WebGIS Claude Skills is a repository-backed engineering skill system for building reliable GIS software with AI coding agents.

It is designed to help an agent move from:

Requirement → Architecture → Design → Data Contracts → Implementation → Validation → Security → Performance → Documentation → Persistent Memory

The goal is not simply to generate GIS code. The goal is to make the agent reason about spatial correctness, scientific validity, scalability, security, cost, reproducibility, production operations, and validated learning.

GIS applications fail in ways that ordinary software assistants often miss:

• incorrect CRS or measurement assumptions
• invalid geometry and raster misalignment
• spatial/temporal leakage in GeoAI
• huge GeoJSON payloads and browser overload
• unbounded spatial APIs
• insecure GeoServer/OGC exposure
• provider-secret leakage
• expensive cloud choices without requirements
• scientifically invalid methodology chosen because it improves a metric

This repository turns those concerns into skills, contracts, anti-patterns, decision matrices, fixtures, executable evaluations, persistent project memory, and a controlled evidence-driven learning pipeline.

The visual is intentionally part of the repository homepage so the project communicates its architecture before a visitor reads the details.

                        USER REQUIREMENT
                               │
                               ▼
                    ┌────────────────────┐
                    │ PROJECT ORCHESTRATOR│
                    └──────────┬─────────┘
                               │
              ┌────────────────┼────────────────┐
              ▼                ▼                ▼
        ARCHITECTURE       DESIGN SYSTEM    DATA CONTRACTS
              │                │                │
              └────────────────┼────────────────┘
                               ▼
                        IMPLEMENTATION
                               │
                               ▼
                    ┌────────────────────┐
                    │ QUALITY EVALUATION │
                    ├────────────────────┤
                    │ GIS Correctness    │
                    │ Scientific Integrity│
                    │ API / Security     │
                    │ Scalability        │
                    │ Performance        │
                    │ Regression         │
                    └──────────┬─────────┘
                               ▼
                         VERIFIED STATE
                               │
                               ▼
                        PROJECT MEMORY

A core rule of the system is:

The system must never modify scientific methodology merely to artificially improve model metrics.

A higher F1, IoU, accuracy, R², or other score does not automatically mean a better model when the evaluation protocol changed or became less valid.

The system instead records:

• methodology
• scientific rationale
• evaluation protocol
• dataset/model versions
• before/after metrics
• validation evidence
• experiment lineage

Ambiguous methodology changes are escalated rather than silently accepted.

The repository can evolve from experience across many projects without treating GitHub as a raw memory database.

The controlled model is:

Project memory → Sanitized observation → Learning candidate → Deduplication / aggregation → Privacy + generalization checks → Validation / regression → Versioned repository knowledge

Project memory stays project-scoped by default. Raw conversations, private source code, proprietary datasets, secrets, customer data, and sensitive locations are not repository learning inputs.

A new observation must not silently overwrite an existing skill. If it belongs to an existing capability, the skill evolves through a versioned, validated change. A separate skill is created only when the responsibility is genuinely distinct.

See learning/README.md, learning/schema.md, learning/PRIVACY.md, and learning/contribution-protocol.md.\n\nThe contribution boundary is local-only by default. Explicit opt-in is required before sanitized evidence may leave a project boundary. Contribution manifests are deterministically validated in CI; a valid manifest is still not accepted repository knowledge.

The system explicitly separates:

Native CRS ≠ Display CRS ≠ Analysis CRS

This allows datasets with different native coordinate reference systems to coexist in the same map without mutating their source data.

Permanent reprojection is treated as a deliberate data transformation, not a side effect of visualization.

The repository includes executable evaluation building blocks for:

CRS
 ↓
Geometry
 ↓
Raster
 ↓
GeoAI
 ↓
Scientific Integrity
 ↓
Web GIS Scalability
 ↓
Spatial API
 ↓
Security
 ↓
Performance
 ↓
Learning Safety
 ↓
Regression

Evaluation results distinguish:

• PASS
• FAIL
• BLOCKED
• ESCALATE

A plausible patch is never treated as verified without the required validation.

