| name | data-analysis |
| description | Methodology recipes for multi-step data questions over the bundled DB. Load when the user asks for a complex breakdown, comparison, or anything that needs more than one SQL pass. |
Data analysis recipes
The core rules (delegating to sql-agent, parsing the artifact-id contract,
artifact visibility) are already in your main instructions — this skill
only contains extra patterns for non-trivial questions.
Picking a dimension when the user is vague
If the user says "show me trends" or "what's interesting in the data",
pick a dimension yourself rather than asking back. Default options for the
Chinook DB:
- time — invoice date by month/quarter; good for trend questions
- geography — billing country; good when user mentions regions
- genre / artist — when user mentions music
- customer — when user mentions revenue / spend
Pick the dimension that gives the cleanest story. State the dimension you
picked in the first line of your reply.
Multi-step delegation pattern
For questions that need two or more SQL passes (e.g. "top 10 X then break
each by Y"):
- First
task(subagent_type="sql-agent", description=...) for the
filter/seed list. Parse artifact_id=art_a; columns=....
- Second
task(subagent_type="sql-agent", description=...) referencing
the IDs from step 1 explicitly in the description so the sub-agent can
embed them in a WHERE … IN (...) clause.
- Process / chart in
python_exec, joining the two artifacts on the
shared id column.
Never try to do both passes in one SQL call — the sub-agent is allowed
LIMIT 200 and the join can blow past it.
Validating before you present
Quick sanity checks after read_artifact:
- Row count matches the LIMIT you asked for, or the question intent.
- Numeric totals are plausible (no all-zero, no unexpected negatives).
- No mostly-null columns.
If something looks off, refine via another sql-agent call rather than
papering over it in pandas.
