atmospheric-science-research

SKILL.md

Atmospheric Science Research

Keep the control plane small: choose one mode, load only the needed references, and keep compute separate from plotting.

Goal

Create a scientifically rigorous, visually appealing, and informative figure suitable for publication in a top-tier atmospheric science journal.

Modes

  • Full: Plan -> Execute -> RR
  • Execute-lite: Execute -> RR
  • RR-only: RR

Choose a mode first:

  • use Full for a new project, new dataset, new processing method, or any task that may change the scientific conclusion
  • use Execute-lite for a constrained patch to labels, ticks, units, legends, panel letters, colors, levels, spacing, filenames, or export settings; rerender the affected outputs, then continue into RR
  • use RR-only when the task is to inspect or revise existing figure outputs; stay in RR until the result is PASS or BLOCKED
  • if the science-change boundary is unclear after a light inspection, use Full

Reference loader

Load only what the current step needs:

Do not load every reference file up front.

1. Project structure

  • Use data_processed/ and figN/ only when the project has no established layout.
  • Save reusable derived results in the project's processed-data location when one exists; otherwise use data_processed/.
  • Keep figure work near the existing figure layout when one exists; otherwise use figN/.

2. Plan

For the Plan step:

  • load references/plan.md
  • load references/readme-template.md
  • inspect the data only enough to design the work
  • stop and ask the user if a scientific ambiguity remains
  • write or update a short Chinese README.md aligned with the actual project structure

3. Execute

For all computation work:

For all figure work:

If subagents are available and the work can be split cleanly, divide it into compute and plot responsibilities:

  • compute agent: ingest data, preprocess, run diagnostics, and save reusable outputs
  • plot agent: read prepared outputs, render figures, and apply plotting standards
  • integrate their outputs through reusable intermediates

If subagents are unavailable, keep the same split in separate modules or scripts.

Execution rules:

  • choose the smallest workflow and the simplest defensible implementation that fully answers the scientific task
  • keep heavy computation out of plotting scripts
  • use the minimum code that solves the problem
  • never recompute heavy diagnostics when only figure styling changed
  • save reusable intermediate results once, then reuse them
  • when modifying an existing project, touch only the files, comments, and formatting needed for the task; match the local style
  • for Execute-lite, keep the patch and rerender scope to the minimum relevant code and affected outputs
  • when no project structure exists, organize files using this default structure, using fig1/ as the local example:
    • processed data in data_processed/*.nc
    • plotting scripts in fig1/plot_*.py
    • optional preprocessing or intermediate computation scripts in fig1/compute_*.py
    • figure outputs in fig1/fig1_*.png and fig1/fig1_*.svg

4. RR (review & revision)

For the RR step:

  • load references/review.md
  • load references/plot-standards.md
  • open and review the actual output PNG, not only the code
  • if a figure fails plotting standards or obvious physical-sense checks, patch the minimum relevant code
  • repeat until the result is PASS or BLOCKED

5. README follow-through

Use one README.md across the project:

  • in Full, write the plan version after data inspection and before execution. in Execute-lite or RR-only, keep the existing README.md and update it at final handoff only when paths, outputs, or the image summary changed
  • update the delivery version at final handoff
  • keep the README path section aligned with the actual project directories and outputs
  • for later modifications, append only the new version notes and image summary

6. Web and uncertainty rule

Use search only after checking the local code, docs, and project notes when formulas, methods, or package usage remain unclear, or when debugging stalls without enough local evidence.

Just stop and ask the user when the uncertainty and ambiguity remains.

7. Keep the control plane minimal

Do not force the full flow for a tiny patch. Do not load every reference file up front. Do not create extra QC, gate, audit, or planning documents. Do not rewrite working compute code for a figure-only issue.

Weekly Installs
5
First Seen
8 days ago
Installed on
opencode5
gemini-cli5
antigravity5
claude-code5
github-copilot5
codex5