manuscript-writing

name	manuscript-writing
description	LLM guidance for staged academic manuscript writing in ResearchOS.

Manuscript Writing Guidance

Use this guidance during T8 writing and revision.

Core Principle

Write the paper as a staged research argument, not as one long generation. Tools organize resources, inventories, section plans, assembly, and mechanical audits. The Writer LLM is responsible for the scientific story, claims, section prose, and venue-aware positioning.

For UTD, Management Science, and Information Systems Research style work, make the main argument follow real-world information problem → computational problem → technical solution. Establish the phenomenon and its decision consequence first. Explain why the information structure or existing modeling choice naturally creates a computational difficulty. Only then introduce the technical design as a reasoned response. Carry design rationale, management relevance, and decision value through Method and Analysis instead of adding them as an isolated final paragraph.

Publication Prose and Citation Preferences

Treat these as editorial standards for writing and review. When they conflict with factual accuracy, verified evidence, or a target-venue template, preserve the fact and express it naturally.

• Aim for the clean, precise, and restrained prose expected in strong UTD/FT50/CCF-A work. Do not use empty importance claims, stock transitions, or repeated statements to make a section look longer.
• In ordinary body prose, do not rely on dashes, colons, or label sentences such as Problem:, Gap:, Insight:, or Implication:. Definitions, equations, tables, contribution lists, and a venue template may require limited punctuation of that kind, but the argument itself should use complete sentences that state causality, contrast, qualification, or consequence.
• Do not construct an argument as a sequence of parallel short sentences. Let each paragraph develop one claim through a readable order such as definition, mechanism, evidence, boundary, and implication. Move to a new term, paper, result, or practical consequence only after explaining why the preceding discussion leads there.
• Avoid semicolons and fragmented parallel clauses in ordinary prose. Prefer a complete causal, contrastive, or qualifying sentence. Define every term, construct, mechanism, and symbol at first use and keep the same wording and notation throughout the paper.
• Keep the section hierarchy compact. Do not create a subsection or \paragraph{} for every artifact, claim, or paper. Use a heading only when the argument changes function in a way that helps the reader navigate the paper.
• Explain a technical or theoretical term when it first matters to the argument. When useful, use a real or conditional such as scenario to make an abstract mechanism legible. An illustrative example must not be presented as data, a result, or a fact that has not been verified.
• Every citation must be real, available in the bibliography, and semantically matched to the precise claim it supports. Verify the paper's subject, method, setting, and evidence level in its note card, citation pool, or source record before citing it. Prefer genuinely relevant work from UTD/FT50/CCF-A venues, leading marketing, economics, management science, information systems, and computer-science outlets, together with foundational, highly cited, influential, and important recent work. Venue quality never permits a background citation to support a mechanism, causal effect, empirical result, business implication, or contribution that it does not actually establish.
• Treat citation selection as a source-reading decision, not a formatting operation. Do not infer support from a title, author reputation, venue, abstract snippet, or model memory. If the available source record cannot verify the intended wording, remove the citation and dependent strong claim, or use an explicitly bounded motivation statement. Broaden coverage through verified work that represents a distinct development stage, research tradition, comparison dimension, or evidence boundary, never through an unrelated citation bundle.
• Before writing Introduction or Related Work, inspect the section evidence supplement. If the existing library still misses a precise research stream, historical stage, or nearest-work comparison, run one bounded targeted_literature_supplement. Use its abstract notes for discovery and bounded context only, and upgrade the source before making a mechanism, result, causal, or novelty claim.
• Integrate verified literature, project evidence, and scholarly model reasoning into one natural argument. Model reasoning may improve explanation, connect evidence, derive design implications, motivate a hypothesis, and make the prose coherent without a visible provenance label. It cannot fabricate a citation or silently supply a bibliographic identity, empirical fact, numerical result, attributed prior-work finding, established causal claim, research consensus, resource availability claim, or external novelty verdict.
• Match contribution verbs to the accepted contribution character. Use introduce or a priority claim only when the novelty audit supports it, improve or extend for an improvement, and adapt or transfer for exaptation. Do not present a routine change as a core contribution.
• Keep tense stable. Describe the method and enduring definitions in the present tense, completed experimental procedures in the past tense, and general facts in the present tense. Prefer active voice when it makes responsibility clearer.
• Avoid stock AI prose such as delve into, it is worth noting, plays a crucial role, in the realm of, shed light on, pave the way, and generic rapid-development openings. The deterministic craft audit may flag these phrases, but the Writer should avoid them before audit rather than relying on repair.
• Keep repeated numerical results at the source precision across sections. A change in rounding, denominator, population, or evidence status requires an explicit reason and the corresponding source artifact.

