J•JEV Field Guide
Home/Jev agent skill routing: retrieve, verify, or load nothing

After the basics · Worked guides

Jev agent skill routing: retrieve, verify, or load nothing

Distinguish creating a slide deck from editing an existing one using candidate retrieval and applicability checks.

Source checked 2026-10-0215 minutes
What you will build

Build a capability catalogue with a valid none outcome.

Before you start · Jev agent tool calls: selection, arguments, permissions, and fallback

Design it step by step

01

Define verbs and deliverables

Invented request: create a product deck from written notes. Creating and editing are different. Include goals, existing file types, outputs, and constraints; a PPT keyword is insufficient.

02

Make catalogue boundaries visible

Store IDs, descriptions, input/output types, prerequisites, and exclusions. App-controlled versions and paths matter. Truncating every description to “handle PPT” destroys the distinction.

03

Retrieve, then inspect fuller descriptions

The official approach ranks candidates and verifies using more detail. An illustrative three-candidate budget needs testing. Check creation support and prerequisites; rank is only reading priority.

04

Allow none and clarification

If all candidates require a file but the user has notes only, reject them or expand retrieval. Clarify ambiguous requests. A suggestion does not authorize installing a plugin.

05

Measure completed tasks too

Track recall, wrong loads, unnecessary loads, and task completion separately. Record selected versions and execution outcomes; test creation, editing, conversion, no-skill requests, and stale catalogues.

Worked example · Invented by this site

ObservationJudgmentApplication action
New deck from notesCreation capability fitsCheck full conditions
Existing deck, one text editEditing capability fitsAvoid rebuilding
No candidate fitsNo applicable skillReturn none or expand

A copyable design draft

Original examples. JSON illustrates request or input structure; Python calculates invented scores without an API call. Verify current interfaces and task policy before integrating.

# Invented candidate contract.
{
  "task": "create a new deck from notes",
  "available_inputs": ["plain_text"],
  "candidates": [
    {"id": "deck-author", "inputs": ["plain_text"], "creates_new": true},
    {"id": "deck-editor", "inputs": ["pptx"], "creates_new": false}
  ],
  "none_allowed": true
}
Design a decision

Common mistakes

  1. Selecting by extension only.
  2. Treating rank as confirmation.
  3. Always forcing a skill.

TRY / THINK / COMPARE

Think first, then compare

The top candidate needs an existing file, but the user has only notes. What next?

Show explanation

Verify creation support; reject an incompatible candidate and seek an authoring capability or return none.

Before handing it over

  • Catalogue retains prerequisites and exclusions.
  • IDs and versions traceable.
  • Verification can reject all.
  • Selection and completion evaluated separately.

Common questions

Is a recommendation a load command?

It is a suggestion subject to availability and permission checks.

Are more candidates always better?

No. Validate the recall–cost tradeoff.

Sources and further reading

Inspired by official patterns and Datawhale practice topics. Explanations, examples, and exercises are independently written. These are teaching designs, not live API runs or benchmarks.