Paper-Agent-Skills

Evidence-aware research writing skills for AI agents. Revise a manuscript against supplied experiment records and references, with traceable suggestions, focused edits, and validation reports.

Includes academic paper revision, rebuttal preparation, and presentation storyboarding. The host agent performs the review; bundled Python tools package the evidence and validate edits. The tools require no model API key.

Available in the skills.sh directory.

Installation

Python 3.10+ and Node.js are required for this installation path. Install all three skills into the shared agent directory, which DSH can discover:

npx skills add Scodive/Paper-Agent-Skills --agent universal --skill '*' --copy --yes
python3 -m venv .paper-agent-venv
.paper-agent-venv/bin/python -m pip install -r .agents/skills/academic-paper-writing/scripts/requirements.txt

Each skill includes its executable tools and readable runtime. A source checkout is not required after installation. Other supported agents can be selected through the skills CLI.

As a DeepSeek Harness (dsh) plugin

The same three skills also ship as a DSH plugin bundle, which registers them as a bundled skill provider in the model's skill catalog:

dsh plugin --profile web add dsh-paper-agent-skills

Either installation path works. The plugin bundle requires DSH 0.2.1 or newer and needs no build step.

Usage

Open an agent session in your working directory and provide:

Use academic-paper-writing to revise paper.tex against results.csv and references.bib. Use .paper-agent-venv/bin/python. Complete preparation, full-text review, edit application, and checks. Save the run to runs/review-01 and retain unresolved evidence gaps.

Read the outputs in runs/review-01/results/:

• suggestions.md — recommendations and unresolved items.
• revised.md or revised.tex — the revised manuscript.
• changes.diff — the exact edits.
• checks.md — mechanical checks and evidence limitations.

Inputs: Markdown, single-file TeX, or text-based PDF; experiment records in JSON/CSV; references in BibTeX/JSON. See input formats.

Validation

Observed results on the supplied synthetic fixture; engineering checks and live agent execution are reported separately.

Evaluation	Observed result	Scope
Automated test suite	40/40 passed	Ingestion, integrity, rejection rules, and standalone bundles
Remote installation	3/3 skills; 49 bundle files matched	Public GitHub source installed with skills CLI
GitHub Actions	2/2 jobs passed	Python 3.10 and 3.13; run
Invalid response scenarios	9/9 rejected	Unknown evidence, unsupported numbers, stale edits, and related cases
Fixed-response replay	5 → 0 mechanical findings; 2 pending items retained	Pre-authored response, not a new model run
Live DSH skill loading	3/3 skills loaded	DSH 0.2.2-rc.1; skills installed by skills CLI 1.7.6
Live DSH paper revision	5 → 0 mechanical findings; 6 edits; 3 pending items retained	One independently generated synthetic-case response

The live DSH process completed successfully and its artifacts passed an independent recheck. Rebuttal and slide skills were loaded but did not receive end-to-end behavioral evaluation. These results do not establish writing-quality gains, scientific correctness, or acceptance-rate improvements.

Distribution results · Engineering results · DSH results · DSH-generated response · Methodology

Development

python3 -m pip install -e '.[collection]'
python3 scripts/build_skill_bundles.py --check
python3 scripts/validate_release.py --out output/validation.json

For a local output preview: python3 examples/revision/demo.py --out output/demo-01. This replays a pre-authored synthetic review. Source workflows and additional commands are documented in the detailed guide.

MIT License. Collected papers retain their original licenses. Experimental records are not reproduced, references are not independently verified, and PDF layout is not preserved by the tools.