GeoAI Skills

Geospatial AI should know when not to answer

A command can succeed while the geographic claim is still wrong.

GeoAI Skills is a vendor-neutral collection of 18 Agent Skills for the methodological layer between a user request and a geospatial claim. The skills do not replace GDAL, PostGIS, Earth Engine, ArcGIS, QGIS, Python libraries, or MCP servers. They tell an agent which preconditions must hold, which checks must run, what evidence must be reported, and when the available data cannot support the requested conclusion.

Two measurements are shown above and they are deliberately not pooled. The routing card covers 167 cases on Claude Code 2.1.214, claude-sonnet-5, suite efe27d8c1736…. The behavior package covers 92 clean pairs on Claude Code 2.1.222, the same model name, and suite 520bc41dd4c0…. Both describe one runtime/model pair and are not universal or model-independent claims. The behavior figures are model-judged, uncalibrated, and not yet human-verified.

Success is a defensible claim — or a justified stop.
Install · See the boundary · Inspect the evidence · Static poster

The animation uses a public, synthetic request. It demonstrates the method the skills encode; it does not claim that every runtime will produce identical text.

Quick start

Install all 18 skills for Codex:

npx skills add muend/geoai-skills --skill '*' -a codex

Or install one specialist for Claude Code:

npx skills add muend/geoai-skills \
  --skill remote-sensing-analysis \
  -a claude-code

Then ask naturally:

Can May 2024 and September 2025 Sentinel-2 scenes support a defensible changed-hectares claim?

The relevant specialist should test observation comparability before proposing pixel arithmetic. In this example the season mismatch blocks a defensible two-date area claim, so the useful answer is not a fabricated number: it is the reason to stop and the evidence needed to continue.

What makes this different

There are real neighboring projects. The difference is center of gravity, not a claim that every other skill collection is interchangeable or inferior.

What is unusual here is the combination:

1. Claim preconditions, not only API knowledge. A result must survive CRS, unit, validity, comparability, leakage, uncertainty, and provenance checks.
2. A safe failure mode. Missing evidence produces a narrower claim, clarification, provisional plan, or refusal—not invented certainty.
3. Routed specialists. One large geospatial prompt is not loaded for every task; domain ownership and collisions are explicit and tested.
4. Published misses. Routing evidence, paired behavior evidence, known regressions, execution errors, costs, suite hashes, and the failed quality gate remain visible.

Choose OpenMapStack when you want its executable open-stack project contract. Choose a platform collection when that platform is the problem. Use GeoAI Skills when the hard question is whether the method and evidence can support the claim at all. These projects can complement one another.

What changes when the skills are present

The skills complement tools and runtimes. They are the method and verification layer that tells an agent when not to trust an apparently successful operation.

Evidence with boundaries

First paired behavior measurement

The first behavior package compares skills enabled and disabled under the same runtime, model, prompts, and non-skill tool configuration. Of 93 responses in each arm, 92 formed clean matched pairs.

Two facts must travel together:

• The paired effect is large in this run. The enabled arm wins 65 of the 70 discordant criterion comparisons, and observed critical spatial failures fall from 13 cases to 4.
• The absolute quality level is not good enough. The enabled arm covers 48% of pinned criteria, but only 18/92 cases (19.6%) meet every criterion, and the critical-failure rate is 4.3%. The repository's Phase 3 gate applies to that case-level pass rate: it requires at least 85% plus less than 2% critical failures. It fails both conditions.

We published the gate we failed.

This is a single, different-family model judgment (gpt-5.6-sol, low reasoning) of responses from claude-sonnet-5. It is uncalibrated and has no completed human verification. Forty-seven arm-blinded human review packets are prepared; until adjudication lands, these figures are evidence about this run, not a claim of human-verified overall answer quality. Raw response text remains unpublished to avoid contaminating that review; hashes and lengths preserve identity.

Read the behavior evidence package for provenance, ties, criterion structure, the excluded termination case, cost, tokens, and limitations. The historical routing-only label behavior_quality-not_evaluated applies to the earlier routing card—not to the repository's current evidence inventory.

Four evidence layers, kept separate

The routing card and behavior package use different Claude Code versions and different suite hashes; do not combine their metrics. The external subset is independently authored and reported separately. It does not copy upstream datasets, prompts, or reference implementations, and its results are never pooled with native metrics. See its frozen suite and offline result protocol.

The routing-only run recorded 99.17% precision, 96.77% recall, and 96.41% full-route accuracy on suite efe27d8c1736…, with zero control activations. It also recorded one false positive, four false negatives, one incomplete route, and eleven execution errors; all remain in the card. Routing says which skill loaded, not whether the answer was correct.

Read BENCHMARK.md for routing and known defects, and BEHAVIOR.md for the separate behavior card. See EVALUATION.md for the provider-neutral prepare → execute → judge → score → compare protocol and ROADMAP.md for the open evidence and release gates.

If you want measured, pre-registered evidence like this to keep being published, star the repository so the work is easier to find and sustain.

The 18-skill stack

Install the full suite for cross-skill routing, or cherry-pick one specialist. Every skill must remain safe and useful alone; sibling references are advisory and critical safeguards have local fallbacks.

