Distillery

Stop burning 25k tokens every time Claude needs to check a table name.

The Problem

You ask Claude to write a query. Claude calls list_tables on Supabase MCP. Your context window screams. 25,000 tokens gone. For a table name.

The Fix

Distill your schema once into a ~300 line grep-friendly file. Claude searches that instead.

Before: "What tables handle risks?" → 25k token MCP call
After:  "What tables handle risks?" → grep, 0 tokens, instant

Install

/plugin marketplace add Snack-JPG/Distillery
/plugin install distillery@Snack-JPG-Distillery

Restart Claude Code.

Use

Generate your schema index:

/distillery:distill

This creates schema-index.md in your project and updates CLAUDE.md so future sessions know about it.

Then just ask naturally:

• "What tables handle users?"
• "What columns are in project_risks?"
• "Find tables with organization_id"

The skill intercepts MCP calls and searches your index instead.

What You Get

A file like this:

# Schema Index

## Users & Auth
users | Core accounts | id, email, role, organization_id
user_sessions | Active sessions | id, user_id, token_hash, expires_at

## Projects
projects | Main entity | id, name, status, owner_id
project_risks | Risk register | id, project_id, title, impact_score

One line per table. Grep it. Done.

It Stays Fresh

The skill tells Claude to update schema-index.md after schema changes. For automatic reminders, add this hook to your project's .claude/settings.local.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "mcp__supabase__apply_migration",
        "hooks": [
          {
            "type": "command",
            "command": "echo '📝 Schema changed - update schema-index.md with the table/columns you just created or modified.'"
          }
        ]
      }
    ]
  }
}

When Claude runs a migration, it gets reminded to add the new table directly - no need to re-query all 318 tables. Just one line edit.

# Quick test after install
grep "user" schema-index.md

Shows the value prop immediately.

Real World Test

Ran this on PulseAI - a 543k LOC enterprise SaaS with 318 database tables:

Metric	Before	After
Tables	318	318
Schema lookup tokens	~32,000	0
Output file	-	490 lines
Domains auto-organized	-	40+

The AI grouped tables into logical domains: Organizations, Projects, RAID, AI Insights, Pulse AI System, Financial, Resources, and 35 more. One line per table, grep-searchable, foreign keys visible.

Why

Approach	Tokens	Speed
Raw MCP list_tables	~25,000	Slow
Distilled index	~0	Instant

---

Built by Austin because context windows aren't infinite and MCP calls aren't free.