Light-skills

Turn research questions, evidence, experiments and papers into one project you can keep moving forward.

Website · English · 简体中文

Install · Skills · Workflow · Project layout · Data figures · Method diagrams · Feedback

Light-Science (formerly Light-skills) is a research skills pack for AI agents. It connects literature, ideation, design, experiments, analysis, journal selection, writing, figures and review. Project memory keeps the reasons behind research decisions, and a leader skill coordinates the next step, so every revision can be traced back to a concrete source, version and artifact.

Start from scratch, or bring an existing paper, dataset, codebase or set of reviewer comments and go straight to the stage you need.

The skill instructions are written in Chinese, and several skills reply in Chinese by default. Ask your agent to answer in English if you prefer.

Installation

Let your agent install it

Send this to your agent:

Install the research skills from https://github.com/Light0305/Light-Science.
Ask me first whether to install globally or for the current project, then install the skills I need from the skills table in the README.

Use the skills CLI

Windows, in PowerShell:

npx.cmd skills@latest add Light0305/Light-Science
npx.cmd skills@latest add https://github.com/Light0305/Light-Science/tree/master/light-chart/skills/light-chart

macOS / Linux:

npx skills@latest add Light0305/Light-Science
npx skills@latest add https://github.com/Light0305/Light-Science/tree/master/light-chart/skills/light-chart

The first command lists the top-level skills; the second adds Chart, which lives in a nested folder. Pick the same agent and install scope both times, otherwise Chart ends up somewhere else. The Diagram entry keeps its four sub-skill folders, and each sub-skill can also be installed on its own. For a project-level install, change into the target project first. You can also install a single skill:

npx skills@latest add Light0305/Light-Science --skill light-write

More detail on install locations, dependencies and invocation is in the installation guide (Chinese).

Start from a real question

Use light-leader to move this project forward: I want to study <specific question>, and I already have <papers/data/code>.
The project lives at <project path>, and the goal is <a paper / a software copyright and related outputs>.
First read the existing project memory and the actual materials, find the most pressing problem, then complete the step that can be done now.

To continue an existing project:

Continue <project path>. Read Memory and the latest handoff first, then check the adopted versions of the manuscript, figures and analyses.
This round, handle <research question or reviewer comments>, and write the results, the reasons for each decision and the next step back into the project.

Replace everything in angle brackets before sending. For a focused task, just say “use light-chart to plot this data” or “use light-write to revise this section”.

Skills

15 main skills plus 4 dedicated Diagram entries: 19 in total.

The Diagram entries can be used on their own or called by the main entry as needed:

light-analysis was named light-anaysis in v2.0.0 and earlier. If you installed the old name, reinstall it under the new one; old project records that mention light-anaysis refer to the same skill.

How a project moves forward

Research is allowed to go backwards: analysis may expose problems in execution, a weak design or a gap in the research question itself, and review may call for more analysis, more experiments or a different target journal. Data acquisition, patents and software copyrights join when the project needs them, and can be prepared early when publication timing matters.

The leader restores, then verifies, then acts. It reads the relevant project memory and handoffs, goes back to the papers, code, run results and adopted figures to check the actual state, and decides which skill takes the next step. Only when several steps need coordinating does it keep goals, dependencies, progress, acceptance criteria and a recovery entry in Workflow/.

Memory lets research pick up where it left off. Why the direction changed, which plan failed and which constraints still hold are all recorded with their sources. When a source changes or sources conflict, it reminds you to re-verify; the scientific facts themselves stay in each stage’s original files. Real progress from research or revisions is written back to memory, and the next round starts from there.

Project layout

The skill install directory holds capabilities; the research project directory holds results. Below are the main parts a project may use. Start from the actual task: files and folders appear only when there is content for them. An existing project structure takes priority. The full rules are in the research project layout spec (Chinese).

