Trajectory - Inner Map
Slug: project-requirements-document-vault
Source: /Users/nitishchauhan/Downloads/Project Requirements Document - Vault.pdf (extracted: ~/.cache/gemma14-sources/Project_Requirements_Document_-_Vault.pdf.txt)
Index: Index - Project Requirements Document Vault
Pipeline: CONSTITUTION - Publishable Asset Pipeline · 00 - Master Index
Mode lean: Authority (product thesis + anti-scope)
This source is a product-spec PRD, not a multi-turn chat. The “prompt spine” is the authored decision sequence inside the document: what Nitish forced into order so the product could not drift into second-brain territory.
1. Prompt spine (decision order in the PRD)
- Name the product and the single job: universal conversation capture + export for cognitive closure.
- State anti-vision hard: not KM, not second brain, not AI memory platform. Export-first utility only.
- List the real user pain: multi-platform chats, fear of loss, tab hoarding, slow copy-paste, tools that chase memory instead of capture.
- Lock core philosophy: never interrupt thinking; capture accurately; interpretation optional; capture mandatory.
- Freeze five design principles: Capture First, User Control, Low Cognitive Load, Universal Support, Modular Architecture.
- Name primary users (heavy AI users, researchers, writers, developers, students, creators, knowledge workers).
- Define the primary flow end-to-end: finish chat → Vault → queue → browse more → multi-site queue → sidebar review → export together → mental loop closed.
- Specify V1 features that serve that flow only: universal capture, sidebar, queue, export formats, metadata, assets, local buffer, generic webpage capture.
- Fence V1 out of scope with a harsh exclusion list (summaries, embeddings, RAG, graphs, generation, recommendations, auto-tagging).
- Sketch V1.5 as optional 48-hour temporary cloud vault (safety without hoarding).
- Park future modules as independent (MCP, self-hosted archive, plugins); capture never depends on AI processing.
- Technical direction: React/TS/Tailwind, MV3, IndexedDB, modular extractors, normalize → queue → export.
- Success criteria centered on reliability, formatting fidelity, multi-export, one-minute understandability, fewer fear-tabs.
- End with the permanent feature gate: capture-and-close-the-loop vs turn-into-another-KM-system.
2. Start state
Initial itch is not “I need a better notes app.”
It is: valuable AI conversations are scattered, fragile, and unfinished in the mind, so people keep tabs open as a prosthetic memory. Existing tools respond by building second brains and AI memory layers. That increases ambition and still fails the basic job: get the conversation out safely, then let the person decide later.
Conflict at the door:
- Want one universal capture surface across ChatGPT, Claude, Gemini, Grok, Perplexity, Cursor, Poe, DeepSeek, and whatever comes next.
- Refuse to let that surface become Notion-with-embeddings.
- Need a product identity that is export-first cognitive closure, not knowledge management cosplay.
3. Transitions and impulses
P1 → P2 (Name the job → anti-vision)
- Identity cut: if you do not say what it is not, feature requests will define it for you.
- Scope fear: KM language (“second brain,” “memory platform”) is the default gravitational well for anything that touches notes or chats.
- Product honesty: capture without interpretation is enough for V1; “smart later” is optional, not the core.
P2 → P3 (Anti-vision → problem catalog)
- Lived symptom: tabs left open as unfinished mental loops, not laziness.
- Fragmentation: same fear across many AI sites, no unified export path.
- Market mismatch itch: competitors optimize AI memory; the missing utility is reliable capture.
P3 → P4 (Problems → philosophy)
- Protect thinking time: the extension must never interrupt flow with prompts or rewrites.
- Role split: human interprets; machine preserves.
- Binary discipline: optional interpretation, mandatory capture.
P4 → P5 (Philosophy → design principles)
- Operationalize the thesis: “capture first” becomes never summarize / never rewrite / never modify.
- Trust boundary: user decides what matters; product does not rank, tag, or recommend in V1.
- Maintenance realism: modular per-site extractors so layout churn does not sink the product.
P5 → P6 to P7 (Principles → users + primary flow)
- Who feels the pain first: heavy multi-platform AI users (research, writing, code, study, content).
- Director mode on the loop: success is not “saved somewhere”; success is mental loop closed after multi-item export.
- Queue as staging, not storage: collect across sites, then one export action.
P7 → P8 (Flow → V1 features)
- Only features that serve the flow: sidebar as temporary workspace, multi-select queue, faithful formats, metadata, assets, buffer/retry, generic fallback.
- Reject permanence creep: queue is not a library.
- Faithfulness over cleverness: preserve images, tables, code, links when possible.
P8 → P9 (Features → harsh out-of-scope)
- Anti-creep peak: summaries, embeddings, RAG, graphs, LinkedIn/Substack generation, recommendations all feel “smart” and all pull identity off capture.
- Future parking lot is not V1 fuel: saying “later” out loud protects the core.
- Identity gate: if it turns Vault into KM, it is out.
P9 → P10 to P11 (Out-of-scope → V1.5 + future modules)
- Safety without hoarding: 48-hour temporary cloud vault, then auto-delete.
- Independence rule: future AI modules must never make capture depend on processing.
- Hoarding anxiety: permanent free cloud storage would reintroduce digital guilt the product is trying to reduce.
P11 → P12 to P14 (Future → tech, success, gate question)
- Engineering follows philosophy: modular extractors, normalizer, queue, exporter; React/TS/Tailwind; MV3; IndexedDB.
- Success as behavior change: fewer fear-tabs, one-minute UI clarity, multi-queue export works.
- Permanent decision rule: every feature request hits the loop-closure vs KM question first.
4. Inner state arc
Start: scattered AI work + fear of loss + tab prosthetics; irritation at tools that sell memory instead of capture.
Shift 1: product is not storage ambition; product is closing the mental loop.
Shift 2: philosophy hardens into non-negotiables (never rewrite; user decides; low load; universal; modular).
Peak heat: the exclusion list. Naming what V1 will not do is the real product design move. Without it, Vault becomes another second brain demo.
Secondary relief: V1.5 temporary vault and future modules are allowed only if they stay independent and do not corrupt capture.
End hunt: a durable gate question and success criteria that measure reliability and loop-closure, not feature count. The document is hunting product identity under pressure, not a kitchen-sink roadmap.
5. Speech habits visible in this source only
- Short anti-definitions stacked for force: “This is not… / This is not… / This is not…”
- Single-sentence job statements repeated as north stars (“capture… close the mental loop”).
- Absolute capture language: never summarize, never rewrite, never modify, always preserve.
- Flow drawn as down-arrows ending in “Mental loop is closed.”
- Harsh exclusion lists as identity tools, not afterthoughts.
- One recurring feature-gate question as the constitution of the product.
- Modular file examples (
extractors/chatgpt.ts,generic.ts) as engineering speech for the same modular idea. - No chat banter, no “bro,” no playful digression. Spec tone: declarative, fenced, principle-first.
6. Write-as-Nitish steering notes (for Final only)
- Open from the tab-as-prosthetic / unfinished loop feeling, not from a generic PRD summary.
- Thesis line early: cognitive closure via export-first capture; refuse second-brain gravity.
- Keep value blocks close to source language for principles, flow, V1 features, out-of-scope, V1.5, tech stack, success, gate question.
- Authority mode: teach the decision rule; do not sermonize “playful learning.”
- Unique opening: no Gemma 13 clone phrases; no “Wait. Before the map”; no “binary is boring.”
- No em dashes or en dashes in new prose.