Route, Don’t Flood: Using db/SKILL.md to Steer Oracle‑Aware Assistants

If you’ve tried to make an assistant “Oracle‑aware,” you’ve likely hit the same wall: you paste a stack of links, the model blends vendor‑neutral habits with stale Oracle guidance, and the answers get vague when you need precision. The fix isn’t more context—it’s better routing. One Oracle skill at a time, in the right order, so every next move follows Oracle’s own sequence.

In Article 1 we outlined an operating model for AI on Oracle Database: route → act → trust. This piece goes deep on the first verb. It shows how to use db/SKILL.md in the public Oracle Database Skills repository as your front door; how to route by persona or by task; and how to enforce progressive discovery so your assistant loads exactly one file at a time. This is a no‑execution article; we stay in the routing lane and leave tools to Article 3 (SQLcl MCP). Prerequisite: you only need to browse

Slow‑query diagnosis

Begin by reading the plan that actually ran. In Oracle that means capturing the executed plan from the cursor (for example, DBMS_XPLAN.DISPLAY_CURSOR(NULL, NULL, 'ALLSTATS LAST')), not relying solely on an EXPLAIN PLAN estimate. If the plan looks sane but runtime stalls, pivot to wait events. If rows and joins are misestimated, fix statistics and cardinality before widening the window with AWR.

• Route (skills under db/performance): explain‑plan → wait‑events → optimizer‑stats → awr‑reports


• Availability note: AWR/ASH access depends on environment and licensing.

Agent‑safe schema change

Discover what exists before you propose change, add loss‑prevention so a single bad predicate can’t harm a table, and make each step idempotent so a retry won’t double‑apply. Only then move into an audited migration workflow.

• Route (skills under db/agent and db/migrations): schema‑discovery → destructive‑op‑guards → idempotency‑patterns → schema‑migrations

A compact prompt that enforces progressive discovery

A small prompt can prevent freelancing by enforcing a loop: route, read one file, summarize, decide. That discipline keeps the context precise and affordable.

Prompt

Open db/SKILL.md. I’m [persona or task].

Which single skill file should I load next, and why?

Summarize that file in 5 bullets, then propose the next decision (read another file or stop).

Do not execute any actions.

Progressive routing loop