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.