Vault Project Requirements

Or: the product is not “save the chat.” The product is closing the mental loop.

Source: /Users/nitishchauhan/Downloads/Project Requirements Document - Vault.pdf
Phase: 3 (Final)
Trajectory: Trajectory - Inner Map
Index: Index - Project Requirements Document Vault
Pipeline: CONSTITUTION - Publishable Asset Pipeline · 00 - Master Index
Slug folder: Work/project-requirements-document-vault/
Status: draft
Mode: Authority

How this one actually started

It started with tabs that would not die.

Not because the page was still loading. Because the conversation was still open in the head. A useful exchange on Claude. A half-finished architecture dump on ChatGPT. A research trail on Perplexity. A Cursor thread that finally made the bug make sense. Closing the tab felt like throwing away unfinished thought. So the tabs stayed. The browser became a prosthetic working memory. The fear was not “I might want this someday.” The fear was “if I close this before I capture it, the loop never closes.”

Most tools answer that fear by getting smarter: summaries, embeddings, second brains, AI memory layers. This PRD answers by getting smaller and stricter.

Working title: Vault.
Job: universal conversation capture and export for cognitive closure.
Not: a knowledge management system.
Not: a second brain.
Not: an AI memory platform.

Export-first. Faithful preserve now. Decide later whether the thing deserved more attention.

That is the whole identity fight. Everything below is how the document fences that identity so feature gravity cannot drag it into KM cosplay.

The real problem is unfinished mental loops

Modern AI work is multi-site by default. The same week can span:

ChatGPT
Claude
Gemini
Grok
Perplexity
Cursor
Poe
DeepSeek
Qwen
Mistral
OpenRouter
whatever ships next week

Those threads often hold real ideas. The failure mode is not lack of intelligence features. It is capture friction:

fear of losing important conversations
conversations scattered across websites
no unified export workflow
manual copy-paste that is slow and inconsistent
tabs left open because forgetting feels expensive
existing solutions that lean into AI memory instead of reliable capture

Result: unnecessary cognitive load and unfinished mental loops.

If you misread the problem as “I need a better archive,” you will build a library. If you read it as “I need to close the loop without interrupting thinking,” you will build a capture utility. Vault is the second bet.

Core philosophy: capture is mandatory

The extension should never interrupt thinking.

It should quietly do one job:

Capture everything accurately.

Interpretation is optional.
Capture is mandatory.

That split is the product. Humans judge value. The tool preserves the raw material so judgment can happen later without the tab tax.

Design principles (V1 constitution)

1. Capture First

Never summarize.
Never rewrite.
Never modify.
Always preserve.

If the tool starts “helping” by compressing meaning at capture time, it is already competing with the user’s future judgment.

2. User Control

The extension never decides what matters.
The user decides.

No auto-ranking. No silent prioritization. No “we thought this was important.”

3. Low Cognitive Load

Every interaction should reduce mental effort.
No unnecessary prompts.
No complicated settings.
Minimal clicks.

If capture feels like a project, people will keep using tabs as memory.

4. Universal Support

Avoid building only for one AI site.
Support conversation websites whenever possible.
Fall back to full webpage capture if structured extraction is unavailable.

Platform churn is permanent. Universality is the only durable strategy.

5. Modular Architecture

Every supported website should have an independent extractor.
That keeps maintenance sane when layouts change.

One broken site should not take down the whole product.

Who this is for

Primary users in the PRD:

heavy AI users
researchers
writers
developers
students
content creators
knowledge workers

Shared trait: multi-platform conversation load high enough that copy-paste and tab-hoarding become daily tax.

Primary user flow (success ends at closure)

The document does not end success at “file saved.” It ends success when the mental loop closes.

User finishes an AI conversation.
Clicks the Vault extension.
Chooses Add to Queue.
Continues browsing.
Adds multiple conversations from different websites.
Opens sidebar.
Reviews queued conversations.
Selects desired items.
Exports everything together.
Mental loop is closed.

The queue is a temporary staging surface so multi-site work can become one clean export action. It is not permanent storage. That distinction matters: if the queue becomes a library, the product has already drifted.

Primary features (V1)

Universal conversation capture

Structured extraction when possible across major AI chat platforms.
If structured extraction fails, export the webpage faithfully.

Reliability beats cleverness. A perfect extractor that fails silently is worse than a generic capture that still closes the loop.

Displays:

conversation title
source website
capture time
selected status

Allows:

multi-select
remove
rename
reorder
search within the current queue

Again: temporary workspace, not archive UI.

