SkillAgentSearch skills...

ts-paper-write

Stage 3 of the ts-paper suite. Draft every section of a Traitement du Signal paper as LaTeX body files (sections/<id>.tex + abstract.tex) from the blueprint and refs.bib, following strict per-section recipes, the no-fabricated-results rule, and IIETA citation style. Write ALL sections in one pass

Install / Use

npx skills add Spark-To-Paper-Skills/spark-to-paper-skills --skill ts-paper-write

Installs into whichever agent you are using.

About this skill
๐Ÿ“„

SKILL.md

Installable skill definition

Quality Score

91/100

Supported Platforms

Universal

Tags

Our assessment of ts-paper-write

ts-paper-write scores 91/100 on our quality scale, 368th of 1,155 Content & Media skills we index (top 32%).

Its SKILL.md is 18 KB long, well organised into 8 sections with 2 code examples: a thorough specification that gives an agent plenty to work with.

With 1,126 GitHub stars, it is one of the more widely adopted skills in the catalogue.

Substance
30/30
Structure
18/20
Description
15/15
Adoption
13/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated about 3 months ago, so ts-paper-write is actively maintained.
  • Our last check on 2026-10-29 found the source still online.
  • It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
  • Its trust signals score 100/100, with no cautions. These come from repository metadata, not a code audit โ€” read the skill file before letting an agent act on it.

Safety scan

No issues found

Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful.

AI review by kimi-k2.7-code on 2026-10-04. Automated pattern scan on 2026-10-04. It catches known dangerous patterns, not every risk โ€” read a skill before letting an agent act on it.

ts-paper-write compared with similar skills

All 4 of these similar skills score higher than ts-paper-write; compare them before choosing.

SkillScoreStarsUpdatedFormat
ts-paper-write (this skill)by Spark-To-Paper-Skills911.1k3mo agoSKILL.md
siyuanby siyuan-note10047.3ktodayMCP Server
x402by browser-use100121.0k3d agoSKILL.md
algorithmic-artby anthropics100177.9k55d agoSKILL.md
pptxby anthropics100177.9k55d agoSKILL.md

Frequently asked questions

How do I install ts-paper-write?
Run npx skills add Spark-To-Paper-Skills/spark-to-paper-skills --skill ts-paper-write. The install tabs above show the steps for each supported agent.
Which AI agents does ts-paper-write work with?
It is written for Universal, as a SKILL.md file. Other agents that read the same format can often use it too.
Is ts-paper-write safe to use?
Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful. It is MIT-licensed and scores 100/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is ts-paper-write still maintained?
The repository was last updated about 3 months ago, so ts-paper-write is actively maintained.

name: ts-paper-write description: > Stage 3 of the ts-paper suite. Draft every section of a Traitement du Signal paper as LaTeX body files (sections/<id>.tex + abstract.tex) from the blueprint and refs.bib, following strict per-section recipes, the no-fabricated-results rule, and IIETA citation style. Write ALL sections in one pass. Use when writing the body of a TS proposal paper.

ts-paper-write โ€” draft all sections (LaTeX body)

Quality is the goal. Write all sections in one holistic pass when you can (you hold the whole paper in context, so terminology stays consistent and nothing is re-explained) โ€” but a section that isn't genuinely good gets rewritten until it is. Output one LaTeX body file per section into sections/<id>.tex, plus sections/abstract.tex. Aim for prose a TS reviewer would accept: precise, specific, well-argued โ€” not filler that merely hits the word count.

The one rule that overrides everything: no fabricated results

This is a proposal โ€” there are no real experiments. Forbidden anywhere in a sentence: any concrete metric/percentage/score/delta (18.3%, 0.72 F1, +0.05, "beats SOTA by X points") and stand-ins ([X]%, XX.X, [TBD]). Write forward-looking: "We evaluate X on three benchmarks and report results in Table~\ref{tab:main_results}; we expect Y because Z." The only place -- may appear is inside a results table cell, never in prose.

What the number-audit catches in code vs self-review: in proposal mode draft_lint hard-fails these prose number forms โ€” percentages (18.3%), decimal scores (0.72, and any decimal โ‰ฅ 1 such as 1.8), signed deltas (+0.05), x-multipliers (2.5x), word-form magnitudes (doubles, three times, ninety percent), and placeholder tokens ([X]%, XX.X, [TBD]). Bare integers are treated as design constants/counts and are not audited (so "256 dimensions" / "Section 3.2" / a year never false-flags) โ€” so any fabricated bare-integer result is a judgment item you must self-review, the code will not catch it.

