Many people understand “reducing AI flavor” as one thing: run a detection tool and see whether the score goes down. But anyone who has actually written a draft knows that the score itself cannot help you revise. What is truly useful is locating the problem—which sentence, which sentence pattern, which word makes the text look machine-written. MuchSci’s Personal Agent Assistant takes the latter path: instead of giving a score, it first diagnoses, then locates, then rewrites item by item.
The key to understanding it is to separate “detection” from “rewriting.” Detection judges risk; rewriting is the actual hands-on fix. This tool does the latter, and it breaks rewriting into a process completed by multiple agents working together.
01 Reducing AI flavor is not about “detection”; it is about “sentence-pattern fingerprints”
Let us start with a common misunderstanding: whether the AI flavor is strong often depends not on the content but on sentence-pattern fingerprints.
▲ Figure 1 | Machine traces usually lie not in the content but in expression habits at the sentence-pattern level
If a paper’s research direction, core argument, and experimental data are provided by the author, there is usually no problem at the content level. But at the sentence-pattern level, machine preferences easily appear: too many em dashes, highly repetitive connectives, and clusters of judgment sentences. These features do not affect the factual correctness of the paper, yet they give readers the intuition that “this passage was not written by a human.”
So the first step in rewriting is to separate “whether the content is correct” from “whether the sentence patterns look human.” The Personal Agent Assistant handles the latter—repairs at the level of sentence patterns and word choice. It does not promise any detection result; it only makes the text more like something written by a human in terms of expression.
02 The Personal Agent Assistant’s execution chain
Open the MuchSci Personal Agent Assistant interface, and you can see the complete execution chain of a rewriting task.
▲ Figure 2 | The execution chain is visible: after Exec and TodoWrite are completed, a docx and a preview pdf are delivered
The screenshot shows a real task: first execute `Exec`, then execute `TodoWrite`, and both steps are marked “completed.” Two files are then delivered—the main deliverable is `/data/..._降AIGC版.docx`, and the other is a rendered preview pdf for checking the layout. This process visibility makes rewriting not a black box but a procedure that can be reviewed.
Next, the tool gives a diagnostic conclusion: “First, the diagnosis: where the AI flavor in this piece is concentrated.” It points out that “the most typical machine trace is not the content but the sentence-pattern fingerprint,” and lists the high-risk items. The specific item visible in the screenshot is “Em dash overload: 13 occurrences of —— in the full text, almost every paragraph uses it for a ‘follow-up jab’ extension.”
03 How multiple agents divide the work
The Personal Agent Assistant does not rely on a single agent to do everything; instead, it splits the task among several roles.
▲ Figure 3 | Diagnostic, rewriting, and review roles divide the work to reduce self-confirmation bias
The diagnostic agent is responsible for reading the original text, identifying sentence-pattern fingerprints, and listing the high-risk items. Its output is not a score but a table: what the trace is and where it appears in the original text.The rewriting agent takes the list and handles it item by item: splitting overly long sentences, replacing high-frequency connectives, and softening absolute statements.The review agent rescans after the rewrite to confirm whether the high-risk items have dropped to zero and whether the meaning has been distorted.
The value of this division of labor is that diagnosis and rewriting are not done by the same role, so the bias of “checking oneself” is less likely. The review stage can also surface unresolved issues for the author to judge.
04 The method of “diagnose first, then locate, then rewrite item by item”
Breaking the process down a little further, it is roughly the following four steps.
▲ Figure 4 | The key to reducing AI flavor is not the detection score but locating the specific sentence and then revising it
The first step is to read the original text. What is read here is not the topic but statistical features at the sentence-pattern level: em dash frequency, connective distribution, and judgment-sentence density.The second step is to provide a diagnosis. The tool lists high-risk items, such as “Em dash overload: 13 occurrences in the full text” in the screenshot.The third step is to revise item by item. For each high-risk item, the tool rewrites the relevant sentences, preserving the original meaning while changing the expression.The fourth step is to review again. After revision, it reassesses to confirm that the traces have significantly decreased or dropped to zero.
To illustrate the diagnostic results, a table can be added to the main text:
| Trace | Manifestation in original text |
| Em dash overload | 13 occurrences in the full text; almost every paragraph uses it for a “follow-up jab” extension |
| Transition-word pileup | Connectives such as “first,” “second,” and “in addition” appear at high frequency |
| Clustering of judgment sentences | Many expressions of the “is” and “means” type |
The point of this table is not to scare people but to translate the abstract “AI flavor” into specific problems that can be located and revised.
05 Boundaries
As usual, let us put the boundaries on the table.
① This article discusses the functional mechanism, not a promise of results. The Personal Agent Assistant performs repairs at the sentence-pattern and word-choice levels and cannot guarantee any detection result; different detection tools use different criteria, and their results may also be inconsistent.
② It does not change the truthfulness of the content. Whether the research question, experimental data, and cited literature are reliable remains the author’s responsibility. The tool only handles expression and does not touch the substantive judgments of the paper.
③ The permissible scope of AI participation in paper writing is subject to the current regulations of the user’s school or institution. In scenarios involving academic integrity, users are advised to confirm the relevant regulations in advance and not to submit directly based only on tool output.