🇨🇳 中文版

Google Earth Engine Master Utility (GEEMu).

Google Earth Engine Master Utility (GEEMu) is a compact full skill folder for OpenAI Codex and Claude Code. It helps write Google Earth Engine (GEE) Python workflows with geemap. It is designed for research-grade remote sensing tasks where the assistant must first check the local environment, understand the study design, reason about data and boundaries, then write reproducible GEE Python code.

• OpenAI Codex: install GEEMu as a local skill folder, then ask Codex to use GEEMu for GEE/geemap tasks.
• Claude Code: install the same folder under the Claude skills directory, then ask Claude Code to use GEEMu.
• The skill is file-based. It does not require a separate vector-search server, embedding model download, private credentials, or the original development workspace.

GitHub repository:

https://github.com/sadassimov/geemu-skill

Install the complete folder — do not skip any files. The skill depends on the whole package, especially the local knowledge base gee_vector_db/chunks.jsonl and gee_vector_db/documents.jsonl (tens of MB): these JSONL files are what the retrieval workflow searches, so the skill will not work without them. git clone and GitHub's "Download ZIP" already include every file — just don't hand-pick files, exclude large/.jsonl files, or use a shallow/sparse/partial checkout.

For OpenAI Codex on Windows PowerShell:

git clone https://github.com/sadassimov/geemu-skill.git "$env:USERPROFILE\.codex\skills\GEEMu"

For Claude Code on Windows PowerShell:

git clone https://github.com/sadassimov/geemu-skill.git "$env:USERPROFILE\.claude\skills\GEEMu"

For OpenAI Codex on macOS/Linux:

git clone https://github.com/sadassimov/geemu-skill.git ~/.codex/skills/GEEMu

For Claude Code on macOS/Linux:

git clone https://github.com/sadassimov/geemu-skill.git ~/.claude/skills/GEEMu

1. Open https://github.com/sadassimov/geemu-skill.
2. Download the repository as a ZIP file.
3. Unzip it and rename the folder to GEEMu.
4. Put the folder here for Codex: ~/.codex/skills/GEEMu.
5. Put the folder here for Claude Code: ~/.claude/skills/GEEMu.

Windows users can use these equivalent locations:

• Codex: %USERPROFILE%\.codex\skills\GEEMu
• Claude Code: %USERPROFILE%\.claude\skills\GEEMu

GEEMu uses a layered design:

• SKILL.md: the main instruction layer. It tells the assistant how to start every GEE task, including project ID, authentication, geemap import, proxy reminders, research design, data-layer reasoning, boundary complexity, and run records.
• references/: focused reference layers for local environment checks, proxy setup, research design, data selection, administrative/vector boundaries, and export strategy.
• gee_vector_db/: a portable local knowledge database stored as JSONL text. It is searched by lightweight keyword scoring.
• awesome-gee-community-datasets/community_datasets.csv: a structured community dataset table used as one data-recommendation source.
• templates/: standard RUN.md and DATA_LAYER.md records for documenting parameter choices, data transforms, scale, region, export settings, and assumptions.
• examples/: complete reference tasks, including Sentinel-2 NDVI, tiled export, geemap local download, community dataset recommendation, NSW dNBR/NBR recovery, Landsat water-frequency mapping, Random Forest regression (GEE + scikit-learn), stratified 256x256 sample chips, and deep-learning water inference.
• scripts/check_local_environment.py: a deterministic checker for local Python, geemap, Earth Engine credentials, project ID, and geemap vector helper availability.

When a user asks GEEMu to write GEE code, the expected flow is:

1. Confirm or request the Google Cloud Project ID.
2. Check whether ee, geemap, credentials, and vector helpers are available.
3. Remind users in restricted networks to configure proxy environment variables before online GEE access.
4. Define the research question, study area, time windows, calculation scale, output products, and export destination.
5. Choose boundary data deliberately: administrative boundary, uploaded vector, local vector converted by geemap, or drawn geometry.
6. Assess boundary complexity and decide whether simplification, buffering, clipping, or tiled export is needed.
7. Choose data layers and transformations, then explain what each value or band represents before scaling, masking, thresholding, or exporting.
8. Write GEE Python code using geemap where appropriate.
9. Record the run and data layer decisions in Markdown.

