Skills
Data dictionary skill
Your tables, columns, and traps: agents query with the map instead of guessing.
Last reviewed July 17, 2026
A single-file skill template that gives agents your data dictionary: table grains, column meanings, null semantics, and the known query traps.
SKILL.md
# Data dictionary Apply when writing SQL or analyzing data in this workspace. ## Tables - orders: one row per order. status values: pending, completed, refunded, cancelled. Revenue queries filter status = completed ONLY. - customers: one row per account. country is nullable for pre-2023 signups (unknown, not domestic). - events: one row per tracked event; deduplicate on event_id, the pipeline can double-deliver. ## Conventions - All timestamps UTC; report in the requester's timezone and say which. - Money columns are integer cents; divide by 100 only at display time. - Soft deletes: rows with deleted_at set are excluded from every metric unless the request names them. ## Known traps - Joining orders to events fans out; aggregate events first. - Test accounts: exclude customer ids listed in the internal-accounts note. When a request needs a table or column not listed here, say so and stop; never invent columns.
Set it up
- 1
Copy the skill body
Replace the sample tables with your real grains, enums, and traps; the traps section pays the rent.
- 2
Create the skill
Add it in the workspace Skills page; every SQL-writing chat now starts with the map loaded.
- 3
Maintain it on misses
When an agent writes wrong SQL for a documented reason, the fix goes in this file, not in the chat.