ProjectName/
├── README.md                          # Research question, stage, conclusions and next step
├── project.json                       # Optional cross-stage index
├── Memory/
│   └── memory.json / events/ / current.md
├── Workflow/
│   └── workflow.json                  # Goals, tasks and recovery entry, when needed
├── Literature/
│   ├── manifest.json / Searches/
│   └── Year_Author_ShortTitle/
│       └── meta.json / paper.pdf / paper.md / paper.zh.md / note.md / images/
├── Idea/IDEA-0001_Topic/
│   └── context.json / ideas.json / decision.md / cards/
├── Design/PROT-0001_Plan/v01/
│   └── design.json / plan.md / protocols/
├── Experiment/
│   ├── Code/ / Environment/ / Data/    # Code, environment or controlled references
│   └── EXP-0001_Batch/
│       └── campaign.json / execution-report.md / runs/R001/
├── Analysis/AN-0001_Question/
│   └── analysis-plan.json / analysis-report.md / code/ / derived/
├── Journal/JRN-0001_Round/
│   └── selection-report.md / light-write-brief.md
├── Write/PAPER-0001_WorkingTitle/
│   └── outline.md / claim-evidence.md / sections/ / drafts/ / build/
├── Figure/
│   ├── Methods/FIG-0001_MethodFigure/  # Intent, prompts, references, native files and exports
│   ├── Mechanisms/FIG-0002_Mechanism/
│   └── Data/FIG-0003_DataFigure/       # Data mapping, code, caption and exports
├── Review/REV-0001_ManuscriptVersionRound/
│   └── review.md / packet/
├── Acquisition/ACQ-0001_Target/case/
├── Patent/PAT-0001_Invention/
│   └── case.json / work/ / drafts/ / figures/
└── SoftwareCopyright/SWC-0001_SoftwareVersion/
    └── 正式资料/<version>/              # Final filing materials (Chinese file names required by the filing)
        ├── <software>_功能说明.docx     # Feature description
        ├── <software>_软件代码.docx     # Source code
        └── <software>_软件说明书.docx   # User manual

Ideas, designs, runs, analyses and figures can all keep several rounds of versions, all serving the same core output.

Data figures

Choose the chart from the scientific question and the data structure; keep experimental units, repeated measurements, missing values and uncertainty; then handle layout, color, annotation and export. Python and R are both supported, including multi-panel figures, omics, follow-up, signals, ecology, spatial fields and image panels.

01 · Measured EEG Raw time series and band power	02 · Ordination and heatmap Community composition along an environmental gradient	03 · 3D field and slices Volume data, spatial calibration and matching views
04 · Temperature field and time series Real coordinates, year-over-year comparison and data support	05 · Plant response, repeated measures Individual trajectories and pairing kept	06 · Growth and follow-up Longitudinal change and the actual number of observations
07 · Raincloud with raw points Distribution shape, quartiles and individual observations	08 · Method comparison and heatmap Organizing reported values, hierarchy and annotation	09 · Survival and number at risk Follow-up support and uncertainty

Data sources and terms for each figure are in the showcase notes and manifest.

Method and mechanism diagrams

Read the project materials first to pin down the scientific question, objects and relationships the figure must explain. Then, figure type and reading path first, domain second, pick visual references from the library, from the user or from related papers. Extract layout, hierarchy, color and arrow language into a structural blueprint and a high-quality prompt, generate the image for real, and check labels, relationships and evidence boundaries.

When the figure needs further editing, break the chosen reference into objects, text, layers and relationships, and rebuild it along one of three routes: Illustrator, PowerPoint or Visio, aiming to match the reference’s composition and scientific relationships. The native file is saved, reopened and checked with the matching editing behaviour before previews are exported.

Generated reference vs. native reconstruction

These are real cases. The left column is the generated reference; the right column is the actual export from the matching tool. Click an image to view it at full size; the native files can be downloaded and inspected.

Tool and case	Generated reference	Native reconstruction export
Illustrator Lithium–sulfur battery mechanism AI file
PowerPoint CRISPR–Cas9 PPTX file
Visio MonoFormer architecture VSDX file

Feedback and support

Use cases, problems, missing domain support and reproducible suggestions are all welcome.

• Open an issue: describe the skill you used, what you expected and what happened, and attach minimal reproducible material.
• Email: [email protected]
• GitHub: @Light0305
• Related projects: Light-UI-UX · Light-WeChat
• AI Shop：LightChuan-Shop

Support the author	WeChat	Official account

License

Original code and documentation are under the MIT License. Third-party code, adapted resources, paper figures, example data and galleries keep their own licenses, so the pack as a whole is not uniformly MIT. See third-party notices (Chinese).