Use prompts like these with Codex or Claude Code after installing the skill.

Use GEEMu to write a geemap-based Earth Engine Python workflow for the California wildfire region in the United States. First check the local Python environment, Earth Engine authentication, geemap availability, proxy reminder, and Google Cloud Project ID. Use GADM administrative boundaries to define the California study area, then narrow the analysis to the selected burned region or fire perimeter if one is provided. Use a pre-fire 3-year median NBR as the baseline, then evaluate post-fire recovery for 1 year after the fire. Compute dNBR to detect burned areas, then compute an NBR recovery ratio map relative to the baseline and post-fire condition. Use an appropriate scale for regional analysis, consider boundary complexity and whether tiled export is needed, export the recovery map, and write RUN.md plus DATA_LAYER.md documenting all choices.

Use GEEMu to create a Sentinel-2 NDVI workflow for a user-provided study area. Check authentication and project ID first, choose a cloud-masking strategy, explain the scale and band transformations, make a median composite, export the NDVI raster, and record the run decisions.

Use GEEMu to design a tiled export workflow for a large or complex boundary. Assess geometry complexity, recommend simplification or tiling, preserve the meaning of the boundary, and generate Earth Engine Python code that exports manageable tiles with documented scale and CRS choices.

Use GEEMu to write a geemap workflow that downloads ESA WorldCover 2020 directly from Earth Engine to a local GeoTIFF. Check the local environment, authentication, proxy needs, and Google Cloud Project ID first. Let the user provide either an uploaded ROI asset or a simple lon/lat bbox. Use ESA/WorldCover/v100, select the Map band, preserve categorical class values with nearest-neighbor resampling, choose a safe scale and dtype, and explain when direct local download should be replaced by tiled or Drive export. Write RUN.md and DATA_LAYER.md documenting the ROI, output path, class data, scale, CRS, and risks.

Use GEEMu to recommend candidate GEE datasets for a drought, climate, fire, or vegetation-monitoring task. Search the local knowledge database and the community dataset table, compare spatial scale, temporal coverage, variables, limitations, and export suitability before writing code.

Use GEEMu to map multi-year surface-water frequency from Landsat surface reflectance for my study area: mask clouds, compute NDVI/NDWI/MNDWI/EVI, flag water, remove hill shadow / high HAND / steep slopes, aggregate a yearly water frequency, classify it into none/seasonal/permanent, overlay GAIA impervious surface, and export. Use Landsat Collection 2.

Use GEEMu to train a Random Forest regression combining GEE and a local scikit-learn model: build a Landsat composite with spectral indices, sample it at my labelled points, export a training CSV, fit RandomForestRegressor, convert the trees to an Earth Engine classifier with geemap.ml, and apply it.

Use GEEMu to generate 256x256 training chips by stratified sampling: use ESRI 10 m LULC water as the label and a Sentinel-2 median as the image; sample 100 points over my ROI and download label + image chips in parallel.

Use GEEMu to run deep-learning surface-water inference: export Sentinel-2 tiles over my ROI with a fishnet, run a pretrained WatNet model on each tile, write a binary water-map GeoTIFF, and show how to upload the result with geeup.

GEEMu's own files (SKILL.md, references/, templates/, examples/, scripts/, and this README) are released under the MIT License; see LICENSE.

Bundled third-party material keeps its original terms: the Awesome GEE Community Datasets table (CC BY 4.0), Earth Engine documentation excerpts (CC BY 4.0 / Apache 2.0), geemap (MIT), and other sources represented in the local knowledge database. Review THIRD_PARTY_NOTICES.md before reuse or redistribution.