Service Desk

Task 1 asks an agent to resolve a support request with account, requester, entitlement, owner, and history data. Its collocated Task.md keeps the exact request, focal records, accepted result, and Verifier criteria.

The audit proves reusable project facts:

• the Harness uses search_cases, get_case, and update_case;
• case visibility depends on team membership;
• entitlement and account state can change the valid action;
• updates write an event and change the case row in one transaction;
• reset rebuilds SQLite from a fixed structured fixture; and
• the Verifier can compare raw case and event state without exposing it.

Create .agents/skills/service-desk-world/SKILL.md. Add only these proven facts, their source paths, and routes to any reusable script.

After Task 2

Task 2 uses the same product but tests a different capability: it requires the agent to connect a stale knowledge article, a recent incident, and a customer message before making an update.

The second audit finds more reusable knowledge:

• documents need stable structured facts before prose generation;
• article freshness is decided by effective_at, not insertion order;
• incident-product and case-product links must stay complete;
• generated messages need fact-ID review for unsupported claims; and
• known-good, valid-alternative, wrong, shortcut, collateral, and corrupt- evidence fixtures expose different Verifier faults.

Add a short reference for the data relationships and a reusable fixture validator only because two distinct Tasks now use them. Update SKILL.md to route agents to those items.

Resulting project skill

Its root skill states when the knowledge applies and gives exact project paths, commands, state rules, and known limits. It can contain project-specific Task, Environment, data, Verifier, Harbor, and calibration guidance. It does not say only “follow eval-engineering,” and it does not copy broad guidance without a project-specific adaptation.

The skill does not include either Task’s request, selected records, expected answer, or exact rubric. Those stay with each Task. It also does not add empty folders, a catalog, a manifest, or a file per concept. New content must reduce rediscovery or prevent a demonstrated defect in another Task.

Evidence of value

Task 3 tests whether the World Skill works. A new author should be able to find the correct tool contract, create relationship-complete data, reset state, and reuse the raw-state Verifier boundary without studying Tasks 1 and 2 in full. If a rule fails on Task 3, narrow or correct it and cite the new evidence.