ChatGPT?

NOT REALLYA prototype of the interface is possible, but the durable product depends on difficult systems beyond the UI. Build a safer adjacent workflow instead.

BUILD FIRST

A non-sensitive or read-only version of the core experience
A manual review step wherever correctness or trust matters
A narrow companion tool that removes a real piece of work

LEAVE FOR LATER

Regulated decisions, payments, or irreversible actions
Claims of correctness, compliance, or uptime that the MVP cannot support
Sensitive integrations until security, support, and auditability are ready

RESEARCH NOTES

What this tool actually does.

Checked 2026-09-06

ChatGPT is a conversational AI product for multi-turn questions, writing, coding, file-based work, and web-assisted answers. Its Projects feature groups chats, uploaded files, and custom instructions into a focused workspace, while the product’s capabilities vary by account, model, and enabled tools.

SOURCES

RESEARCHED BUILD PROMPT

Give this to your coding agent.

Build a focused, private document-research assistant inspired by the project workflow of ChatGPT. Do not copy ChatGPT's brand, name, logo, proprietary assets, or UI.

Core workflow
- Let a user create named workspaces with separate chat history, custom instructions, and uploaded text/PDF/Markdown files. Stream multi-turn answers and show which workspace files support each answer.
- Index extracted text into a per-user retrieval store, retrieve relevant passages for each question, and render inline citations that open the source file and page/section when available. Keep a clear “not found in sources” response instead of guessing.
- Provide conversation search, rename/archive/delete/export controls, file processing progress and errors, model/provider selection, temperature/token settings, and an explicit setting for whether chats/files are retained.
- Use a provider adapter for an external LLM API and make prompts, retrieved chunks, model name, and citations inspectable in a developer/debug view.

Implementation and scope
- Use React, a Node or Python API, Postgres/SQLite metadata, object storage for uploads, a background ingestion job, and a vector index with strict workspace isolation. Add prompt-injection defenses for uploaded content.
- Out of scope: training a foundation model, broad multimodal parity, autonomous browser/computer actions, image/video generation, enterprise admin, and claims of ChatGPT-level model quality.
- Use seeded documents and a fake model adapter so the full upload → retrieve → answer → cite flow works without paid credentials.

Include a README with setup, provider configuration, file limits, privacy/retention behavior, retrieval design, and important limitations.

Visit the official product page