Research Paper

Research Paper

A production-grade agent skill that turns any compatible coding agent (Claude Code, OpenCode, Cursor, Cline, Codex, Aider, Amp, Antigravity, and 50+ others) into a multi-agent research system:

Orchestrator → Researcher → Methodologist → Analyst → Visualizer → Writer → Citation engine → Validator → Reviewer → Publisher.

It produces full, citation-heavy, visually rich, publication-ready outputs in arXiv / IEEE / ACM / Nature / Harvard styles, plus literature reviews, theses, technical whitepapers, survey papers, and policy briefs.

This file is the entry point. It is intentionally compact. Heavier guidance (instructions, workflows, engines, validators, rubrics) lives in the topic folders below and is loaded on demand via Claude Code's filesystem tools (progressive disclosure).

1. When to activate

Slash commands (preferred)

Common options (any command): --style [harvard|apa|ieee|mla|chicago|nature|arxiv-numeric], --format [arxiv|ieee|acm|nature|harvard|...], --depth [quick|standard|comprehensive], --sources [N], --visualizations [auto|N|none], --audience [academic|technical|executive|general].

Natural-language triggers

• "Write a research paper / academic paper / scientific paper on …"
• "Do a literature review / systematic review on …"
• "Format this draft as IEEE / ACM / arXiv / Nature / Harvard / APA …"
• "Write a thesis chapter / dissertation chapter on …"
• "Produce a whitepaper / survey paper / policy brief on …"
• "Analyze this dataset and write up the findings as a paper."
• "Add citations / bibliography / references in <style>."
• "Peer-review this draft / validate the methodology."

Do NOT activate for

Blog posts, marketing copy, tweets, casual answers, or single-paragraph explanations. Those are handled normally without this skill.

2. Operating principles (read every time)

0. Anchor to TODAY's date FIRST. Before any planning, search, or writing, determine today's actual date — via system clock (date -u +%Y-%m-%d), runtime context, or asking the user. Never default silently to the training-data cutoff. Year ranges (--years last-3) are computed from today, not from the model's training year. Full protocol: instructions/freshness.md.
1. Plan before writing. Always start with the research plan in orchestration/pipeline.md. Never jump into prose.
2. Progressive disclosure. Only read the file you need for the current step. Never preload the whole skill.
3. Evidence first. Every non-trivial claim is backed by a citation, dataset, equation, or explicit derivation.
4. No hallucinated citations. Never invent DOIs, page numbers, authors, or volumes. Mark gaps with [CITATION NEEDED] or [UNVERIFIED] and surface them in Known gaps.
5. Reproducibility. Datasets, code, environments, seeds, and hyperparameters are documented end-to-end.
6. Dual register. Maintain academic rigor and a "Plain-English summary" for non-specialists (see prompts/simplification-prompts.md).
7. Visual-by-default. Comparisons, trends, distributions, structure, geography, and processes always get a figure or table (see visualization_engine/decision-engine.md).
8. Self-review. Run the simulated peer-review pass (review_pipeline/) and the publication checklist (quality_control/publication-checklist.md) before delivery.
9. No silent failures. Anything missing surfaces in a Known gaps block at the end of the paper.
10. Multi-agent ready. For long papers, dispatch sub-agents per orchestration/agents.md.

3. Top-level workflow

intake → plan → lit-review → methodology → data-analysis →
visualization → drafting → citations → validation → review → ship

Each step has a dedicated playbook. Read it, do the step, persist the artifact to disk, move on. Detailed master pipeline: orchestration/pipeline.md.

4. Format selection

When the user does not specify a format, infer it:

If still ambiguous, ask once, then proceed.

5. Citation style selection

Map domain → default style if not specified:

• CS / engineering / physics → IEEE numeric
• ML / AI / preprint → author–year (Harvard / APA-compatible)
• Biology / medicine / Nature → Nature numeric superscript
• Social science / business / humanities → Harvard (or APA)
• Law / history → Chicago

Style rules: citation_engine/citation-styles.md. Per-style modules: citation_engine/styles/. The deterministic formatter is toolchains/format_bibliography.py.

6. Visualization decision (summary)

Full rules: visualization_engine/decision-engine.md and visualization_engine/visualization-guide.md. Rendering happens via toolchains/generate_charts.py; if Python is unavailable, the skill falls back to Markdown tables + Mermaid diagrams — never silently skips a planned figure.

7. Tooling expectations

This skill works in three tiers, gracefully degrading:

If a tier is missing, the skill detects it and adapts — no silent failures. See toolchains/README.md for setup.

Output formats

The skill produces Markdown by default. For other formats, run the output converter:

python toolchains/convert_output.py --input paper-final.md --to pdf --out paper.pdf
python toolchains/convert_output.py --input paper-final.md --to docx
python toolchains/convert_output.py --input paper-final.md --to html
python toolchains/convert_output.py --input paper-final.md --to tex
python toolchains/convert_output.py --input paper-final.md --to epub

Supported targets (via Pandoc): md (always), html, docx, pdf (needs LaTeX), tex, rtf, epub, odt, pptx.

Self-test:

python toolchains/convert_output.py --self-test

The user can also request a non-Markdown output directly:

/research "topic" --output paper.pdf
/research "topic" --output paper.docx

8. Output contract

Every artifact this skill produces includes, at minimum:

1. Title — specific, ≤ 15 words.
2. Authors / Affiliation block — placeholders if not provided.
3. Abstract — 150–300 words, structured.
4. Keywords — 4–8.
5. Plain-English summary — 5–10 sentences.
6. Numbered sections following the chosen template.
7. At least one figure and one table for any paper > 1500 words (unless purely theoretical and explicitly opted out).
8. In-text citations in the chosen style.
9. Full reference list with DOIs / URLs.
10. Limitations section.
11. Future work section.
12. Reproducibility statement (data, code, environment, seeds).
13. Appendices for derivations, hyperparameters, prompts, raw outputs.

Anything missing is surfaced in a final Known gaps block — never silently swallowed.

9. Long-context strategy

For papers > ~10,000 words:

• Persist every artifact to disk before moving on (paper-spec.md → outline.md → bibliography.yaml → methodology.md → analysis/findings.md → figures-plan.md → sections/<NN>-<name>.md → paper-draft.md → paper-cited.md → paper-final.md).
• Read only the section being drafted (plus the outline) at any time.
• Cross-section consistency is enforced by the outline + a final cover-to-cover read pass.

Full strategy: long_context/strategy.md.

10. Multi-agent orchestration

For deep / parallel runs, dispatch sub-agents:

Topology: orchestration/agents.md.

11. Failure handling

• Missing data → synthetic illustrative dataset, clearly labeled.
• Unverifiable source → [UNVERIFIED], listed in Known gaps.
• Conflicting evidence → explicit "Contradictions in the literature" subsection.
• Out-of-scope request → narrow scope, list dropped sub-topics in Future work.
• Token / context pressure → see §9.

Full failure-handling matrix: orchestration/failure-handling.md.

12. Where to look next

• Plan a paper → orchestration/pipeline.md
• Pick a template → templates/
• Write a section → prompts/writing-prompts.md
• Add citations → citation_engine/, workflows/citation-pipeline.md
• Make charts → visualization_engine/, workflows/visual-generation-pipeline.md
• Validate stats → methodology_engine/statistical-methods.md, toolchains/statistical_validation.py
• Self-review → review_pipeline/, rubrics/academic-quality.md
• Ship it → quality_control/publication-checklist.md

Always prefer reading the specific file you need over re-reading this one.