Atmospheric Science Research
Deliver reproducible science and figures that communicate evidence at publication size.
Working contract
- User instructions take precedence over skill defaults. Preserve the project's paths, backend, and conventions; use Python/xarray/Matplotlib for a new workflow unless the user chooses otherwise.
- Infer answers from files and prior decisions. Ask only when missing information changes the scientific result or a required deliverable. State reversible assumptions and continue independent work; routine style choices do not require approval.
- Keep one brief record of the question, data/period, baseline, estimator, uncertainty, panel roles, and export target in the existing plan or project notes. Establish the question before analysis; write the conclusion only when supported by results.
- Treat data integrity and honest uncertainty as requirements. Treat fonts, palettes, layout, and directory examples as defaults, subject to the actual journal and user brief. Never manufacture observations or significance to complete a figure.
- Default to PNG and editable PDF exports. Honor user or journal requirements for other formats with an appropriate exporter.
Load on demand
Read only applicable references, once unless changed. Paths are relative to this skill.
| Task |
Reference |
| Inspect, compute, or change scientific results |
compute.md |
| Create or revise a figure |
plot-standards.md |
| Maps, vectors, pressure levels, or terrain |
maps.md |
| Verify a rendered figure |
review.md |
| Handoff or update analysis documentation |
readme-template.md |
Workflow
Run only requested stages. Compute-only work ends with numerical checks and provenance; a plot patch reuses valid data.
- Inspect. Read metadata first: variables, dimensions, units, calendar, coordinates, coverage, chunking. Use an available header tool or a short Python probe; inspect bounded samples for ambiguities. Batch independent reads and return summaries, not arrays.
- Compute, if needed. Resolve scientific choices; implement with the compute reference. Preserve raw inputs. Save reusable derived data with provenance before plotting. Reuse only outputs whose inputs and scientific settings still match. For long jobs, draft plotting against the known output schema while computation runs.
- Compose. Give each panel an evidence role; share encodings for comparable quantities. Set physical dimensions before typography. Use the plot reference and, where relevant, the map reference. When the project lacks a suitable exporter, optionally use figure_export.py for scoped Matplotlib defaults and fixed-canvas PNG/PDF exports.
- Verify and deliver. Open the rendered figure, check relevant numerical invariants and final exports, then follow the review reference. Fix observed issues in one batch per iteration; stop after the checks pass. Deliver applicable code, derived/source data, requested exports, captions, and concise QA/provenance notes. Record partial results and blockers when work cannot finish.
Patch scope and effort
For figure-only changes, read existing code and data provenance, reuse valid outputs, and render only affected figures. If a scientific bug is found, fix and verify the affected computation within scope; ask only for a missing scientific decision.
Keep checks proportional: a label edit needs visual/export checks; a changed estimator needs numerical verification. Do not rerun expensive analysis or broaden testing after relevant checks pass without new evidence. Report results, consequential assumptions, and blockers concisely in the user's language; keep code, comments, and logs in English.
1---2name: atmospheric-science-research3description: Analyze atmospheric and climate data, compute reproducible diagnostics, and create or revise publication figures from NetCDF, GRIB, HDF, Zarr, or derived tables. Use for climate analyses and manuscript maps or plots; not generic illustration.4---56# Atmospheric Science Research78Deliver reproducible science and figures that communicate evidence at publication size.910## Working contract1112- User instructions take precedence over skill defaults. Preserve the project's paths, backend, and conventions; use Python/xarray/Matplotlib for a new workflow unless the user chooses otherwise.13- Infer answers from files and prior decisions. Ask only when missing information changes the scientific result or a required deliverable. State reversible assumptions and continue independent work; routine style choices do not require approval.14- Keep one brief record of the question, data/period, baseline, estimator, uncertainty, panel roles, and export target in the existing plan or project notes. Establish the question before analysis; write the conclusion only when supported by results.15- Treat data integrity and honest uncertainty as requirements. Treat fonts, palettes, layout, and directory examples as defaults, subject to the actual journal and user brief. Never manufacture observations or significance to complete a figure.16- Default to PNG and editable PDF exports. Honor user or journal requirements for other formats with an appropriate exporter.1718## Load on demand1920Read only applicable references, once unless changed. Paths are relative to this skill.2122| Task | Reference |23|---|---|24| Inspect, compute, or change scientific results | [compute.md](references/compute.md) |25| Create or revise a figure | [plot-standards.md](references/plot-standards.md) |26| Maps, vectors, pressure levels, or terrain | [maps.md](references/maps.md) |27| Verify a rendered figure | [review.md](references/review.md) |28| Handoff or update analysis documentation | [readme-template.md](references/readme-template.md) |2930## Workflow3132Run only requested stages. Compute-only work ends with numerical checks and provenance; a plot patch reuses valid data.33341. **Inspect.** Read metadata first: variables, dimensions, units, calendar, coordinates, coverage, chunking. Use an available header tool or a short Python probe; inspect bounded samples for ambiguities. Batch independent reads and return summaries, not arrays.352. **Compute, if needed.** Resolve scientific choices; implement with the compute reference. Preserve raw inputs. Save reusable derived data with provenance before plotting. Reuse only outputs whose inputs and scientific settings still match. For long jobs, draft plotting against the known output schema while computation runs.363. **Compose.** Give each panel an evidence role; share encodings for comparable quantities. Set physical dimensions before typography. Use the plot reference and, where relevant, the map reference. When the project lacks a suitable exporter, optionally use [figure_export.py](scripts/figure_export.py) for scoped Matplotlib defaults and fixed-canvas PNG/PDF exports.374. **Verify and deliver.** Open the rendered figure, check relevant numerical invariants and final exports, then follow the review reference. Fix observed issues in one batch per iteration; stop after the checks pass. Deliver applicable code, derived/source data, requested exports, captions, and concise QA/provenance notes. Record partial results and blockers when work cannot finish.3839## Patch scope and effort4041For figure-only changes, read existing code and data provenance, reuse valid outputs, and render only affected figures. If a scientific bug is found, fix and verify the affected computation within scope; ask only for a missing scientific decision.4243Keep checks proportional: a label edit needs visual/export checks; a changed estimator needs numerical verification. Do not rerun expensive analysis or broaden testing after relevant checks pass without new evidence. Report results, consequential assumptions, and blockers concisely in the user's language; keep code, comments, and logs in English.
zsy1207