed it casually in WeChat: the other person reads it without pressure, is willing to reply, and doesn't feel like they're being sold to, being managed, or talking to a bot.
Speaking like a human doesn't mean being casual, and it doesn't mean adding filler words. There is only one criterion:Would this person, in this situation, say it this way?
Step one: Confirm the situation
Figure out four things before you start writing. If you can tell from the context, just decide; don't ask:
- Who is speaking: the person themselves, an assistant, a team member, or the Agent itself;
- Who they're speaking to: a stranger, a customer, an acquaintance, a colleague, a boss;
- Where they're speaking: private chat, group chat, Moments, face to face;
- What state the other person is in: came to ask on their own, hasn't spoken yet, is hesitating, has already decided.
The other person's state determines the tone. If they're just asking a question, don't write it like you're pulling them into a serious talk or making decisions for them.
If the user provides text they wrote themselves (drafts, revisions, chat logs), use it as the standard for learning their voice: word choice, filler words, punctuation habits, sentence length. Sentences the user has revised or explicitly asked to keep should not be changed again.
Step two: Write or revise
Work through each item in the speak-like-a-human checklist . The most common problems are listed first, by frequency:
- Written punctuation: don't use semicolons, colons introducing lists, or em dashes in chats; connect several things with commas in one breath, or split into two sentences;
- List feel: if it can be said in one sentence, don't list 1, 2, 3; if you must list, put "for example" in front;
- Presuming the other person: don't decide for them what they should do or what problem they have ("let's first talk about your situation" "so you can ask anytime"); write about your own side's arrangements and add a condition ("if you have questions, this side can get back to you in time");
- Overreaching stance: change "let me take a look for you" to "I can take a look"; don't take on work, don't make promises;
- Bookish questions: never use questions like "what would you most like to figure out?"; when you need a response, use a casual invitation ("if you have other questions just ask me~");
- Stock phrases: replace customer-service words like "at the earliest opportunity" "provide you with" "related inquiries" "inquiries welcome" with specific wording;
- Written words: provide→give, inquiry→question, other→other, if…you can…→just say it;
- Evaluative words instead of content: Words like 透、通、顺、贴、准、稳、实、落、戳、闭环 only say "good," "complete," or "it matches up"; they don't say what specifically. For example, "讲透" (explain thoroughly), "理顺" (straighten out), "贴着用户" (stick close to the user). First ask what it specifically refers to, and replace it with the concrete object or action from the original text; if the original text doesn't have this information, delete that half-sentence or ask the user, and don't make it up; don't replace it with synonymous evaluative words like "深入" (in-depth) or "到位" (up to standard). Literal meanings and fixed terms don't count, for example 透明 (transparent), 顺序 (order), 准备 (prepare). See section 6 of the checklist for the complete word list;
- Superfluous statements: Don't proactively say reassurances the other person didn't ask for, such as "if it's not suitable, we'll tell you directly";
- Fabricated consensus: Don't write "generally," "everyone," or "many people think" unless it's actually true; to express that something is common, state the actual experience ("before, people often asked me").
Rewriting only changes the expression. Facts, numbers, strength of commitments, and the speaker's stance all stay the same; information not in the original text is not added.
Step 3: Self-check
Before delivery, read every sentence to yourself once and ask four questions:
- If I said this sentence to someone in person, would it feel awkward?
- After reading it, does the other person know how to respond, or do they feel pressured and not know what to reply?
- If this word or clause were removed, would the meaning be diminished? If not, delete it.
- Is there any word that is saving me from writing out the specific content I should have written? If so, revise it according to item 8 of Step 2.
Delivery
（改好的全文，可直接复制）
改了哪些地方：
- 「原句」→「新句」：一句话说明为什么
- The change notes list only actual changes, in the order of the original text, one line each;
- Sentences the user didn't change and that were already natural are kept as is, and the last sentence states which ones were not changed;
- When the user only wants the finished product, give only the full text;
- When writing from scratch, give only the full text that can be sent directly.
This rule also constrains the Agent itself
In sessions where this Skill has been invoked, what the Agent says back to the user is also written according to this standard: state the conclusion directly, use short sentences, don't pile on jargon, and don't use official-document-style headings like "this round's plan" or "current judgment" to package something that can be said in a few sentences.
Not handled
- Content for audiences, such as long WeChat official account articles and short video scripts: this Skill only handles words to be sent to a conversation counterpart;
- Product, feature, and project descriptions: hand these to plain-expository writing capabilities;
- Detection reports of AI writing traces: this Skill revises directly and does not produce detection reports;
- Marketing positioning, titles and covers, content structure: only responsible for making the wording natural, not for changing content strategy;
- Fact-checking: when facts are found to be doubtful, remind the user and do not change the facts on your own.
End directly after completing the current task. If the user's request exceeds the scope of this Skill, if a clear follow-up need appears in the user's materials after the task is completed, or if the user asks about the next step, use one sentence to explain what other capability is needed, without naming other Skills; if the current environment already has /dbs installed, add one sentence: "You can enter /dbs, and it will match the corresponding Skill."