ai4s-research/ai4s-skills/paper-writer

Paper Writer

Overview

End-to-end research paper builder. Single stage, full quality from the start — there is no skeleton phase to enrich later. The agent (Claude Code / Cursor / Aider / Codex / …) does the writing using its own tools (WebFetch, WebSearch, Write, Bash). This skill has no Python runtime; it is purely a procedure + reference playbooks + a LaTeX template.

The substantive work is decomposed into reference playbooks under references/:

Read the relevant reference _before_ writing, not after.

The full pass does not fit in a single turn. The bibliography is built across ~20+ small WebFetch/WebSearch batches; sections are drafted one per turn; figures are generated one at a time. Read `references/00-incremental-execution.md` before starting — it is the only execution mode that actually completes without losing work.

When to Use

• User asks to "write a paper" on a specific topic.
• User wants Abstract + Introduction + Related Work + Method + Experiment + Results + Conclusion.
• User has experiment results (a results.json) and wants them formatted into a paper.

When NOT to Use

• User wants only a literature survey → the literature-survey skill.
• User wants only the experiment package → the experiment-suite skill.
• User wants only direction/topic exploration → the research-explorer skill.
• User wants the full multi-skill pipeline → the ai4s-agent skill (which invokes this skill as one stage).

Workflow

Step 1 — Understand requirements

Confirm with the user:

• Topic — specific enough to motivate a title; if too broad, narrow it before proceeding.
• Experiment provenance — measured (user supplied a results.json produced by the experiment-suite skill or compatible) or simulated. Default is simulated; in that case the disclosure footnote must flag it (see references/06-experiment-provenance.md).
• Language — default Chinese in conversation; the paper itself is English unless the user requests otherwise.

Always tell the user that human review by a domain expert is recommended before any scientific publication or production use.

Step 2 — Set up the run directory

Create a timestamped working directory and copy the template. Runs never overwrite each other.

TOPIC="<topic>"
SLUG=$(python3 -c "import re,hashlib,sys; t=sys.argv[1]; n=re.sub(r'[\\s_]+','-',re.sub(r'[^\\w\\s-]','',t.lower().strip())).strip('-')[:40].rstrip('-'); h=hashlib.sha1(t.encode()).hexdigest()[:8]; print(f'{n}-{h}')" "$TOPIC")
TS=$(date +%Y-%m-%d_%H%M%S)
RUN=output/paper-writer/$SLUG/$TS/paper
 
mkdir -p "$RUN/sections" "$RUN/figures"
cp -r templates/paper/. "$RUN/"
ln -sfn "$TS" "output/paper-writer/$SLUG/latest"

In commands below $RUN = output/paper-writer/<slug>/latest/paper.

The template provides only main.tex (title placeholder), an empty sections/ skeleton, an empty figures/, and compile.sh. Everything substantive is produced in Step 3 below.

Step 3 — Build the paper (REQUIRED — this is the whole job)

Open references/00-incremental-execution.md first. Then carry out the five tracks below across many turns, persisting state to $RUN/ after every batch.

3.1 Bibliography — 200+ real entries

Open: references/01-bibliography-expansion.md.

First choose and record the temporal profile from that reference. AI4S and similarly fast-moving fields default to at least 60% of references from the current calendar year and previous two years; an explicitly recent window uses the stricter recency-led profile. Then plan 15–25 query angles. For each angle: WebSearch → pick candidates → WebFetch each candidate's abstract / arXiv API URL → extract canonical title/authors/year/venue/url → append a BibTeX entry to $RUN/bibliography.bib. Every entry must originate from a URL fetched in this session. Memory entries are forbidden.

Hard stop: do not draft prose until the bibliography has ≥ 200 entries, contains no unknown keys, and passes check_bibliography_freshness.py for the recorded profile.

3.2 Figures — 4–8 publication-grade

Open: references/02-figures-publication-grade.md.

Decide what the paper needs based on its claims and evidence:

• Architecture / pipeline diagram only when the paper introduces or compares a

real method whose mechanism needs explanation.

• Quantitative comparison plots when measured or explicitly simulated results

support them.

• Heatmap / multi-panel ablation only when the data justifies it.

Generate each figure into $RUN/figures/. Save the matplotlib / TikZ source alongside the PDF so each figure is reproducible. If the experiment-suite produced a figures/manifest.json, reuse those figures by symlink or copy — don't redraw what's already produced.

3.3 Sections — 7 substantive .tex files

Open: references/03-section-playbook.md.

Draft each section per its playbook (length, structure, citation density, equation requirements, anti-patterns). Cite real entries from the bib built in 3.1.

Order: introduction → related_work → method → experiment → results → conclusion → abstract last (you only know the paper's shape after writing the rest).

3.4 Layout discipline

Open: references/04-layout-discipline.md.

• Put each figure or table in the section whose prose first introduces or

interprets it, immediately after that paragraph in the source. Do not collect artifacts in a fixed section or force one float placement across the paper.

• Wrap tables and figures in standard LaTeX floats with standalone captions;

choose [htbp], [tbp], or [p] from the artifact's size and narrative role.

• Use ~\cite{} and ~\ref{} (non-breaking space).
• Let LaTeX assign citation, figure, table, equation, section, and algorithm

numbers from 1 in first-appearance order. Never type display numbers manually.

• Set \author{AI4S Agent} and attach a \thanks footnote that always recommends human review, and additionally flags simulated numerics when applicable.

3.5 Compile + quality gate

cd "$RUN"
pdflatex -interaction=nonstopmode main.tex
bibtex main
pdflatex -interaction=nonstopmode main.tex
pdflatex -interaction=nonstopmode main.tex

Open: references/05-quality-gate.md.

Run all G1–G8 hard gates and S1–S4 soft gates. If a hard gate fails, fix and re-run; do not ship a paper that fails G1–G4. If you cannot honestly clear a gate (e.g., bibliography stalled at 156 entries because the topic is niche), say so explicitly instead of padding.

Step 4 — Deliver

Report to the user:

1. output/paper-writer/<slug>/latest/paper/main.pdf — final PDF.
2. output/paper-writer/<slug>/latest/paper/ — complete LaTeX project (reproducible).
3. Stats per the report format in references/05-quality-gate.md (pages, bib size, total \cite{}, figure count, table count, provenance, compile warnings).

Cross-skill data flow (path convention)

If a sibling skill has already run for the same topic, reuse its outputs by path:

• output/literature-survey/<slug>/latest/bibliography.bib → seed $RUN/bibliography.bib (still bring it up to 200+ in 3.1 with WebFetch).
• output/experiment-suite/<slug>/latest/results.json → the source of the numbers cited in 3.3 / 3.5; its simulated flag controls the disclosure clause in 3.4.
• output/experiment-suite/<slug>/latest/figures/*.pdf (+ manifest.json) → reuse in 3.2 rather than redrawing.

The slug formula in Step 2 is the contract; all four skills compute the same slug for the same topic.

Important rules

• No LLM SDK in this skill. No import anthropic / import openai. The agent runs the procedure; the skill is just SKILL.md + references + template.
• No fabricated citations. Every BibTeX entry must trace back to a URL fetched this session. Real or weaker claim — never fake reference.
• Simulated numbers stay visibly labelled. Title \thanks + abstract disclosure paragraph + per-caption disclosure. Do not let the simulated label disappear during drafting.
• Honest stop > padding. If the topic is too niche for 200+ real citations, say so to the user instead of inventing entries.
• Real-paper scope is 8–14 pages with 200+ references. For workshop / blog format, adjust scope explicitly with the user up front.