Output format (LaTeX body files)

  • Each sections/<id>.tex contains body content only โ€” NO \documentclass/\usepackage/\begin{document}/\title, and NO top-level \section{} (the assemble stage adds the canonical ALL-CAPS heading). Use \subsection{...} / \subsubsection{...} for structure; never put a number in a heading (LaTeX auto-numbers); never \label a heading.
  • One physical line per paragraph. Write each paragraph (and each \item/\caption) as a SINGLE line โ€” do NOT hard-wrap prose mid-sentence at ~80/100 columns. Separate paragraphs with one blank line. A single newline is just a space in LaTeX, so wrapping is invisible in the PDF, but a hard-wrapped .tex reads as choppy "broken" writing to a human browsing the source. (The assemble stage also runs reflow_tex.py as a net, but write it clean from the start.)
  • No bold in running prose (TS house style). Emphasis only in captions. No \textbf/\begin{itemize} mid-paragraph except the intro contribution list.
  • Math: inline $...$; display as numbered \begin{equation}...\end{equation} (no \[...\], no starred envs). Always use LaTeX commands for symbols ($\Delta$, $\alpha$, $\leq$) โ€” never raw Unicode (the #1 compile-killer). Split display lines >~60 chars with align/split (two columns are narrow).
  • Citations: emit \cite{bibkey} immediately after the supported claim's sentence; group adjacent ones as \cite{k1,k2}. Only cite real bibkeys present in refs.bib. Claims about your own method get no citation. No citations in abstract or headings. Cross-domain ban: cite only same-task / nearest-technical lineage.
  • Write substantive, flowing full paragraphs โ€” connected prose, NOT a comma-separated list of clauses. Punctuate an appositive/enumeration with a colon or em-dash, never a bare comma: write "a pipeline: retrieval, planning, drafting" or "a pipeline โ€” retrieval, planning, drafting โ€” in which โ€ฆ", NOT "a pipeline, retrieval, planning, drafting, in which โ€ฆ" (that bare-comma comma soup is exactly what reads as fragmented "็ขŽๅฅๅญ"). Vary sentence length, use real transitions, keep terminology consistent (use the blueprint's terminology names exactly). Write human from the start โ€” avoid AI tells. draft_lint hard-fails (ai_tell) only the CONTEXT-FREE canned phrases in code: "it is worth noting", "plays a crucial role", "a testament to", "tapestry", "delve into", "realm of", "paradigm shift", "navigating the landscape", "ever-evolving", โ€ฆ plus in order to. The CONTEXT-DEPENDENT cohesion tells โ€” em-dashes, sentence-initial firstly/moreover/furthermore/additionally,, not only โ€ฆ but also, rule-of-three triplets, leverage/utilize โ€” are judgment-only (deliberately NOT code-caught: a regex can't tell a legitimate appositive em-dash from abuse, and banning the token outright produced the comma soup). You and ts-paper-refine's de-AI pass keep the good uses and remove the abuse by taste. Banned filler: "In this section we", "As mentioned above", "Recall that".

Per-section recipes

abstract (sections/abstract.tex) โ€” one unbroken paragraph, 150โ€“220 words (ts_iieta band), no heading, no citations, no math, all acronyms spelled out. Four beats: (1) problem + why it matters, (2) precise technical gap, (3) method + the specific mechanism/insight, (4) evaluation scope + hedged outcome. Qualitative outcome claims only ("consistently improves") โ€” never a number.

introduction โ€” exactly 5 paragraphs: (1) background + real-world stakes, โ‰ฅ2 cites; (2) gap/attack on prior methods, โ‰ฅ4 cites each followed by a concrete limitation ("X cannot handleโ€ฆ"); (3) our method + key insight, 0 cites/0 numbers; (4) validation scope, hedged, no numbers, may reference Table~1/Section~4; (5) a lead-in sentence then a 3-item contribution list (\begin{itemize} with exactly 3 \item, no heading), each naming a specific component + 1โ€“2 competing methods (cited) + the concrete difference. โ‰ฅ8 cites total; ~800โ€“1200 words (ts_iieta band) โ€” substantive paragraphs, not terse.

related_work โ€” by theme, never chronology. Use the template's theme_subsections count of theme \subsection{}s (ts_iieta = EXACTLY 3: Theme A = task/benchmark lineage, Theme B + Theme C = two technical-mechanism / adjacent lineages) โ€” followed by a plain closing paragraph with NO \subsection heading (draft_lint enforces exactly that many \subsections โ€” no more, no fewer; the closing paragraph itself is a recipe requirement you satisfy, not a code check). Each theme: per representative paper, one technique sentence + one contrast/limitation tied to your setting; end each theme with an explicit technical gap ("However, these do not addressโ€ฆ, which we resolve byโ€ฆ"). 5โ€“9 cites/theme (โ‰ˆ15โ€“27 total) โ€” this section carries the bulk of the well-read 40โ€“50 reference set; the closing paragraph contrasts the single nearest paper and bridges to Method.

