sci-using-kit

English | 中文

An evidence-bounded SCI manuscript workflow skill family for Codex.

sci-using-kit gives research-writing agents a superpowers-style entry point and a set of branch skills for literature vetting, paper revision, evidence-to-discussion writing, figure generation, comparative positioning, reviewer preflight, LaTeX editing, and submission packaging.

The goal is not to auto-generate a paper. The goal is to keep claims, citations, experiments, and manuscript prose aligned.

Why This Exists

Scientific writing agents often fail in predictable ways:

• They let literature-search context pollute the main writing context.
• They turn weak evidence into strong claims.
• They compare against prior work with incompatible tasks, hardware, datasets, or metrics.
• They polish prose by deleting necessary caveats.
• They create attractive but unsupported figures with wrong language, invented labels, misleading axes, or private infrastructure details.
• They publish source packages with private paths or local-only assumptions.

This kit gives agents a structured path:

sci-using-kit
  -> classify project phase
  -> emit Routing Decision
  -> choose a branch skill
  -> preserve prompt kernels
  -> report evidence, risks, and verification

Installation

Windows PowerShell

git clone https://github.com/keithhegit/sci-using-kit.git
cd sci-using-kit
.\scripts\install.ps1

macOS / Linux

git clone https://github.com/keithhegit/sci-using-kit.git
cd sci-using-kit
bash scripts/install.sh

By default, the scripts copy all folders under skills/ into:

• Windows: %USERPROFILE%\.codex\skills
• macOS/Linux: $HOME/.codex/skills

You can also copy the skill folders manually.

Starting Workflow

In a new Codex thread, start with:

Use $sci-using-kit for this SCI manuscript task.

The entry skill should produce:

**Routing Decision**
- Project phase:
- Task type:
- Primary branch skill:
- Additional gate skills:
- Evidence bundle to read:
- Output expected:

Project phases:

• kickoff
• drafting
• revision
• arXiv preflight
• SCI Track B
• repo release

Skills Library

Skill	Use when
sci-using-kit	Starting or resuming any SCI paper task.
sci-issue-intake	A request, reviewer comment, or advisor note is vague or multi-step.
sci-literature-vetting	Adding, checking, or repairing citations and related work.
literature-review-search	Searching and organizing references into citation-safe literature tables.
sci-paper-revision	Rewriting or polishing manuscript prose without strengthening claims.
sci-evidence-discussion	Turning tables, logs, figures, HIL/simulation outputs, or measurements into results/discussion text.
sci-figure-generation	Planning, prompting, generating, validating, captioning, and packaging evidence-bounded SCI/arXiv figures.
sci-comparative-positioning	Claiming novelty, speedup, efficiency, improvement, or value over prior work.
sci-reviewer-preflight	Auditing a manuscript before arXiv, advisor review, or SCI submission.
sci-latex-editing	Editing live LaTeX sources, Overleaf exports, figures, citations, labels, or appendix files.
sci-submission-preflight	Preparing arXiv, journal, conference, source zip, metadata, or public-release packages.

Methodology Core

This kit uses a superpowers-style shell with an SCI evidence-writing soul.

It preserves these prompt kernels:

• State recovery: important claims, evidence, decisions, and open questions should survive across sessions.
• Literature isolation: literature-heavy work should produce compressed matrices, not pollute the main writing context.
• Claim/evidence discipline: do not fill evidence gaps with confident prose.
• Evidence-to-discussion separation: distinguish observation, interpretation, scope, limitation, and next evidence.
• Figure-as-claim discipline: treat labels, axes, units, visual comparisons, and captions as claim surfaces that need evidence and privacy checks.
• Comparative legitimacy: no "better/faster/improved" claim unless task, data, hardware, metric, and evidence level are comparable.
• Reviewer perspective: find the objection before the reviewer does.
• LaTeX safety: make section-scoped edits and avoid changing structure and scientific content in the same pass.

The full prompt-kernel reference lives at:

skills/sci-using-kit/references/prompt-kernels.cn.md

Repository Layout

sci-using-kit/
  README.md
  README.cn.md
  LICENSE
  .gitignore
  skills/
    sci-using-kit/
    sci-issue-intake/
    sci-literature-vetting/
    literature-review-search/
    sci-paper-revision/
    sci-evidence-discussion/
    sci-figure-generation/
    sci-comparative-positioning/
    sci-reviewer-preflight/
    sci-latex-editing/
    sci-submission-preflight/
  scripts/
    install.ps1
    install.sh
    validate_skills.py
  .github/workflows/
    validate.yml

Version Notes

v0.2.0 - SCI figure generation branch

• Added sci-figure-generation for evidence-bounded graphical abstracts, system/workflow diagrams, result snapshots, image-generation prompts, captions, and Overleaf figure assets.
• Added Figure-as-Claim discipline: figure labels, axes, units, visual comparisons, and captions must be checked against evidence.
• Added safeguards for English-only arXiv figures, mixed-unit visualizations, image-generation hallucinations, and private infrastructure leakage.

v0.1.0 - Initial SCI skill family

• Introduced the entry skill and core branch skills for literature vetting, manuscript revision, evidence-to-discussion writing, comparative positioning, reviewer preflight, LaTeX editing, and submission packaging.

Validation

python scripts/validate_skills.py

On Windows PowerShell:

$env:PYTHONUTF8='1'
python scripts\validate_skills.py

Important Boundary

This repository intentionally does not include:

• private manuscripts
• local research logs
• unpublished project data
• local Windows paths
• paper-specific drafts
• workflow audit reports

Only reusable skill folders and public helper scripts belong here.

License

MIT License. See LICENSE.