The system is designed to do more than document bad patterns.

Anti-pattern
     ↓
Detection
     ↓
Classification
     ↓
Explain risk
     ↓
Select correct pattern
     ↓
Auto-fix when safe
     ↓
Validate
     ↓
Regression protection

Examples include:

• giant GeoJSON
• client-side million-feature overload
• unbounded spatial queries
• degree-based planar distance
• mixed-CRS analysis
• raster grid misalignment
• spatial ML leakage
• temporal leakage
• provider secrets in the browser
• unrestricted WFS
• raster full-download workflows
• map reinitialization
• N+1 spatial requests

webgis-claude-skills/
├── skills/
│   ├── core/
│   ├── webgis/
│   ├── spatial/
│   ├── remote-sensing/
│   ├── geoai/
│   ├── 3d/
│   ├── desktop/
│   ├── realtime/
│   ├── infrastructure/
│   ├── quality/
│   └── scientific-integrity/
├── profiles/
├── decision-matrices/
├── contracts/
├── anti-patterns/
├── fixtures/
├── evals/
├── learning/
├── templates/
└── project-memory/

The repository currently keeps proven existing paths stable while the broader target architecture is established through compatible additions rather than a risky bulk reorganization.

The orchestrator can classify work as:

• Web GIS application
• GIS portal
• GeoAI platform
• Remote-sensing platform
• GIS dashboard
• Desktop GIS
• GIS API/service
• GIS data pipeline
• 3D GIS
• Real-time GIS
• Enterprise GIS
• Scientific/research GIS
• Commercial GIS SaaS

Hybrid projects are supported.

The system does not hard-code a single framework.

Technology choices are requirements-driven and can evaluate, where relevant:

Frontend: React/Vite, Next.js, Vue/Nuxt, Angular, Svelte/SvelteKit

Maps: MapLibre GL JS, OpenLayers, Leaflet, Cesium

Backend: FastAPI/Python, Node/TypeScript, .NET, Java/Spring, Go

Spatial database: PostGIS, SQLite/GeoPackage

Raster: COG/GDAL, object storage, STAC, raster services

Point cloud: LAS/LAZ, COPC, PDAL, 3D Tiles

Deployment: VPS, containers, managed infrastructure, Kubernetes, on-premises, hybrid

The decision engine records constraints, alternatives, evidence, cost/licensing implications, and exit considerations.

The default policy is:

Use the best viable free/open-source option first.

Paid services are considered when a concrete requirement justifies them, such as scale, SLA, proprietary data, production GPU throughput, or managed reliability.

This repository is licensed under Apache License 2.0.

Third-party components, if incorporated later, should retain their original license and attribution requirements.

The repository can be installed into an existing code project without requiring Python for the skill distribution layer.

For a normal persistent installation, use npm:

npm install -g webgis-claude-skills
webgis-claude-skills install --agent all

Update later with:

npm update -g webgis-claude-skills
webgis-claude-skills update --agent all

Once the npm package is published:

npx webgis-claude-skills install --agent all

Until the public npm package is published, the same CLI can be run directly from this GitHub repository:

npx github:Muhammad-Tariq/webgis-claude-skills install --agent all

Install for one client:

npx webgis-claude-skills install --agent claude
npx webgis-claude-skills install --agent codex
npx webgis-claude-skills install --agent cursor
npx webgis-claude-skills install --agent opencode

Install the complete repository support bundle:

npx webgis-claude-skills install --full

The npm installer has two modes. The default is lightweight and does not install Python GIS libraries:

npm install -g webgis-claude-skills
webgis-claude-skills install --agent all

For GIS development, opt into a managed Python virtual environment:

webgis-claude-skills install --agent all --with-gis

This creates an isolated .webgis-claude-skills/venv/ and installs:

NumPy · Pandas · Shapely · PyProj · GeoPandas · Rasterio

The managed environment keeps the user's existing Python projects untouched. Advanced users can explicitly use their existing system Python instead:

webgis-claude-skills install --agent all --with-gis --python-system

The same options work through npx:

npx webgis-claude-skills install --agent all --with-gis

Verify the installation:

npx webgis-claude-skills verify
npx webgis-claude-skills doctor

Update later:

npx webgis-claude-skills update --agent all

The installer supports the portable .agents/skills/ layout plus native/compatible locations for Claude Code, Cursor, and OpenCode. Codex uses the .agents/skills/ project location. The compatibility registry is maintained in integrations/agent-support.json.

