muend.
Agent skills · GeoAI knowledge infrastructure

geoai-skills

Stop silent CRS, spatial-leakage, validity, unit and uncertainty failures before they ship.

Activev0.4.1MIT2 Awesome lists
Versionv0.4.1
Skills18
Evaluation167 routing cases / 92 behavior pairs
DistributionCodex / Claude / Copilot / skills.sh
ListingsAwesome Claude Code / Awesome Geospatial
LicenseMIT
Method

Methodological checks before a geographic claim

Eighteen independently installable skills cover the geospatial lifecycle: data acquisition, remote sensing, spatial statistics, ML validation, PostGIS, cartography and guarded ArcGIS automation. They require coordinate systems, units, validity, comparability, leakage controls, provenance and uncertainty to be made explicit.

v0.4.1 packages those skills for Codex, Claude Code, GitHub Copilot and compatible Agent Skills runtimes. Runtime archives stay separate from evaluation cases and private development material.

Community listings

Included in two curated Awesome lists

Awesome Claude Code lists the project under Research & Scientific Inquiry. Awesome Geospatial includes it under Agent Skills. Both entries were checked directly in the maintainers’ source lists on 2 November 2026. These are editorial listings, not benchmark rankings.

Specialists18

Independently installable, vendor-neutral skill packages.

Routing cases167

99.17% precision; one frozen runtime/model experiment.

Behavior pairs92

302 pinned criteria; evaluated separately from routing.

Curated lists2

Awesome Claude Code and Awesome Geospatial.

Evaluation

Routing and behavior are separate evidence

The routing run records 99.17% precision, 96.77% recall and 96.41% full-route accuracy on Claude Code 2.1.214 with claude-sonnet-5. The later behavior experiment uses 92 clean matched pairs and a model judge: criterion coverage changes from 28.1% to 48.0%, while critical spatial failures change from 14.1% to 4.3%.

The behavior project gate remains unmet: all-criteria success is 19.6% against an 85% requirement, and critical failures are 4.3% against a below-2% requirement. A 47-case blinded human-review sample is prepared but not adjudicated. The published behavior results are model-judged, uncalibrated and not human-verified.

Evidence boundary

The routing and behavior packages use different runtime versions and suite hashes. Neither supports a universal model-independent quality claim. Package and source links above expose the evidence and its limitations.

Open to collaboration

Build AI systems that hold up.

© 2026 Muhammed Enes Duran Static site · no trackers
Technical diagram