AI-Assisted Academic Paper Writing: From "Squeezing Toothpaste" to "Skeleton First, Fill Later"

Core problem: The biggest pain in AI paper writing is not inability to write, but writing the wrong things — baseline regressions without parallel-trend tests, introductions that never answer "why should readers care," and de-AI rewrites where version three looks more AI-generated than version one. The root cause: AI knows how to write a sentence, but not what a complete paper's skeleton looks like. Skills are that skeleton.

This post is a translated summary of a Chinese forum guide; all claims below belong to the original author.

Key points

1. The core contradiction: AI can form sentences but not structure

• Asking an AI to write a DID paper yields the baseline regression, then fragments only when prompted for parallel trends, placebo tests, heterogeneity, etc.
• The fix: give AI a methodology playbook — e.g., a DID skeleton of Introduction (Keith Head five elements) → Literature review → Data & variables → Identification strategy → Baseline results → Robustness (4–6 checks) → Heterogeneity & mechanisms → Conclusion.

2. Writing-stage skill suites

• academic-research-skills (1,790 stars): five-stage pipeline (research → write → review → revise → finalize) with journal-specific style calibration and citation hallucination checks against arXiv/CrossRef/Google Scholar.
• stats-paper-writing-agent-skills: includes check-tex (LaTeX/reference validation) and check-bib (BibTeX validation, DOI completion).
• claude-scientific-skills (8,799 stars, 140+ skills): 28+ database interfaces (OpenAlex, PubMed, Semantic Scholar), IMRaD generation, literature review, statistical analysis.
• academic-paper-skills: for humanities/social sciences, using a Strategist (argument framework) + Composer (content filling) dual-agent design.

3. Introduction writing: Keith Head's five elements + four agents

• Formula: Hook → Question → Antecedents → Value-added → Roadmap, with common failure modes (vague hooks, oversized questions, literature listing without a gap, overly modest contributions, roadmap that repeats the table of contents).
• introduction-writer uses Strategist → Drafter → Reviewer → Reviser. The Reviewer must run independently of the Drafter — the same agent writing and reviewing would inherit the author's bias.

4. Revision: three-agent editing + six-step anti-degradation protocol

• copy-edit-master: Structure Editor (Strunk & White + McCloskey), Line Editor, Quality Reviewer, with automatic Git commits per stage for rollback.
• ai-revision-guard six steps: anchor the draft as baseline → classify each edit (grammar/style/content/format) → execute in batches → report changes → cap at ≤2 rounds per section → run a 7-item homogenization checklist.
• Three degradation patterns detected: escalating word replacement ("important → critical → fundamental"), stance neutralization ("we find" → "studies show"), and unique argumentation flattened into generic templates.

5. De-AI polishing: a five-tool toolkit

• chinese-de-aigc: targets CNKI/Wanfang/VIP/Turnitin Chinese detectors; rewrites four-character clichés, function-word stacking, explicit connectives, absolute assertions, symmetric paragraph structures, and uniform sentence length; scores on five dimensions (specificity, rhythm, hedging, implicit cohesion, researcher voice).
• humanizer_academic: 23 AI-tell patterns in 5 categories for English papers.
• skill-deslop (5-dimension scorecard) and stop-slop (3-layer detection) distinguish legitimate disciplinary conventions from genuine AI tells.
• avoid-ai-writing: auditable four-stage output for compliance-sensitive contexts.

6. Citations and final formatting

• Zotero MCP: direct AI access to your library for PDF review, citation tracing, annotation.
• claude-code-my-workflow (Emory Econ 730, reportedly adopted by 15 research groups): 14 agents / 28 skills / 6 hooks, triple XeLaTeX compilation, adversarial review, and three quality thresholds (≥80 advisor-ready, ≥90 submittable, ≥95 top-journal).

Honest limitations (Section 7 of the original)

• Hallucination checks can miss misattributed authors/years/venues and unindexed niche literature — citation verification is necessary but not sufficient; humans remain the final gatekeeper.
• De-AI rewriting risks erasing legitimate academic conventions; distinguishing convention from tell requires domain knowledge.
• Four-agent architectures quadruple API calls, tokens, and latency — acceptable for a 500–800 word introduction, costly for full 8,000–15,000 word papers.
• The Emory workflow targets empirical economics; it does not fit theory papers and needs heavy adaptation for humanities or natural sciences.

Takeaway

Writing a paper is not producing sentences — it is persuading referees with logic. Plan the skeleton first, then fill in content, with a Git checkpoint at every step: structure planning (Strategist/Keith Head) → drafting → structural and line editing → de-AI polishing → degradation guard → citation/formatting. Skills do not replace your judgment; they structure it so the AI knows which direction to execute.

Reference links

• Main repository: https://github.com/brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research
• CoPaper.AI: https://copaper.ai
• academic-research-skills: https://github.com/Imbad0202/academic-research-skills
• claude-scientific-skills: https://github.com/K-Dense-AI/claude-scientific-skills
• academic-paper-skills: https://github.com/lishix520/academic-paper-skills
• chinese-de-aigc: https://github.com/brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research/tree/main/skills/48
• humanizer_academic: https://github.com/matsuikentaro1/humanizer_academic
• skill-deslop: https://github.com/stephenturner/skill-deslop
• stop-slop: https://github.com/hardikpandya/stop-slop
• avoid-ai-writing: https://github.com/conorbronsdon/avoid-ai-writing
• Emory Econ 730 workflow: https://github.com/pedrohcgs/claude-code-my-workflow
• Keith Head introduction formula: https://keithhead.wordpress.com/research/
• Zotero MCP: https://github.com/54yyyu/zotero-mcp