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.
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.
| Public project | Its strongest center of gravity | How GeoAI Skills differs |
|---|---|---|
| OpenMapStack | An open-first, reproducible GIS project contract with a CLI, templates, artifact validation, and worked output | GeoAI Skills is not an execution framework. It supplies separately routed method specialists across proprietary and open stacks, with claim-narrowing and refusal conditions before or around execution. |
| geospatial-skills | Separately installable skills for tools, formats, catalogues, viewers, and large-scale pipelines | GeoAI Skills organizes around analytical decisions and failure modes—leakage, comparability, inference, uncertainty, measurement, and safe mutation—rather than a tool catalogue. |
| GIS Agent Skills | Practical ArcGIS/GIS workflow, readiness, schema, publishing, and post-run checklists | GeoAI Skills spans the full data-to-claim lifecycle and evaluates cross-skill routing boundaries as well as behavior. |
| MapLibre Agent Skills and Mapbox Agent Skills | Deep platform-specific mapping implementation; MapLibre also distinguishes eval-verified and provisional skills | GeoAI Skills is platform-neutral and extends beyond application delivery into remote sensing, spatial inference, geostatistics, accessibility, ML validation, LiDAR, terrain, and databases. |
| GeoMaster | One broad, example-rich geospatial science skill | GeoAI Skills splits ownership across 18 narrow specialists so activation, collisions, negative routes, and domain boundaries can be tested independently. |
What is unusual here is the combination:
- Claim preconditions, not only API knowledge. A result must survive CRS, unit, validity, comparability, leakage, uncertainty, and provenance checks.
- A safe failure mode. Missing evidence produces a narrower claim, clarification, provisional plan, or refusal—not invented certainty.
- Routed specialists. One large geospatial prompt is not loaded for every task; domain ownership and collisions are explicit and tested.
- 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
| A plausible shortcut | GeoAI Skills guardrail |
|---|---|
| Measure area in EPSG:4326 because the operation returns a number. | Select and record an appropriate projected CRS; verify units before reporting area or distance. |
| Randomly split spatial samples and report a high validation score. | Audit spatial and group leakage; use blocked validation and show geographic error structure. |
| Count changed pixels from two convenient dates. | Check season, sensor, processing level, registration, mutual masks, threshold sensitivity, and error-adjusted area uncertainty. |
| Sum overlapping spatial intersections. | Dissolve or deduplicate overlap before measurement and preserve an auditable accounting path. |
| Render a map and assume it communicates honestly. | Check projection, classification, palette accessibility, legend semantics, uncertainty, and export metadata. |
| Run a destructive local GIS mutation immediately. | Inspect first, plan the mutation, require an explicit gate, verify outputs, and retain recovery evidence. |
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.
| Measure | Skills enabled | Control | Paired result |
|---|---|---|---|
| Criterion coverage | 145/302 (48.0%) | 85/302 (28.1%) | 65 wins / 5 losses / 232 ties; p = 2.2e-14 |
| Cases meeting every criterion | 18/92 | 6/92 | 42 wins / 3 losses / 47 ties; p = 8.7e-10 |
| Critical spatial failures | 4/92 (4.3%) | 13/92 (14.1%) | 3.25× fewer observed failures |
| Forbidden-behavior violations | 0 | 0 | no separation |
| Skill activation | 92/92 | 0/92 | clean control arm |
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 unp