How Hermes routes a spatial question through skills and verified tools

A spatial question often looks simple while hiding several independent operations. The GIS skill library helps Hermes choose and validate that workflow; the actual execution depends on the tools registered in the active profile.

This article reflects Hermes GIS Skills v0.1.0. Tool names and schemas vary by installation.

1. A skill sets the workflow

For a question such as “Which zoning district covers this approved address?”, a relevant skill should require:

• an authoritative zoning source;
• a geocoding step or approved input geometry;
• the expected spatial relationship and spatial reference;
• bounded output fields and counts;
• source, timestamp, and caveat reporting;
• no write or publication unless separately approved.

The skill does not grant access or register an Esri tool. It contributes process and quality checks to the model context.

2. Hermes inspects available tool schemas

The active profile may expose public web access, an ArcGIS REST integration, a geocoder, a local Python script, or another approved tool. Hermes must use the names and argument schemas actually registered for that session.

request
  → load relevant GIS skill
  → inspect available tool schemas
  → choose public/read-only execution where possible

Do not publish a made-up helper such as intersects('1234 J Street') as ArcGIS SQL. An address is not geometry, and ArcGIS where clauses do not geocode it.

3. Geocode, then query geometry

A defensible sequence is:

1. Geocode the approved address and retain the geocoder source and match quality.
2. Confirm the point geometry and spatial reference.
3. Query the authoritative zoning layer with that geometry and an ArcGIS spatial relationship supported by the service.
4. Request only the district fields needed for the answer.
5. Validate result count, layer authority, and any boundary ambiguity.

If the configured integration cannot perform these steps, Hermes should return the plan and commands rather than claiming the lookup ran.

4. Return evidence, not just prose

A useful final answer includes:

Input
- approved address or geometry
- geocoder and match quality

Source
- zoning service and layer URLs
- owner or authority evidence
- inspection timestamp

Query
- geometry and spatial reference
- spatial relationship
- requested fields and result count

Result
- district code and label
- boundary or overlap caveats
- no write requested or performed

5. Keep corrections in the right layer

A reusable correction such as “always validate paging before full extraction” may belong in a skill. A one-off address, private layer ID, client result, or credential does not. Keep project evidence in sessions or notes and sanitize any improvement before contributing it to the public repository.

The advantage of the skill-first approach is not a magic GIS toolset. It is an inspectable workflow that keeps source validation, execution boundaries, and reporting requirements visible from the start.