Read-only DB Access

Some questions only rows can answer: did that trigger fire, what does this view actually return, is that column populated, how does the query behave at volume. This bundle makes asking them routine without making writing them possible.

What's inside

Path	Purpose
`skills/readonly-db/`	Setting up the read-only role and the `db:q` runner, and the five rules for handling what comes back.
`rules/database-safety.md`	Fires on `server/`, `db/`, `drizzle/` and `prisma/` — read-only exploration, no destructive SQL, no dumps, migrations from CI.
`CLAUDE.md`	A pointer block to paste into your project's `CLAUDE.md`.

The shape of it

Two independent guards, because either one alone is a single point of failure:

A dedicated role. <app>_claude_ro has SELECT and nothing else, with sensitive tables revoked explicitly and a statement_timeout so a bad join fails instead of hanging. Its credential lives in its own env var, separate from the app's.
A one-statement runner. {{pm}} run db:q -- "SELECT …" accepts only SELECT, WITH, EXPLAIN, SHOW, TABLE and VALUES, rejects a second statement, wraps the call in BEGIN READ ONLY, refuses to run outside development, and caps its own output.

The allowlist is there for the error message. BEGIN READ ONLY is what makes it true.

It is not a sandbox, and the skill says so plainly: any session in the repo can still reach the read-write credential the app uses. This removes a class of accident on the exploration path — it does not contain a session that goes looking.

Install

unzip readonly-db.zip
mkdir -p .claude/skills .claude/rules
cp -R readonly-db/skills/. .claude/skills/
cp -R readonly-db/rules/. .claude/rules/
cat readonly-db/CLAUDE.md >> CLAUDE.md

Then ask Claude to run the readonly-db skill: it writes the migration that creates the role and the db:q script, and wires the npm script. Put the role's credential in whatever secret manager the project already uses — never in a committed file.

Placeholders

{{pm}} in the skill and the rule is your package manager. The setup CLI renders it; by hand, replace it with pnpm, npm, yarn or bun.

What it deliberately does not do

No pg_dump/pg_restore — a dump is every revoke undone at once.
No production connection. Reproduce on a branch or a local copy.
Nothing read from the database is written into a file, a commit message or a PR body. Git history is permanent and gets read back into context on every later prompt.

Companion bundles

nuxt — its rules/database.md covers the Drizzle side (schema layout, generated migrations, query conventions) and agrees with this one on the safety rules.