Queue

Collect multiple conversations before exporting.
Exists only to simplify workflow.
Not permanent storage.

Export formats

Markdown (.md)
JSON
HTML
Plain text
ZIP package (optional)

Metadata on every export

title
URL
website
timestamp
export timestamp
browser
model (when detectable)
conversation identifier (if available)

Metadata is how future-you can place the capture without forcing organization at capture time.

Asset capture

Preserve when technically possible:

images
tables
code blocks
Markdown formatting
hyperlinks
PDFs
attachments

Faithful conversation means more than text blobs.

Local buffer

If export fails: store temporarily, retry automatically.
Never lose captured conversations because of network interruptions.

Trust is the product. One silent loss and people go back to fear-tabs.

Generic webpage capture

Also works outside AI sites:

capture webpage
capture selected text
capture conversation
queue for export

Universal support is incomplete without a non-AI fallback.

Out of scope for V1 (identity fence)

These are intentionally excluded, not forgotten:

AI summaries
auto-tagging
embeddings
vector databases
knowledge graphs
RAG
semantic search
content generation
mind maps
LinkedIn generation
Substack generation
automatic categorization
AI recommendations

They can belong in future versions. They do not belong in core V1.

This list is not anti-ambition. It is anti-identity-theft. Every one of those features is a magnetic “smart product” story. Together they pull Vault toward the exact category the vision rejects.

Version 1.5: temporary cloud without hoarding

Optional cloud-based temporary vault:

Export → Upload → Available for 48 hours → User chooses save permanently / download / ignore → Automatic deletion after 48 hours.

Purpose in the PRD’s own words: temporary security without encouraging long-term digital hoarding.

That is a careful compromise. Safety net for device failure or handoff, without turning Vault into infinite free storage that restarts guilt and backlog culture.

Future versions (independent modules only)

Potential later capabilities:

AI-assisted summaries
content extraction
knowledge graph generation
search across exported conversations
MCP integration
self-hosted archive
plugin ecosystem
API access

Constraint that protects the core:

These should remain independent modules.
Capture must never depend on AI processing.

If capture requires a model call, the product has broken its own philosophy. Offline reliability and non-interruptive capture come first. Intelligence layers are optional attachments, not load-bearing walls.

Technical direction (enough to keep the thesis honest)

Frontend: React, TypeScript, Tailwind CSS

Extension: Chrome Manifest V3, WebExtension-compatible architecture

Storage: IndexedDB, Chrome Storage API

Export tooling: JSZip, Turndown, DOMPurify

Pipeline shape:

Content Script
  → Website Extractor
  → Normalizer
  → Sidebar Queue
  → Exporter
  → Local Storage
  → (Optional VPS Upload)

Extractors stay modular, for example:

extractors/
  chatgpt.ts
  claude.ts
  gemini.ts
  grok.ts
  perplexity.ts
  cursor.ts
  reddit.ts
  generic.ts

This is the engineering expression of the same product rule: origin-specific mess gets normalized once; queue and export do not care which site produced the capture; one site’s DOM churn does not rewrite the whole system.

Success criteria

Vault V1 succeeds if it can:

capture conversations reliably from supported platforms
preserve formatting with minimal loss
export multiple queued conversations in a single action
support generic webpage capture
keep the interface simple enough that a new user understands it within one minute
reduce the user’s need to leave tabs open out of fear of losing information

That last line is the behavioral north star. Feature completeness is not the scoreboard. Fewer fear-tabs is.

The permanent feature gate

Whenever a feature request appears, ask one question first:

Does this help users capture and close the loop, or does it turn Vault into another knowledge management system?

If the answer is the latter, it belongs in a future version, not in the core product.

That question is the PRD’s real artifact. Specs age. Site extractors rot. Formats get new siblings. The gate question is the durable constitution: cognitive closure first, KM ambition later and only as independent modules.

What I am actually defending here

I am not defending “less AI” as a moral position.

I am defending a job-to-be-done that keeps getting buried under smarter stacks:

multi-platform conversations are valuable and fragile
fear of loss creates tab prosthetics and unfinished loops
capture must be accurate, quiet, and universal
interpretation can wait
the product wins when the mental loop closes after export
every tempting KM feature must pass the loop-closure test or stay out of V1

Vault (working title) is an export-first utility for cognitive closure.
If it becomes a second brain, it failed its own requirements document, even if the demo looks impressive.

Curiosity Rooms

Explorer

Final - Vault Project Requirements