supabase（Supabase 后端开发）

Supabase Database Operations

Execute database operations on Supabase: queries, writes, migrations, and diagnostics.

MCP is optional. This skill works with MCP (auto), Supabase CLI, psql, or Dashboard. See BACKENDS.md for execution options.

Scope

Applies to:

• Database actions on Supabase: query/statistics/export, write (after confirmation), migration (DDL), type generation, query logs/advisors

Does not apply to:

• Project-side integration (env/client code/data access layer) → Use workflow-ship-faster (Step 6) for project setup; this skill handles DB-side actions only

Called by:

• workflow-ship-faster uses this skill as DB operation foundation

Postgres Best Practices (Built-in)

Ship Faster vendors Supabase's Postgres best practices under references/postgres-best-practices/.

Consult it when:

• Writing/reviewing/optimizing SQL queries
• Designing indexes, schema changes, or RLS policies
• Diagnosing performance, locking, or connection issues

Source of truth:

• Full guide: references/postgres-best-practices/AGENTS.md
• Individual rules: references/postgres-best-practices/rules/*.md

When proposing changes, cite the relevant rule file path (for example: references/postgres-best-practices/rules/query-missing-indexes.md) and keep changes minimal.

Security Rules (Must Follow)

1. Read first: Always check schema before any operation
2. Default LIMIT 50: All SELECT queries default to LIMIT 50, unless user explicitly requests more
3. Write operation confirmation: INSERT/UPDATE/DELETE must before execution:

- Display the SQL to be executed - State expected number of affected rows - Await explicit user confirmation

1. No bare writes: UPDATE/DELETE without WHERE condition → refuse directly, do not execute
2. Batch threshold: Affecting > 100 rows → force double confirmation + suggest SELECT count(*) first
3. DDL via migration: Schema changes must use migrations, not direct DDL
4. Production environment: Write disabled by default; only allow when user explicitly says "execute on prod" and double confirms
5. Sensitive fields: email/phone/token/password are masked or not returned by default, unless user explicitly requests

Operation Flow

1. Parse requirements → restate objective
2. Unsure about tables/fields → first query schema (information_schema or list_tables)
3. Plan SQL → present to user
4. Read-only → execute directly
5. Write operation → confirm before execution → verify affected rows → report result

File-based Pipeline

When integrating into multi-step workflows, persist artifacts to disk:

runs/<workflow>/active/<run_id>/
├── proposal.md                # Requirements / objective
├── context.json               # Known tables/fields/IDs
├── tasks.md                   # Checklist + approval gate
├── evidence/sql.md            # SQL to execute (write ops written here first)
├── evidence/result.md         # Conclusion + SQL + results
└── logs/events.jsonl          # Optional tool call summary (no sensitive data)

Output Format

• Language: English
• Structure: Conclusion → Key numbers → Executed SQL → Result table (max 50 rows)
• Overflow handling: Truncate + show total count + optional export/pagination

Example:

✅ Query complete: 142 new users in the last 7 days

Executed SQL:
SELECT DATE(created_at) as date, COUNT(*) as count
FROM user_profiles
WHERE created_at > NOW() - INTERVAL '7 days'
GROUP BY DATE(created_at) ORDER BY date DESC;

| date       | count |
|------------|-------|
| 2025-01-09 | 23    |
| 2025-01-08 | 31    |
| ...        | ...   |

Error Handling

Example Workflows

Read: Simple Query

User: Get registered user count for the last 7 days, by day

1. Confirm table user_profiles, field created_at
2. Execute aggregation SQL
3. Return: conclusion + numbers + SQL + table

Read: Complex Query

User: Find projects that have runs but all failed

1. Confirm projects, runs tables and status field
2. Present JOIN + aggregation SQL
3. Execute and return results (mask email)

Write: Insert

User: Create a new run for project xxx

1. First check if project exists
2. Present INSERT SQL + expected impact: 1 row
3. Await confirmation → execute → return new record id

Write: Update

User: Change run abc's status to completed

1. First SELECT to verify current state
2. Present UPDATE SQL + WHERE id = 'abc'
3. Confirm → execute → SELECT again to verify

Dangerous: Delete

User: Delete all runs where status = 'failed'

1. First SELECT count(*) WHERE status = 'failed'
2. Present count + DELETE SQL
3. If > 100 rows, force double confirmation
4. After confirmation execute → report deleted row count

Dangerous: DELETE without WHERE

User: Clear the runs table

❌ Refuse to execute
→ Prompt: DELETE without WHERE condition, this will delete all data
→ Suggest: Use TRUNCATE (requires migration) or add explicit condition

Schema Discovery

Get latest schema at runtime:

-- List all tables
SELECT table_name FROM information_schema.tables
WHERE table_schema = 'public';

-- View table structure
SELECT column_name, data_type, is_nullable
FROM information_schema.columns
WHERE table_name = '<table_name>';

For project-specific schema (may be outdated), see schema.md.

Related Files

• BACKENDS.md — Execution options (MCP/CLI/Console)
• SETUP.md — MCP configuration (optional)
• schema.md — Project-specific schema reference