Required Stages

• Resource index: inventory project, literature, hypotheses, novelty audit, experiment results, ablations, figures, tables, code, logs, and bibliography.
• Section plan: decide what each section must prove, which artifacts support it, and what remains missing.
• Outline: build the argument arc before drafting prose.
• Section drafts: draft abstract, introduction, related_work, methodology, experiments, analysis, and conclusion as separate files under drafts/sections/. Limitations are written inside Conclusion as \subsection{Limitations}.
• Assembly: combine section drafts into drafts/paper.tex mechanically, then polish transitions globally.
• Audit: check citations, numeric values, figure/table references, missing sections, and overclaims.
• Review and revise: treat reviewer reports as a change list; update prose, tables, figures, and audit notes.

Section Responsibilities

• Introduction: motivation funnel, precise gap, why existing approaches are insufficient, proposed insight, contribution bullets, headline evidence. Do not oversell.
• Related Work: taxonomy and contrastive positioning from synthesis and bibliography. Every citation must map to a real BibTeX key.
• Methodology: describe the proposed mechanism, algorithm/protocol, implementation choices, and how it differs from baselines.
• Experiments: datasets/settings, baselines, metrics, main results, ablations, seed ensemble, compute budget, and quality controls.
• Analysis: connect ablations and failures back to hypotheses and alternative explanations.
• Conclusion limitations subsection: external-executor evidence boundaries, mock/dry-run status, result-to-claim limitations, external validity, cost, and failure cases.
• Abstract and Conclusion: write after main sections; no new claims. Abstract should not contain formal citations.

Section Depth

Do not turn section-by-section drafting into several short placeholders. Except for the Abstract, each section should be fully developed according to its evidence:

• Methodology should cover artifact overview, component roles, inputs/outputs, algorithm or notation, design choices, and rejected alternatives.
• Experiments should cover RQs, data/splits, baselines, metrics, seeds/compute, main results, ablations, error analysis, and result provenance.
• Related Work should use the smallest coherent set of taxonomy or competing-rationale subsections, not a paper-by-paper laundry list.
• Analysis should explain support for design rationales, alternative explanations, failure cases, sensitivity, and boundaries.
• Introduction and Conclusion should form a complete problem-method-evidence-contribution chain.

Avoid hard word-count floors. If evidence is missing, use natural-language limitations, weaken/remove the unsupported claim, or record the issue in audit/self-check files. Do not leave literal TODO/TBD/LLM_REVIEW_REQUIRED/PLACEHOLDER tokens in final TeX. If evidence is available, develop the section to the density expected by the target venue.

Evidence Rules

• Numbers must come from drafts/experiment_evidence_pack.json, drafts/result_to_claim.json, experiments/results_summary.json, experiments/ablations.csv, or indexed run artifacts.
• Figures must come from existing generated assets or be planned in figure/table registries. If a figure cannot be generated, explain the limitation or remove the reference from final TeX.
• Claims about prior work must cite literature/related_work.bib keys and be traceable to literature/synthesis.md or paper notes.
• Treat synthesis.md [note:...] anchors as evidence provenance only. Convert them to real BibTeX keys from related_work.bib before writing TeX; do not paste [note:...] into paper.tex.
• Introduction and Related Work need visible citation-backed positioning, not only a bibliography at the end. Use representative citations in claim-bearing paragraphs.
• If evidence is external, weak, mock-only, or not fully audited, state the boundary in Conclusion's Limitations subsection and avoid claiming unsupported empirical validation.
• Abstract should not contain formal citations: no LaTeX citation commands, no author-year parenthetical citations, and no numeric citation brackets. Refer to method families or problem classes in the Abstract, then put concrete prior-work citations in Introduction or Related Work.

Tool Boundary

• Use build_manuscript_resource_index, plan_manuscript_sections, plan_manuscript_evidence, build_alignment_matrix, assemble_manuscript, audit_manuscript_claims, audit_writing_craft, and audit_paper_claims for mechanical workflow.
• Do not let tools invent claims, choose final framing, or write the final argument.
• If a tool audit flags a number/citation/figure, fix the source or mark the limitation; do not hide the issue.