Ревью SQL-запросов

Проверяет SQL и ORM-код на корректность (NULL, дубли из join, часовые пояса) и скорость (EXPLAIN, индексы, SELECT *, N+1, keyset-пагинация) и на параметризацию запросов. Используйте при написании, отладке или оптимизации SQL.

Базы данных и SQLПак: Базы данных94 строкиsql-query-review

Установка

Выполните одну команду — она сохранит SKILL.md в папку скиллов агента. Перезапустите агента, и он подхватит скилл сам.

Claude Code — macOS, Linux
mkdir -p ~/.claude/skills/sql-query-review && curl -fsSL https://anymodel.org/en/skills/databases/sql-query-review/SKILL.md -o ~/.claude/skills/sql-query-review/SKILL.md
Codex, Cursor — macOS, Linux
mkdir -p ~/.agents/skills/sql-query-review && curl -fsSL https://anymodel.org/en/skills/databases/sql-query-review/SKILL.md -o ~/.agents/skills/sql-query-review/SKILL.md
Claude Code — Windows PowerShell
New-Item -ItemType Directory -Force "$HOME\.claude\skills\sql-query-review" | Out-Null; Invoke-WebRequest https://anymodel.org/en/skills/databases/sql-query-review/SKILL.md -UseBasicParsing -OutFile "$HOME\.claude\skills\sql-query-review\SKILL.md"
Codex, Cursor — Windows PowerShell
New-Item -ItemType Directory -Force "$HOME\.agents\skills\sql-query-review" | Out-Null; Invoke-WebRequest https://anymodel.org/en/skills/databases/sql-query-review/SKILL.md -UseBasicParsing -OutFile "$HOME\.agents\skills\sql-query-review\SKILL.md"

Другой агент: сохраните файл в его папку скиллов или дайте ему ссылку на файл: https://anymodel.org/en/skills/databases/sql-query-review/SKILL.md

SKILL.md

Исходник

SQL Query Review

Review a query for wrong results first, slow results second, and injection risk always. Identify the database engine and version (Postgres, MySQL, SQLite, SQL Server) before advising; syntax and planner behavior differ. Read the schema (tables, keys, indexes) before judging a query.

Workflow

  1. Get the query, the schema (\d table in psql, SHOW CREATE TABLE, or migration files) and the approximate table sizes.
  2. Check correctness using the list below. Wrong results outrank slow results.
  3. Check safety (parameters, permissions).
  4. Check performance with the real plan, not guesses.
  5. Propose the rewritten query and any index; verify results are identical on sample data.

Correctness checklist

NULL semantics

  • NULL = NULL and x <> 5 are not true when x is NULL. Use IS NULL, IS DISTINCT FROM (Postgres) or <=> (MySQL).
  • NOT IN (subquery) returns nothing if the subquery yields any NULL; use NOT EXISTS.
  • COUNT(col) skips NULLs, COUNT(*) does not. SUM of no rows is NULL; wrap with COALESCE when a number is expected.
  • A filter on the right table in a WHERE clause after a LEFT JOIN turns it into an inner join; move the condition into ON — except WHERE right.id IS NULL, which is a deliberate anti-join and must stay in WHERE.

Joins that duplicate rows

  • Joining a one-to-many table multiplies rows, then SUM/COUNT overcounts. Aggregate in a subquery or CTE first, then join, or use COUNT(DISTINCT ...) knowingly.
  • Confirm the join keys are unique on the side you assume is "one"; check with GROUP BY key HAVING COUNT(*) > 1.
  • Missing join condition produces a cartesian product.

Grouping and ordering

  • Every non-aggregated selected column must be in GROUP BY (or functionally dependent). Do not rely on MySQL's lax mode.
  • LIMIT without a total ORDER BY returns an arbitrary set; add a unique tiebreaker (ORDER BY created_at, id).

Time and types

  • Store and compare timestamps in UTC; use timestamptz in Postgres. State the session time zone before using DATE(ts) or date_trunc.
  • Range filters should be half-open: ts >= '2026-01-01' AND ts < '2026-02-01', not BETWEEN with 23:59:59.
  • Avoid implicit casts and mixing types in comparisons; they change results and disable indexes.
  • Money in integers or numeric, not floats. Case sensitivity and collation affect equality and uniqueness.

Safety

  • All user-provided values go through bind parameters ($1, %s with a driver, ?), never string concatenation or f-strings. For dynamic identifiers (column or table names) use an allow-list.
  • LIKE patterns from users: escape % and _.
  • Application roles should be least-privilege; flag use of a superuser in application code.
  • Do not run UPDATE/DELETE/DDL on a real database without confirmation. For write queries run a SELECT with the same WHERE first and wrap the change in a transaction the user can roll back.

Performance

Read the plan

EXPLAIN SELECT ...;                      -- estimates only, safe
EXPLAIN (ANALYZE, BUFFERS) SELECT ...;   -- Postgres: actually runs the query

EXPLAIN ANALYZE executes the statement. Never use it on INSERT/UPDATE/DELETE against real data unless inside a transaction you roll back, and ask first.

Look for: sequential scans on large tables with selective filters, row estimates far from actual rows (stale statistics: ANALYZE), sorts spilling to disk, nested loops over big sets, repeated subplans.

Indexes

  • Index columns used in WHERE, JOIN and ORDER BY; order composite columns by equality conditions first, then range, then sort.
  • Functions on columns (WHERE lower(email) = ..., WHERE date(ts) = ...) defeat plain indexes; rewrite as a range or add an expression index.
  • A leading wildcard LIKE '%x' cannot use a btree; consider trigram or full-text indexes.
  • Each index slows writes and costs space; do not add one without evidence from the plan. Creating indexes on large live tables needs the safe procedure (see the safe-db-migration skill if available).

Patterns to fix

  • SELECT *: list needed columns; it blocks index-only scans and breaks when columns are added.
  • N+1 from ORMs: one query per row in a loop. Use eager loading (select_related, prefetch_related, includes, joinedload) or a single WHERE id IN (...) query. Detect by enabling query logging for one request.
  • OFFSET pagination slows linearly. Use keyset pagination:
SELECT id, created_at, title FROM posts
WHERE (created_at, id) < ($1, $2)
ORDER BY created_at DESC, id DESC
LIMIT 50;
  • COUNT(*) on huge tables for UI totals: consider approximate counts or caching.
  • DISTINCT used to hide duplicate-producing joins: fix the join.

Output format

  1. Verdict: correct / has correctness bugs / slow, with one sentence of why.
  2. Findings: severity | location | problem | fix.
  3. The rewritten query, and any index DDL as a proposal (not executed).
  4. Plan evidence: before/after key numbers (rows, timing, scan type) if you could measure; otherwise state it was not measured.
  5. Assumptions made about schema, data size and engine version.

Pitfalls

  • Do not tune on an empty dev database; plans differ with data volume.
  • Do not trust a cached warm run for timing; compare repeated runs.
  • Do not rewrite a query into something that changes results without proving equality on sample data.
  • Engine-specific advice must name the engine.

Source: AnyModel skills catalog · License: CC BY 4.0

Подключить через AnyModel API

Скилл работает на той модели, которую использует агент. Подключите Claude Code, Codex или Cursor к AnyModel и запускайте его на Claude, GPT, Gemini и других моделях по одному ключу — оплата за токены, без подписки.

Ещё из пака Базы данных