SkillProof tested Claude skills

SQL Expert

ClickHouse SQL rules: leading-PK filters, qualified names, uniq over uniqExact

av FrankChen021 · FrankChen021/datastoria

Fungerar med konfiguration ★ 7.2/10

SQL Expert — ClickHouse SQL rules: leading-PK filters, qualified names, uniq over uniqExact

Vad den gör

A ClickHouse SQL authoring ruleset that constrains how queries are written: fully qualified table names, no trailing semicolons, mandatory LIMIT on exploration queries, bounded time windows, a filter on the leading primary-key column whenever a secondary column is filtered, uniq() instead of uniqExact(), and smaller table on the right side of joins. Triggers when the user asks to write, fix, or optimize a ClickHouse query, or to pull data from a ClickHouse table. Its schema-discovery and mandatory validate_sql/execute_sql workflow depend on tools supplied by the DataStoria web app, so in plain Claude Code only the SQL-writing rules apply.

Testrapport

Cloned the repo (GitHub API was rate-limited, so I used git), found SKILL.md at resources/skills/sql-expert/, and installed it into a throwaway HOME=$(mktemp -d) — frontmatter parsed as valid YAML with name=sql-expert and a 129-char description. Spot-checked raw fetches: SKILL.md 200 and clickhouse-system-queries/references/system-query-log.md 200, but sql-expert/references/system-query-log.md is 404, so the body's relative reference dangles when the skill is installed alone; no curl|sh, base64, or injection text anywhere in resources/skills/. Trigger judgments, all 5 correct: YES on "Write me a ClickHouse query for the top 10 slowest endpoints last week", "I need a query against our ClickHouse events table showing daily unique users", "Optimize this ClickHouse SQL, it's scanning the whole table"; NO on "Write a Postgres query with a lateral join for each customer's latest order" and "Analyze this CSV of sales figures and tell me which region grew fastest" — though the description's second sentence ("use this when the user needs data, queries, or analysis") has no ClickHouse qualifier and is a real over-trigger hazard. Output test on one MergeTree table (PARTITION BY toYYYYMM(event_date), ORDER BY (event_date, site_id, event_time)): baseline.sql produced `FROM page_views WHERE site_id = 42 AND event_time >= now() - INTERVAL 7 DAY ... uniqExact(user_id) ... LIMIT 10;` while skill.sql produced `FROM analytics.page_views WHERE event_date >= toDate(now() - INTERVAL 7 DAY) AND event_time >= ... AND site_id = 42 ... uniq(user_id) ... LIMIT 10` with no semicolon — four diff hunks, the load-bearing one being the added event_date predicate that the baseline omits, which is what lets ClickHouse prune partitions and use the sparse index when filtering the secondary PK column site_id. I could not verify that with EXPLAIN (no clickhouse, clickhouse-local, or docker on this machine), and I could not execute the skill's own CRITICAL RULE — validate_sql — because that tool ships with the DataStoria web app, hence verdict "setup" rather than "pass".

Testad: 2026-07-21 · Claude Code 2.x (agent harness)

Installation

git clone --depth 1 https://github.com/FrankChen021/datastoria.git /tmp/sql-expert-src
mkdir -p ~/.claude/skills
cp -R /tmp/sql-expert-src/resources/skills/sql-expert ~/.claude/skills/sql-expert
# The SKILL.md body calls tools provided by the DataStoria app, not by Claude Code:
#   validate_sql, execute_sql, get_tables, explore_schema, skill_resource, search_query_log
# Without a ClickHouse MCP server exposing equivalents, Claude cannot run the skill's
# "MANDATORY validation" step - only the syntax/optimization rules in sections 2 and 3 apply.
# The body's `references/system-query-log.md` lives in a sibling skill, not this dir. To get it:
#   cp -R /tmp/sql-expert-src/resources/skills/clickhouse-system-queries ~/.claude/skills/clickhouse-system-queries
# Sibling skills also referenced by the body: visualization, optimize-clickhouse-sql,
#   diagnose-clickhouse-errors, diagnose-clickhouse-clusters, source-code-inspection
# Full experience (with the tools wired up): use the app at https://datastoria.app

Kommandon och exempelprompter

  • /sql-expertClickHouse SQL rules: leading-PK filters, qualified names, uniq over uniqExact

Skills triggas av vanliga förfrågningar — inga kommandon att memorera. Efter installationen aktiverar prompter som dessa skillen (på engelska):

  • Write a ClickHouse query to aggregate daily sales
  • Optimize this slow ClickHouse SQL query
  • Validate this ClickHouse query for correctness

Läs hela rapporten →