# TS Write > Draft every section's LaTeX body from the evidence. Use after ts-plan, before ts-review. - Skill: `openjiuwen-ai/ts-write` (Agent Skill) - Install (CLI): `npx skillmds@latest add openjiuwen-ai/ts-write` - Raw SKILL.md: https://api.skillmd.com/api/skills/openjiuwen-ai/ts-write/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: openJiuwen-ai (https://skillmd.com/u/openjiuwen-ai) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/openjiuwen-ai/ts-write --- # ts-write Before drafting, read back `{PAPER_WORKSPACE}/title.txt`, `{PAPER_WORKSPACE}/contributions.txt`, and `{PAPER_WORKSPACE}/keywords.txt` from ts-plan — the Introduction and Abstract below build on them directly. Write one file per section under `{PAPER_WORKSPACE}/sections/` (the full absolute path — never a bare relative `sections/...`), in this fixed order (id — title — evidence to use — target word count — citations allowed). These targets are deliberately close to a real paper's actual length (adapted from spark-to-paper-skills' own recipe) — hit them with genuine detail pulled from the evidence, not filler. If you run out of real content before reaching the minimum, that's a sign the evidence doesn't support a section this long — say less rather than pad with restatement, hedge-phrases, or generic transitions, and note in your final summary that this section's evidence was thin. 1. `introduction.tex` — Introduction — background evidence — 800-1200 words — yes. End with a lead-in sentence, then `\begin{itemize}` with **exactly 3 `\item`s** — one per line of `contributions.txt`, verbatim or lightly rephrased to fit as a complete sentence — then `\end{itemize}`. No heading before the list. 2. `related_work.tex` — Related Work — background evidence — 400-800 words — yes. Organize by theme, never chronology: **exactly 3 `\subsection{...}`** (each a distinct thread of prior work relevant to this run), each ending with an explicit one-sentence gap statement ("However, none of these address X, which this run's approach does by Y."). Follow the 3 subsections with a plain closing paragraph, no heading, contrasting the single nearest prior approach. 3. `method.tex` — Method — objective/hypothesis + method evidence — 2000-3000 words — yes. Structure: `\subsection{Problem Formulation}` (one framing paragraph) → `\subsection{Notation}` (only if you use mathematical notation — a small table via `\begin{tabular}` mapping symbol to meaning; if the evidence has no math, still include this subsection with a one-sentence note that the method is described procedurally, not symbolically) → one `\subsection{}` per method component named in the evidence → a closing subsection stating the decision rule / objective the method optimizes for, if the evidence names one. If the evidence describes a multi-step procedure, include at least one pseudocode block (a `\begin{enumerate}` numbered list of steps is fine — do not use `algorithm`/`algorithmic` environments unless the evidence itself is already pseudocode-shaped). Include the host-rendered evidence exactly as given if a figure/table is provided in this block, in the subsection it belongs to. 4. `experiments.tex` — Experiments — results evidence — 1000-1500 words — yes. **Exactly 3 `\subsection{}`s, in this order**: `Implementation Details` (setup/protocol facts from the evidence, one paragraph), `Experimental Design` (variants/metrics/baselines and why, one paragraph), `Results` (the host-rendered results table and figure from the evidence, included exactly as given — never redraw them — plus prose grouped by outcome pattern, mentioning every variant and every decision metric named in the evidence, by name). 5. `discussion.tex` — Discussion — discussion/reflection evidence — 900-1400 words — yes. At least one `\subsection{}` (e.g. "Limitations and Failure Modes"): ground it in a real caveat from the evidence — a metric that didn't move, a variant that underperformed, a boundary condition the reflection or plan noted. If the evidence reports no failures at all, discuss the conditions under which the result might not generalize instead of inventing a failure. 6. `conclusion.tex` — Conclusion — discussion evidence — 200-280 words — no citations 7. `abstract.tex` — Abstract — write this one LAST, from `title.txt` and the other six section files you just wrote (read them back), not from a primary evidence block — 150-220 words — no citations. **Prose only — no `\section{...}` heading, no heading of any kind.** The compile step wraps this file's raw content in a proper `\begin{abstract}...\end{abstract}` block itself; a heading you add would show up as extra, wrongly-numbered text inside that block. ## Hard rules for every section - Output LaTeX body content only: a `\section{...}` heading followed by prose (**except `abstract.tex` — see item 7 above, prose only, no heading**). The heading text doesn't need to match anything exactly — the compile step replaces it with the canonical title regardless — but include one so the section reads naturally while you're writing it. No `\documentclass`, `\usepackage`, `\begin{document}`, `\textbf`, or `\textit` — none of these are allowed in prose (`\textbf`/`\textit` in particular: this house style uses no bold/italic emphasis in running text, only in captions). Never put a number at the start of a `\section{}`/`\subsection{}` heading (LaTeX numbers headings itself) and never use a markdown code fence (`` ``` ``) anywhere. - Ground every claim in the evidence you were given. No fabricated numbers, sources, or claims. This run's results are always real and already measured — never hedge with "we expect"/"this should show" for anything already reported as measured. A hyperparameter or design constant quoted from the evidence (e.g. "a learning rate of 0.001") is not a fabricated result and is fine even if it isn't one of the run's reported metrics. - Cite only with `\cite{key}` for keys listed in the evidence's citation list. Never invent a key. - One logical paragraph per physical line — no hand-wrapping mid-sentence. - Escape LaTeX special characters (`_ % & #`) in prose you write yourself. - Never quote non-Latin script verbatim (CJK, Cyrillic, Arabic, etc.) even if it appears in the evidence — the default fonts can't render it. Paraphrase or describe it in English instead. - Avoid AI-writing tells: "it is worth noting," "plays a crucial/pivotal role," "a testament to," "rich tapestry," "delve into," "the realm of," "ever-evolving," "in today's world," "navigating the landscape," "paradigm shift," "game-changing," "in order to." Punctuate an appositive or enumeration with a colon or em-dash, never a bare comma — "a pipeline: retrieval, planning, execution," not "a pipeline, retrieval, planning, execution, which..." (that bare-comma "comma soup" reads as fragmented, not connected prose). - The Experiments section must mention every variant and every decision metric named in the evidence, by name.