I Configured WorkBuddy as a Paper Revision Pipeline to Reduce AI Risk, But Not to Edit Content

Last night my advisor threw out another comment: Chapter 3 has too much "AI flavor," don't touch the plagiarism check rate. I came across a project on GitHub called academic-paper-writer-pro-2, which supposedly added an AI-detection reduction pipeline in its enhanced version. It can rewrite Chinese text and de-AI English LaTeX (LaTeX is a typesetting language commonly used for papers), and emphasizes structured processing of Word and PDF files. I'm in my third year of master's studies. I've been using WorkBuddy for about a month. Usually, I only use it to organize literature packages. This time, I didn't dare to let it directly edit; I only configured it as a "risk pipeline," silently telling myself I'm just referencing it, really.

My environment consists of local thesis files, advisor-annotated Word docs, and CNKI plagiarism check report PDFs, all placed in D:/thesis/revision. Open WorkBuddy, click Workflow on the homepage, then click New, select the Document Processing template, and name it Thesis Revision Package. A workflow is a series of steps, each step called a node, like workstations on an assembly line. Drag in the first Local File node, set the path to D:/thesis/revision, and check Read Only Specified Directory on the right; do not select the entire D drive. A few days ago, I wrote about a local folder path being unreachable, and later found out it was mostly because the authorization scope wasn't set correctly. Here, the interface should display "Connected 3 files." If it shows 0, check if the directory is occupied by cloud sync.

Add Annotation Extraction as the second node. It separates advisor annotations in Word into a table, with fields generally including Location, Original Text, Annotation, and Revision Requirement. I've only been using Word for three days, so the annotation formats are messy; some say "This sentence is too templated," others just say "Rewrite." I let WorkBuddy only extract, not generate body text, outputting it as Advisor Opinion Table. This step is crucial. The most annoying thing about AI collaboration is fragmented output. Structuring opinions first prevents spending ages piecing them together later.

The third node is Risk Marking. Input connects to Advisor Opinion Table and CNKI Plagiarism Report. I set keywords to Repeated Fragments, Suspected Templated Sentences, Excessive Connectors, and Long Attributive Clauses. WorkBuddy provides candidate locations by paragraph, such as "Three consecutive passive sentences here" or "This sentence overlaps with a plagiarism fragment." It doesn't replace directly; it only generates a Suggested Draft. The suggested draft is just a hint, not the final draft. Every time I export the plagiarism report, I look at the high-risk pages first, like a conditioned reflex.

The fourth node is Version Comparison, which finds differences between two file versions paragraph by paragraph and outputs a Revision Comparison Table. I specifically turned off Auto Overwrite Original because once overwritten, if my advisor asks me to revert to the previous version, I'd have to rely on memory. This setting is under Output Options; the default is to generate a new file, don't change it to overwrite.

I also implemented three layers of permissions. In the WorkBuddy project space, I am Editable, my advisor is Read Only, and lab mates are Commentable. Read-only allows viewing but not modifying source files; Commentable allows leaving opinions without polluting the main text. I've been trying Feishu these past few days to sync the Revision Comparison Table for my advisor to see, but I haven't uploaded unpublished experimental data, only structure tables and terminology lists. I've used Doubao Office for about a month, mainly for abstract drafts, but this time I only let it read Markdown text exported from WorkBuddy (a lightweight format using symbols to mark paragraphs and headings), not the full thesis text, to reduce accidental edits.

For daily maintenance, I set three rules for myself. Every day after editing, run the Citation Consistency template in WorkBuddy, exporting fields from Zotero and EndNote, letting WorkBuddy check if titles, years, and page numbers match.

I've used XiaoLvJing for about a month, Scholard for roughly the same duration, but term protection still relies on the Glossary node in WorkBuddy, locking words like Self-supervised, Robustness, and Confidence Level so it doesn't randomly replace them to lower AI detection. Every week, back up the Revision Comparison Table to Tencent Docs, which I've used for about a month; it's stable as a temporary relay.

There are two pitfalls to avoid. Don't throw the thesis body directly into any "AI detection reduction" feature, including the WorkBuddy pipeline I configured. It can find high-risk sentences, but when actually editing, I must go through it myself, especially for data descriptions and method boundaries. Previously, I thought AI research report tools were best as a data cleaning layer; now I judge paper tools the same way. Don't let it generate defense PPTs directly from the thesis. I tried a few days ago, and charts got lost. Later, I changed it to have WorkBuddy extract a Chart List first, manually map fields, then export the outline, so at least I know which chart goes on which page.

Tools can carry bricks for me, but they cannot sign for me. Is the thesis finished? No, but the food delivery arrived.