Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/xxyoudeadpunkxx/decision-rain-library-project/llms.txt

Use this file to discover all available pages before exploring further.

Decision Rain Library Project works with any AI assistant that can read and write Raindrop bookmarks. The most common integration is ChatGPT with the official Raindrop connector. The connector description bootstraps the assistant’s behavior at the start of a session — it does not replace the SYSTEM documents stored in 00_SYSTEM. Think of the connector description as the first instruction and 00_SYSTEM as the full rulebook.

​
Prerequisites

Before connecting an assistant, make sure you have:
  • An active Raindrop account with the five-collection structure already created (see Collections).
  • SYSTEM documents loaded as bookmarks in 00_SYSTEM with governance text in their note or excerpt fields (see SYSTEM Docs).
  • An AI assistant with Raindrop integration support — ChatGPT Plus or Team with connector access, or another MCP-capable assistant.

​
Connecting ChatGPT to Raindrop

ChatGPT’s connector screens change over time, so this is direction rather than a permanent step-by-step UI guide. The general path is:
1

Open ChatGPT settings

Navigate to your ChatGPT account settings and look for the section covering connectors, integrations, or plugins — the label varies by interface version.
2

Find the Raindrop connector

Search for Raindrop in the connector or integration library. The official connector is published by Raindrop.io.
3

Authorize access

Follow the OAuth flow to grant ChatGPT read and write access to your Raindrop account. The connector needs access to all collections you intend to use.
4

Add a connector description

If the connector interface allows a custom description or instruction field, add the recommended description from the section below. This step bootstraps the assistant before it reads 00_SYSTEM.
If the connector screens look different from what you expect, check Raindrop’s official documentation and ChatGPT’s connector help pages for current instructions. The exact UI is not stable enough to document step-by-step here.

​
Writing the Connector Description

A connector description bootstraps the assistant’s behavior before it reads SYSTEM. Keep it short — its job is to orient the assistant, not to replace the full rule set. Use this exact description: 
Use Raindrop as a Decision Library backend, not as generic bookmarks.
New links should first be saved to 00_INBOX.
Before reviewing or moving links, read 00_SYSTEM.
ChatGPT may inspect and propose, but the user decides final placement, tags, and next steps.
Each line does a specific job:

Backend framing

“Use Raindrop as a Decision Library backend, not as generic bookmarks” prevents the assistant from treating your collections like a general-purpose save-everything bookmark folder. It signals that collections represent lifecycle states.

INBOX-first

“New links should first be saved to 00_INBOX” preserves the trace. Every new signal gets captured before analysis begins, so nothing is lost if inspection fails or is incomplete.

SYSTEM read

“Before reviewing or moving links, read 00_SYSTEM” ensures the assistant loads your controlled taxonomy, note template, and decision rules before proposing anything. Without this, the assistant works from defaults.

Propose, don't validate

“ChatGPT may inspect and propose, but the user decides” enforces the operator gate. The assistant cannot promote, archive, or finalize anything — that authority stays with you.

​
MCP and Other Connectors

If you use an MCP-compatible assistant other than ChatGPT, the same principles apply. The assistant needs read and write access to Raindrop collections. The connector description or system prompt should include the same four instructions:
  1. Treat Raindrop as the Decision Library backend.
  2. Save new links to 00_INBOX first.
  3. Read 00_SYSTEM before serious work.
  4. Propose but do not validate — the operator decides.
The specific configuration mechanism differs by assistant and MCP host, but the behavioral contract is identical. SYSTEM documents in 00_SYSTEM remain the canonical rule source regardless of which assistant you use.

​
Conservative Write Behavior

Connector implementations should prefer a preserved pending entry over an overconfident classified entry. If the Raindrop read/write API behaves unexpectedly during a session — slow responses, partial writes, or unclear confirmation — the assistant should stop writing and report the situation rather than retrying with guesses.After any write operation, the assistant should be able to tell you: what was saved, what was changed, and what still requires your review. If it cannot confirm those three things, treat the operation as incomplete.

​
Troubleshooting

Check that the 00_SYSTEM bookmarks have governance text in their excerpt or note field, not just a placeholder URL. Raindrop requires a URL for every bookmark, but the URL destination is not the content — the note field is. If the note is empty, the assistant has nothing to read even if it can access the bookmark.To verify: open each 00_SYSTEM bookmark in Raindrop and confirm the note or excerpt field contains the full governance text from the corresponding docs/ file in the repository.
This means the operator gate is not being enforced. Take these steps:
  1. Remind the assistant of the rule: no entry reaches 20_LIBRARY without explicit operator approval for that specific entry.
  2. Update the connector description to include the propose-don’t-validate instruction if it is missing.
  3. Verify that 00_AGENT_CONTRACT is loaded in 00_SYSTEM with the full authority boundary text in its note field — this document is the primary source of the gate rule.
If the assistant continues writing to 20_LIBRARY without approval, treat any entries there as unvalidated and move them to 10_REVIEW until you have reviewed them.
API errors during a session should not result in silent data loss or silent partial classification. When an error occurs, the assistant should report:
  • what it attempted to save or change
  • what was confirmed written (if anything)
  • what is still pending or unresolved
  • the minimum action needed from you to recover
Do not ask the assistant to retry automatically. Instead, check the affected entries in Raindrop directly, confirm their state, and then resume from a known-good point.

Build docs developers (and LLMs) love

Get started for freeTalk to us