method โ€” \subsection{Problem Formulation} โ†’ \subsection{Notation} (render a small notation table as an inline minipage tabularx, see table pattern below; one intro sentence) โ†’ one \subsection{} per component โ†’ \subsection{Training Objective}. Every subsection opens with prose before any equation. Define every symbol before use; fully specify each learned module (write the actual MLP, not "Enc(C)"); name the differentiable approximation if discrete selection is used. For each component give a Design Rationale (the obvious alternative, why rejected, what your choice enables). Include โ‰ฅ1 pseudocode block for the main procedure โ€” write it as an algorithm+algpseudocode (algorithmicx) environment placed right after the paragraph that first references Algorithm~\ref{...}; use the mixed-case macros \State/\For/\While/\If/\Require/\Ensure (NOT uppercase \STATE โ€” the .sty loads algpseudocode, not algorithmic); keep identifiers short. This is the longest section โ€” target ~2000โ€“3000 words (ts_iieta band): give each component its own \subsection{} with full formalization, the design rationale, and complexity/where-it-helps discussion. Add depth (more real detail), never filler, to reach the band.

experiments โ€” exactly 3 \subsection{}: Implementation Details (1 paragraph: hardware, framework, optimizer, LR, batch, epochs, preprocessing; no equations), Experimental Design (1 paragraph: datasets name/size/domain/why, metrics + what higher means, baselines + why; no equations), Results (place the 3 result tables in order โ€” main_results, secondary_results, ablation_results โ€” each immediately before its discussion; results prose ~600โ€“900 words grouped by outcome pattern, all forward-looking, no numbers). The section targets ~1000โ€“1500 words total (ts_iieta band). Cells are --.

analysis โ€” separate from experiments, no math. ~900โ€“1400 words (ts_iieta band): 2โ€“3 analysis \subsection{}s (Ablation Study, Component Analysis, Sensitivity/When-it-helps) plus a required Failure Mode Analysis (scenario, why it struggles, mitigation). Every comparative observation must reference a specific Table/Figure ("As shown in Table~\ref{tab:ablation_results}, โ€ฆ"). Directional predictions allowed if framed as predictions.

conclusion โ€” exactly one paragraph, no lists/numbers/math, ~200โ€“280 words: summary, key takeaway, one limitation, forward direction. Hedged only ("demonstrates potential", "suggests"); banned: "proves", "guarantees", "revolutionary". Don't copy the intro's contribution wording.

Table & figure LaTeX patterns

Results table (โ‰ค5 cols โ†’ table; >5 cols โ†’ table*), caption goes ABOVE (assemble enforces this), cells --:

\begin{table}[htb]\centering\normalsize\setlength{\tabcolsep}{3pt}
\caption{Main results on <task> across the reported metrics.}\label{tab:main_results}
\adjustbox{max width=\columnwidth}{%
\begin{tabular}{lccc}\toprule
Method & M1 & M2 & M3 \\ \midrule
Baseline A & -- & -- & -- \\
<Method> (Ours) & -- & -- & -- \\
\bottomrule\end{tabular}}
\end{table}

Notation table (inline, in method): \par\noindent\begin{minipage}{\columnwidth}\centering ... \begin{tabularx}{\columnwidth}{@{}l X@{}}\toprule Symbol & Meaning \\ \midrule $x$ & ... \\ \bottomrule\end{tabularx}\captionsetup{type=table}\captionof{table}{Notation.}\label{tab:notation}\end{minipage}\par Figure (placeholder, caption BELOW). Emit a machine-readable FIGURE-SPEC comment so the later ts-paper-figure stage knows what to draw (or to skip). type is renderable {architecture, pipeline, framework, concept, schematic, overview, qualitative, diagram, flow} for a true depiction of the method, or skip-it {results, plot, curve, bar, chart} for a quantitative results figure (which stays a placeholder โ€” a proposal has no real data to plot):

\begin{figure}[htb]\centering
%% FIGURE-SPEC type=architecture
%% DESC: one concise line โ€” the exact boxes, arrows, and data flow to draw, in the paper's own terms
\fbox{\rule{0pt}{3cm}\rule{0.9\columnwidth}{0pt}}
\caption{...}\label{fig:...}\end{figure}

Emit at least template.figures.min (TS = 5) figure placeholders, distributed across method / experiments / analysis (per the blueprint). These are DRAWABLE schematic / conceptual / qualitative diagrams (architecture overview, a component-detail diagram, the formation-graph concept, a qualitative crossing-event scenario with NO numbers, the stratification-protocol concept, โ€ฆ) โ€” each one a real contribution to understanding, not padding. Results figures follow template.results_mode: in proposal mode (no data) emit NO type=results figure at all โ€” quantitative results go in the experiment TABLES (blank --); there is nothing to plot, so don't place or substitute a results figure. In data_aware mode emit results figures too, drawn from the real data. Use an **ex

Truncated for display โ€” read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars1.1k
CategoryContent
Updated2mo ago
Forks19

Languages

Python

Trust signals

100/100

From repository metadata: license, adoption, age and documentation. Not a code audit โ€” see the Safety scan above for what the skill file itself contains.

No cautions
ts-paper-write โ€” Universal Skill: Install & Safety Check | SkillAgent