# Connect a thinking garden to ChatGPT

## What the connection is for

A connected tool allows a conversation to save, retrieve and manage records in a persistent garden. It reduces the switch from exploring an idea to opening another application. The plugin is a bridge to the collection, not the collection itself and not a substitute for deciding what is worth keeping.

You can use the conversational method without a plugin. Add the connection when repeated manual transfer has become a real source of lost material.

## Start with a small toolset

Useful actions are:

- **Search:** find possible duplicates or retrieve material by phrase, title, label or source.
- **Get item:** return a complete record by stable ID before quoting or editing it.
- **Add note:** save an approved, self-contained thought with provenance and optional metadata.
- **Add source:** save an external item separately from the thought it prompted.
- **Link items:** connect two existing records, optionally describing the relationship.
- **Update or remove:** allow correction and deletion with a specific ID and clear response. If the first version lacks these actions, describe the limitation plainly.

Each action should report what it actually did. A proposed note is not saved merely because ChatGPT displayed it in a conversation.

## Define the data contract

An add-note action might accept:

```json
{
  "title": "A specific title",
  "text": "A self-contained paragraph",
  "provenance": "Discussion of a named source; interpretation marked",
  "topics": ["existing broad topic"],
  "labels": ["specific search term"],
  "source_ids": ["existing-source-id"]
}
```

Required fields should be few. Validate text length, known topic names and referenced IDs. Return the new technical ID and, if used, a stable human-facing ID. Do not silently create a link to a source the tool could not find.

## Permission and publication

Separate public reading from authenticated management. A public garden may allow anyone to browse, while only the owner may add, edit or delete. Verify permissions on the server side for every write operation. Tool instructions alone are not access control.

If different gardens hold public research and private personal material, give them separate destinations and clear names. A conversational tool must not guess the destination from a vague keyword. It should confirm where an approved item was saved.

## Write tool descriptions for the actual workflow

Descriptions should say when to use an action and what it does. For example:

```text
Add an approved self-contained note to the owner's public thinking garden.
Use only after the owner has selected or requested saving the note. Preserve
uncertainty and provenance. Return its stable reference and destination.
```

```text
Search notes and sources in the garden before adding a likely duplicate,
or when the owner asks to revisit a previously saved thought.
```

Avoid promising “automatic knowledge extraction” if the working practice involves review, selection and occasional direct entry.

## Implementation outline

The exact product interface and authentication process will change. The durable components are:

1. persistent storage for notes, sources and links;
2. read and write functions with explicit schemas;
3. server-side owner checks for mutations;
4. stable IDs and a policy for merges and deletions;
5. a plugin or MCP connection that exposes the functions to ChatGPT;
6. instructions that distinguish proposing material from saving it;
7. a test that the newly saved item can be found through both the tool and site.

If using ChatGPT Sites, follow the current Sites instructions for exposing the functions and installing the resulting plugin. A working site endpoint is not the same as a selectable, connected plugin in ChatGPT. Test both steps.

## Test with harmless records

Add a short demonstration note and Source, link them, retrieve both by ID, search by an unexpected phrase, revise one, and remove a disposable record. Refresh the site and check that the data persists. Test that an unauthorised visitor can browse only what is public and cannot write. Check ambiguous requests such as “remove that last one” before enabling destructive actions.

The connection is successful when it makes careful capture and correction easier in an ordinary conversation. A green deployment message alone is less persuasive.