spark-to-paper-skills

Citation checks. Editable vector figures. Source-traced numbers.

One orchestrator and 13 composable Claude Code skills turn a one-line idea into a compiled PDF —
real references, editable vector figures, and machine-checked integrity included.
No separate app or orchestration server.

📣 The preprint is on arXiv:2608.11924 — #1 on 🤗 Hugging Face Daily Papers (Aug 13, 2026) and included in the August 2026 Monthly Papers list.

📄 Paper · 🌐 Website · 🏆 Showcase · ✨ Features · 🧭 Compare · 🖼️ Figure Engine · 🔬 Pipeline · 🚀 Quick Start

7 papers across 6 domains — environmental monitoring, energy forecasting, environmental AI, computer vision, clinical AI, and bearing fault diagnosis — generated fully end-to-end with real citations, editable vector figures, and compiled PDF output. Conference-format samples lead the set: the official ICML 2025 style plus two papers in the official NeurIPS 2025 style.

• 2026-08-13 — 🤗 Hugging Face Daily Papers #1. The preprint reached #1 on Daily Papers and included in the August 2026 Monthly Papers list.
• 2026-08-12 — 📄 Preprint submitted. Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill is available as arXiv:2608.11924. See Citation.
• v1.2.0 — PaperBanana+ figure engine. The official PaperBanana renders candidates; the new ts-figure-svg skill learns the render's design language and redraws the figure natively from the paper's facts, iterating against a stdlib-only geometry audit (audit_svg.py) until it passes.
• v1.1.0 — Claude Code plugin support. Restructured as a proper plugin with .claude-plugin/plugin.json. One-command install, auto-loads on session start.
• v1.0.1 — Soft update notification. check_update.py queries GitHub Releases API on each run (24h cache, silent when up-to-date, never blocks).
• v1.0 — Initial release. 13 skills, end-to-end pipeline, hybrid vector figure engine, MIT License.

# Install — auto-loads on next Claude Code session
git clone https://github.com/Spark-To-Paper-Skills/spark-to-paper-skills.git ~/.claude/skills/spark-to-paper-skills

Run ts-paper on this proposal.   ← paste your idea, proposal, or data

The orchestrator auto-routes your input, picks the right mode, and runs the full chain.

_{✓ full  ·  ● partial  ·  – none  |  sources: ARS · Idea2Paper · AutoResearchClaw · AI-Scientist · Kosmos · karpathy/autoresearch · auto_research}

Based on the linked project documentation reviewed in August 2026, the heavier autonomous scientists (AutoResearchClaw, AI-Scientist) match the breadth but ship as standalone Python products. The lighter skills in this comparison (ARS, Idea2Paper) do not cover the same experiment-and-figure path. This comparison is scoped to the listed projects, not the entire ecosystem.

A core differentiator of this project. AI image models produce rasters — but a paper needs editable vector figures whose labels are text and whose logic is the paper's.

ts-figure-svg is PaperBanana+: the official PaperBanana renders candidates, then the figure is redrawn natively — its design language learned from the render, its content re-derived from the paper, repaired against a measuring audit.

figure brief (from the paper's own method text)
   ├─ PaperBanana      → Retriever→Planner→Stylist→Visualizer→Critic   (official, pulled)
   ├─ pick a candidate → scientific correctness first, beauty last
   ├─ learn the STYLE  → palette · type scale · spacing · idiom  →  <label>.style.json
   ├─ redraw NATIVE    → real <rect>/<path>/<text>, content from the PAPER, never a pixel trace
   └─ audit & repair   → iterate until the audit passes and the result is approved
   │
   └─▶ editable SVG · vector PDF   (live text, verified at real column width)

audit_svg.py measures what eyes miss — and it is stdlib-only (Adobe core-14 Times metrics, no renderer, no key, no models):

Fallbacks, in order: native redraw → local raster-preserving optimization → keep the approved PNG. Never a lossy redraw, never a flat boxes-and-arrows regression.

python skills/ts-figure-svg/scripts/setup_paperbanana.py            # pull official PaperBanana + deps
export OPENROUTER_API_KEY="sk-or-v1-..."                            # VLM agents + image generation
python skills/ts-figure-svg/scripts/audit_svg.py --selftest         # the audit needs nothing else

One spark in. One paper out.

