← Back to the app
Setup plan

Claude, Google Drive, one master sheet, and an app on top

How the workspace would be organized so research stays clean and Claude can help without touching the master file.

01

Google Drive structure

One master sheet is the single source of truth. Claude works on copies and drafts; nothing it produces lands on top of the master.

Product Research/
├── 00_MASTER/            ← you edit; Claude reads only
│   └── Product_Database  (Google Sheet: 4 tabs)
├── 01_CLAUDE_WORKING/    ← Claude's drafts, always new files
│   └── 2026-11-19_audit_shampoo.md
├── 02_SOURCES/           ← labels, brand PDFs, screenshots
│   └── P001_Mossline_label.jpg
├── 03_ROUTINES/          ← AM / PM / weekly routines (Docs)
└── 99_ARCHIVE/           ← dated weekly copies of the master
    └── Product_Database_2026-11-16

Naming rule: YYYY-MM-DD_topic for drafts, P###_brand_item for sources. Dates first so files sort in order and duplicates are easy to spot.

02

Normalized database (one Google Sheet)

Each fact lives in one place. An ingredient is described once and linked to every product that contains it, so fixing a note fixes it everywhere.

TabOne row perKey columns
Productsproductproduct_id, brand, product_name, category, size, price_usd, status (Using / Testing / Wishlist / Avoid), my_rating, last_checked, notes
Ingredientsingredientingredient_id, inci_name, common_name, flag_group, note
Product_Ingredientsproduct × ingredientproduct_id, ingredient_id, label_position
Flag_Groupsavoid categoryflag_group, default_on, why_i_avoid, source_link

Data validation on status, category and flag_group (drop-downs) keeps spelling consistent, which is what makes the data "app-ready". The demo app reads exactly these tables.

03

Guardrails so Claude never overwrites the master

Instructions alone are not a lock, so the plan uses three layers. The first one is enforced by Google, not by Claude.

Layer 1: Permissions (hard lock)Claude only gets read access to 00_MASTER. Project knowledge uses read-only snapshots of the master. If you ever connect a tool that can write to Drive, it signs in with a separate Google account that is Viewer on 00_MASTER and Editor only on 01_CLAUDE_WORKING.
Layer 2: Project instructionsA short rules block at the top of the Claude Project (sample below), plus a check-before-create step so Claude searches for an existing file before making a new one.
Layer 3: RecoverySheet version history, plus a small Apps Script that copies the master into 99_ARCHIVE every Sunday. A bad edit costs minutes, not research.
PROJECT RULES (excerpt)
1. 00_MASTER is read-only. Never edit, rename, move or re-upload it.
2. New output = new file in 01_CLAUDE_WORKING named YYYY-MM-DD_topic.
3. Before creating a file, list existing files with the same topic.
   If one exists, propose an update in chat instead of a duplicate.
4. Product audits: use product_id from the master. Never invent IDs.
5. Proposed sheet changes go in a table (product_id, column, old, new)
   for me to paste in. You never change the master directly.
04

Phase 2: from sheet to phone app

Google Sheet
→ published CSV / Apps Script JSON →
Small web app (PWA)
→
Desktop + Android home screen
  • Keep the Sheet as the database. You keep editing where you already work; the app refreshes from it.
  • Front end: a lightweight installable web app (like the demo): search, filters, your avoid list, works offline.
  • Hosting: free static hosting (for example Cloudflare Pages). No monthly cost.
  • Editing from the phone: add a simple form backed by Apps Script that appends rows. Move to a real database (for example Supabase) only if you outgrow the sheet or need multiple users.
  • Privacy: the app can be private (sign-in) if the research should not be public.

Demo only: brands and products are fictional sample data. Ingredient notes are neutral summaries, not medical advice.