Step-by-Step Guide to Writing Manuscript Using SciSpace Agent

Manuscript writing with SciSpace Agent becomes practical when researchers treat it like a workflow engine: curate evidence and research notes first, then feed the agent tightly scoped inputs so it can generate journal-specific sections without inventing content. The core payoff is speed—turning structured notes, references, and trial data into a complete, journal-aligned manuscript draft (including a compilable LaTeX/PDF output)—but only after careful “context engineering” to keep outputs precise.

The process starts after research is essentially done: literature review, research question, hypotheses, and an RCT-style set of research notes with participants, interventions, outcomes, and results. Using a demo topic—“effectiveness of generative AI enhanced feedback on undergraduate academic writing”—the workflow assumes the researcher already has curated notes and a reference set. From there, SciSpace Agent supports the literature review stage by running a query for papers on generative AI in education from 2019 to 2025, selecting relevant databases (including SciSpace Search, Google Scholar, and arXiv), and returning a ranked list of highly relevant papers (97 top results). The agent also surfaces research gaps, and the researcher can inspect abstracts, methods, and conclusions.

Next comes organization and structured extraction. Papers are bookmarked into a SciSpace Library (positioned as a more capable alternative to Zotero), then filtered using custom columns such as outcome type (e.g., writing quality, engagement, academic integrity), intervention type, population characteristics (e.g., undergraduate focus and U.S.-majority studies), and AI tool usage (often GPT variants). Once the shortlist is finalized, the researcher exports the library metadata to CSV. SciSpace Agent can then ingest that CSV and cluster papers into user-defined themes, returning an updated CSV with a theme column—using a generated Python script to perform the clustering.

With references and RCT notes ready, the workflow shifts to drafting. The agent generates a journal-specific outline (introduction, methods, results, discussion, plus supporting materials) with word limits, subsections, and a checklist. A key warning is that dumping excessive or irrelevant context can degrade output quality; better results come from attaching only the necessary CSV columns and concise research notes. The agent then drafts each section in sequence. For the introduction, a notable failure is caught: the first draft does not follow APA 7th edition citation formatting. A counter-prompt for self-checking forces verification against the outline’s citation and formatting requirements, leading to corrected citations, completed references, and adherence to subheadings and word counts.

Methods drafting follows the same pattern: start prompts with the task, attach only the relevant context (research notes and the outline), and explicitly instruct the agent not to invent sample sizes or procedures not present in the notes. Results drafting is more demanding because it may require analysis and figures. Here the agent is instructed to plan the work, use attached datasets (student-level data and AI usage logs), generate tables/figures/diagrams, and avoid fabricating facts. The agent writes Python code for analysis and chart creation, then produces a results section in markdown along with intermediate and final output files.

Finally, the agent generates discussion, abstract, and conclusion using the already-produced introduction/methods/results content as context. The last step merges all sections into a LaTeX “.tex” structure, compiles a PDF (44 pages in the demo), and provides editable markdown/LaTeX files for refinement in tools like Overleaf. The practical lesson is that AI accelerates drafting most reliably when researchers supply curated references, structured notes, and explicit constraints that match journal requirements.