One orchestrator (ts-paper) routes the input, then drives a 7-stage chain plus auto-run experiments:

                         ┌──────────────── optional upstream ────────────────┐
 [corpus.jsonl] ─▶ ts-kg-build ─▶ kg/         (research-pattern KG)
 [raw idea]     ─▶ ts-idea2story ─▶ story + citation seed
                         └──────────────────────────────────────────────────┘
                                                │
 [proposal  OR  proposal + real results]  ─────▶ ts-paper  (Stage 0: ROUTE)
                                                │
   1. ts-paper-plan ──▶ blueprint.json          title · keywords · contributions
   2. ts-paper-cite ──▶ refs.bib                ≥40 REAL refs via WebSearch + Crossref
   3. ts-paper-write ─▶ sections/*.tex          all sections in one holistic pass
   4. ts-paper-refine ▶ right-size + de-AI      scrub + logic self-check
   5. ts-paper-review ▶ adversarial review      multi-reviewer hardening
   6. ts-paper-figure ▶ figures + native SVG    PaperBanana → learn style → redraw + audit
   7. ts-paper-latex ─▶ main.pdf                assemble + compile
                                                │
   8. ts-paper-experiment (AUTO) ─▶ run feasible experiments, fill tables, recompile

Four complementary layers — Claude handles judgment, code provides the deterministic backstop.

python skills/ts-paper/scripts/run_gates.py <workdir> all     # nonzero exit = NOT done

Write to whatever venue you pick — content quality is invariant.

Add a venue by dropping a templates/<name>/ dir with template.json + LaTeX assets — no code changes.

1 orchestrator plus 13 composable skills; all active.

Option A — Claude Code plugin (recommended)

git clone https://github.com/Spark-To-Paper-Skills/spark-to-paper-skills.git ~/.claude/skills/spark-to-paper-skills

Auto-loads on next session. Skills available as /spark-to-paper:ts-paper, etc.

Option B — Try before you install

git clone https://github.com/Spark-To-Paper-Skills/spark-to-paper-skills.git
claude --plugin-dir ./spark-to-paper-skills

Option C — Standalone skills (no namespacing)

cp -r spark-to-paper-skills/skills/ts-* ~/.claude/skills/

Option D — Git submodule (auto-updatable)

git submodule add https://github.com/Spark-To-Paper-Skills/spark-to-paper-skills.git .claude/skills/spark-to-paper-skills

💡 The suite checks GitHub for newer versions on each run. Update: git -C ~/.claude/skills/spark-to-paper-skills pull

Copy .env.example → .env and fill in only what you use.

Run ts-paper on this proposal.

Paste your idea, proposal, or proposal + data. The orchestrator auto-routes, runs the chain, and delivers a compiled paper with page count, sections, references, review outcome, and editable vector figures.

• Claude Code (the suite is a plugin)
• Python 3.10+ with pip install -r skills/ts-figure-optimize/requirements.txt
• LaTeX (latexmk + a TeX distribution) for compilation
• Optional: local raster fallback runtime (~4 GB) · image-model endpoint · LibreOffice

main.pdf exists and is non-trivial · zero LaTeX errors · main.bbl resolved all citations · every \cite{} maps to a complete refs.bib entry · no fabricated numbers anywhere · every figure embedded as an editable vector PDF · the adversarial review stage ran · and run_gates.py <workdir> all exits zero.

A paper produced end-to-end by Spark-to-Paper has been accepted by an SCI-indexed Q2 journal. Until the version of record is public, please cite the arXiv version below. The preprint reached #1 on 🤗 Hugging Face Daily Papers (Aug 13, 2026) and is included in the August 2026 Monthly Papers list.

If Spark-to-Paper helps your research, please cite:

@article{qian2026sparktopaper,
  author  = {Qian, Zhuoyang and Wu, Biao and Wang, Yiran and Yan, Chris D and
             Dai, Desan and Zheng, Liangwei and Jiang, Jin and
             Zhang, Junsheng and Wang, Wenhao},
  title   = {Spark-to-Paper: End-to-End Research Paper Generation
             as a Composable Skill},
  journal = {arXiv preprint arXiv:2608.11924},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.11924}
}

Inspired by:

• 🔬 AI-Scientist (Sakana AI) — Automated research pioneer
• 🦞 AutoResearchClaw (AIMING Lab) — 23-stage autonomous research pipeline
• 📚 academic-research-skills (Imbad0202) — Claude Code research skill suite
• 🧠 autoresearch (Andrej Karpathy) — End-to-end research automation

_{Live star chart paused — GitHub restricted third-party access to stargazer data. It returns once the star-history GitHub App is installed on this repo.}

The model does the reasoning. The code keeps it honest. You get a paper.
_{Built on Claude Code · native editable SVG figure engine · MIT License}