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.

The Decision Rain review loop is the heart of the project. It is a six-step cycle that takes a raw link — a GitHub repo, a tool page, a research spark — and turns it into a bookmark with a written judgment attached. The assistant does the inspection work; the operator makes every consequential call. Understanding where that boundary sits is the key to using the system well.

​
The six-step loop

1

Give ChatGPT a link

You share a URL with the assistant — a repository, a documentation page, a product, a guide, or anything that caught your attention. You do not need to explain why it seemed interesting. Capturing the trace is enough.
2

Save the original link to INBOX

Before doing anything else, the assistant saves the original URL to the 00_INBOX collection in Raindrop. This step is never skipped. Preserving the trace before analysis prevents the assistant from treating capture as classification.
3

Read the page and available sources

The assistant inspects the link using available tools. It reads the page, repository README, documentation, source code samples, pricing page, release notes, and any other official sources it can reach. It also gathers community signals: issues, discussions, forum threads, Hacker News and Reddit mentions, and adoption reports.
4

Write a structured review

The assistant produces a compact review note using the standard template. The review covers what the link claims to provide, what evidence supports or weakens that claim, where the evidence comes from, how confident the assessment is, and what the practical next action is. The note is designed to be scannable later without re-reading the original link.
5

You decide

The assistant presents its proposal — suggested tags, collection, verdict, and next step. You review it and decide whether to accept, correct, or reject it. The assistant may not move an entry, assign final tags, or mark a decision as validated without your explicit approval. A clean proposal is not approval; silence is not approval.
6

Move the entry to the right collection

After you validate the proposal, the assistant moves the entry to the appropriate collection: 10_REVIEW if more analysis is needed, 20_LIBRARY if the decision is confirmed, or 90_ARCHIVE if the link is rejected, outdated, or no longer relevant.

​
The AI proposes — the operator decides

The most important thing to understand about the loop is the authority boundary. The assistant has proposal authority: it can inspect, research, describe, tag suggestions, surface risks, and recommend a decision path. It does not have validation authority. Only you can:
  • mark an entry as a canonical library decision
  • promote an entry into 20_LIBRARY
  • archive an entry
  • approve new tag values
  • change SYSTEM documents
Previous similar approvals, high assistant confidence, clean formatting, and strong source evidence do not count as validation. Every promotion requires an explicit operator action.

​
The review note shape

Every reviewed entry uses a consistent note template so that the key judgment is always in the same place. Because Raindrop may truncate long excerpts, Verdict and Next appear near the top. 
Cataloged: YYYY-MM-DD
Verdict: short practical judgment.
Next: the next useful action.
Claim: what the link appears to provide.
Evidence: what supports or weakens that claim.
Authority: where the evidence comes from.
Truth: how confident the review is.
StackFit: practical judgment of how realistically the operator can get value from the item.
The three fields Claim, Evidence, and Verdict are intentionally kept separate. The claim states what the item appears to offer. Evidence states what the assistant found to support or contradict that. Verdict states the practical judgment after weighing all of it. Collapsing these three into a single opinion defeats the purpose of a structured review.

​
The five collections and how a link flows through them

Collections are lifecycle states, not topical folders. A link moves through them in a defined sequence.
CollectionRole
00_SYSTEMGovernance documents only. Rules, templates, examples, and the agent contract live here.
00_INBOXEvery new link lands here first, without exception, before any analysis.
10_REVIEWLinks that have been analyzed and are worth a closer look before a final decision.
20_LIBRARYEvaluated, operator-validated decisions. Only entries the operator has explicitly approved belong here.
90_ARCHIVERejected, outdated, duplicated, inactive, or historical entries.
The normal flow through the state machine is: 
signal -> 00_INBOX -> analysis proposal -> operator validation -> 10_REVIEW / 20_LIBRARY / 90_ARCHIVE
A direct write into 20_LIBRARY is only valid after explicit operator validation — regardless of how strong the evidence is.

​
What evidence the assistant gathers

The assistant keeps two categories of evidence separate, because they answer different questions. Official evidence describes intended behavior. It includes:
  • documentation and README claims
  • source code and examples
  • release notes and changelogs
  • pricing pages and maintainer notes
  • product pages
Community evidence reveals operational friction. It includes:
  • GitHub issues and discussions
  • forum and mailing list reports
  • Reddit and Hacker News signals
  • adoption friction, setup failures, and pricing complaints
  • real-world breakage reports
If official and community evidence conflict, the conflict must be surfaced in the review before a verdict is written. A useful decision note does not blend these two sources into a single narrative.

​
When evidence is insufficient

The assistant must not classify an entry from insufficient evidence. If it cannot find enough information to write a confident review — or if it cannot read the SYSTEM governance documents from Raindrop — it stops, reports what is missing, and waits for operator guidance.
If you ask the assistant to review a link and it returns a partial note with unresolved questions, that is the system working correctly. Preserving the uncertainty is more useful than filling in confident guesses.

Build docs developers (and LLMs) love

Get started for freeTalk to us