Python is intentionally separate from the npm skill distribution. Use Python when the project or contribution needs GIS/scientific processing, evaluation, GDAL, or remote-sensing tooling.

The standard GIS extra provides NumPy, Pandas, Shapely, PyProj, GeoPandas, and Rasterio.

Python remains fully supported. Once the PyPI package is published:

python -m pip install webgis-claude-skills
webgis-claude-skills install --agent all

For repository development:

python -m pip install -e .

Optional GIS/GDAL/remote-sensing environments:

python -m pip install "webgis-claude-skills[gis]"
python -m pip install "webgis-claude-skills[gdal]"
python -m pip install "webgis-claude-skills[remote-sensing]"

The npm layer does not require Python. Python is available when the project needs a Python runtime, the repository evaluation harness, GDAL, GeoPandas, Rasterio, Shapely, PyProj, Earth Engine, or other Python tooling.

See Installation for the complete distribution architecture.

The repository is intended for coding-agent workflows that can read repository files and skills.

A typical session follows:

1. Load project memory
2. Classify the project
3. Select relevant skills
4. Resolve architecture / technology gates
5. Inspect the existing implementation
6. Implement the smallest coherent change
7. Run domain + correctness + security + performance checks
8. Run relevant evaluation cases
9. Record the verified state
10. Optionally derive sanitized learning candidates
11. Commit and continue from the exact next action

• CRS metadata/invariant checks
• Geometry fixture checks
• Raster metadata/alignment checks
• GeoAI experiment/leakage checks
• Web GIS scalability checks
• Spatial API/security checks
• Performance budget/regression checks

Learning safety currently has a deterministic contract case covering unvalidated automatic promotion. Runtime aggregation/telemetry is intentionally not enabled by this repository.

The harness deliberately refuses to claim numerical correctness when a real computational engine is required but not executed.

• Repository-backed project memory
• Project orchestrator
• Software engineering foundation
• GIS correctness gate
• Technology decision engine
• Project profiles
• Data contracts
• Golden GIS fixtures
• Evidence-driven learning boundary

• Web GIS architecture/frontend/map engine/UX
• PostGIS
• GeoServer
• Spatial APIs
• Remote sensing / GEE
• GeoAI
• Raster / LiDAR / 3D / OGC
• Desktop / real-time
• DevOps / observability / performance

• Anti-pattern remediation
• Executable evaluation harness
• Scientific integrity gate
• Scalability evaluation
• API/security evaluation
• Performance regression checks
• Learning safety case
• Unified evaluation command with pluggable runtime adapters
• Learning candidate validator
• Optional privacy-safe contribution protocol
• CI regression execution for deterministic cases
• Repository health dashboard
• Public skill/version manifest
• Release automation

This project is open to community contributions.

Start here:

• Contribution guide — workflow, validation, privacy, and licensing
• Community guide — contributor onboarding and recognition
• Bug report
• Feature request
• Security policy — private vulnerability reporting

The repository accepts durable contributions such as skills, project profiles, decision matrices, contracts, evaluation cases, fixtures, security/performance checks, documentation, and privacy-safe learning artifacts.

Community contributions follow the same engineering quality gates as the repository itself: provenance, deterministic validation where practical, CI, licensing, privacy, security, and scientific-integrity requirements.

See AUTHORS.md for project authorship and attribution.

Contributions should add durable engineering value.

When adding a skill, contract, anti-pattern, fixture, evaluation, or learning candidate:

1. define the problem and scope
2. add deterministic validation where possible
3. document assumptions and failure modes
4. preserve licensing and provenance
5. avoid secrets and private project data
6. keep learning candidates non-authoritative until validated
7. update project memory
8. verify before claiming completion

Apache License 2.0. See LICENSE.