Installation

Choose the surface you already use.

npx skills add muend/geoai-skills --list

npx skills add muend/geoai-skills --skill '*' -a claude-code

npx skills add muend/geoai-skills --skill '*' -a codex

python tools/build_openai_plugin_bundle.py

claude plugin marketplace add muend/geoai-skills
claude plugin install geoai@geoai-skills

/plugin marketplace add muend/geoai-skills
/plugin install geoai@geoai-skills

/geoai:remote-sensing-analysis

GitHub Copilot

GitHub Copilot can load project skills from .github/skills, .claude/skills, or .agents/skills. With GitHub CLI 2.90.0 or later, preview before installing:

gh skill preview muend/geoai-skills remote-sensing-analysis
gh skill install muend/geoai-skills remote-sensing-analysis

The Skills CLI also provides an explicit Copilot target:

npx skills add muend/geoai-skills \
  --skill remote-sensing-analysis \
  -a github-copilot

See GitHub's Agent Skills documentation for supported locations, pinning, provenance, and security guidance.

python tools/build_skill_archives.py

Release maintainers should use RELEASING.md for the non-publishing candidate workflow, clean-install matrix, evidence requirements, publication gate, and rollback procedure.

Try these prompts

Reject an invalid change claim

Plan a defensible two-date Sentinel-2 forest-loss workflow. One scene is from May 2024 and the other from September 2025. Decide whether this comparison can support a changed-hectares claim.

Expected intervention: identify the phenology mismatch; withhold the hectare claim; request matched-season imagery or a season-aware time series; require registration, mutual masks, sensitivity analysis, and error-adjusted area uncertainty.

Make spatial SQL measurable and index-safe

Design production-safe PostGIS SQL to return flooded area in hectares per parcel for two large EPSG:4326 polygon tables. Include invalid-geometry handling and an EXPLAIN verification plan.

Expected intervention: select a suitable projected measurement CRS, avoid transforming indexed columns inside predicates, repair derived working geometry, deduplicate overlaps, and verify the query plan.

Refuse inconsistent suitability weights

Use this AHP pairwise matrix to build a solar-suitability map. Its consistency ratio is 0.19.

Expected intervention: reject the weights before mapping, identify discordant judgments, require revision, and preserve sensitivity analysis as a deliverable.

How it works

1. Trigger narrowly. Specialist descriptions identify the domain decision that owns the task; negative and collision cases test over-triggering.
2. Load progressively. The agent reads only the relevant SKILL.md, then loads references or scripts when required.
3. Enforce invariants. CRS, validity, leakage, units, uncertainty, provenance, verification, and failure behavior travel across the workflow.
4. Route across specialists. The orchestrator coordinates genuine multi-stage work without replacing specialist judgment.
5. Fail loudly. Missing inputs produce a narrower claim, provisional plan, clarification, or refusal—not invented evidence.

Design principles

1. Fail-loud spatial computing. Successful execution is not proof of a valid geographic result.
2. Methodological honesty. Uncertainty, multiple-testing correction, sensitivity, and error-adjusted estimates are first-class outputs.
3. Anti-leakage by default. Spatial and grouped validation replace random splits whenever geographic generalization is claimed.
4. Tool-pragmatic, vendor-neutral guidance. Open Python, PostGIS/DuckDB, Earth Engine, ArcGIS, and other backends are selected by evidence and scale.
5. Progressive disclosure. Long references and scripts cost no context until the task actually needs them.
6. Measured, not assumed. Source cases, run evidence, suite hashes, errors, costs, limitations, and superseded results remain distinguishable.

geoai-skills/
├── skills/<skill-name>/
│   ├── SKILL.md
│   ├── agents/openai.yaml
│   ├── scripts/
│   └── references/
├── evals/cases/<skill>/       # native cases governed by the committed split
├── evals/external/            # separately reported transfer suites
├── tools/                     # validators, adapters, builders, and gates
├── benchmarks/                # immutable published evidence packages
├── .codex-plugin/             # OpenAI skills-only plugin manifest
├── .claude-plugin/            # Claude marketplace and plugin manifests
├── BENCHMARK.md               # routing card and limitations
├── BEHAVIOR.md                # paired behavior card and open quality gate
├── EVALUATION.md              # provider-neutral evaluation protocol
├── CASE_STUDIES.md            # evidence policy and accepted cases
├── ROADMAP.md                 # canonical current priorities and release gate
└── RELEASING.md               # release and rollback runbook

Contributing

Contributions are welcome. Start with CONTRIBUTING.md. A current task must also respect ROADMAP.md, while questions and reproducible defects belong in the channels described by SUPPORT.md. Project and evidence decisions follow GOVERNANCE.md. A skill change should:

• pass the structural and link validators;
• add or update evaluation cases;
• avoid duplicating canonical cross-cutting rules;
• define verification and failure behavior; and
• keep runtime packages free of benchmark and private development material.

Real-world catches belong in CASE_STUDIES.md only when they include a dated, privacy-safe reproducer, comparison evidence, verification method, and limitations.

License

MIT — use it, fork it, and ship it.