Reflect

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me the networked notes system that replaces Reflect.

The honest starting point: **Obsidian is free, stores plain markdown, and does the backlinks.** Build your own if you want the web version, the assistant and the sync on your own terms.

STACK
- Node 20+ with Fastify, server-rendered HTML with a little vanilla JS
- SQLite through better-sqlite3, WAL mode, FTS5 across notes
- Markdown files on disk as the source of truth, the database as an index
- A local model through Ollama for the assistant
- Caddy in front

WHY FILES AND NOT A DATABASE
- **The notes must outlive the software.** A folder of markdown is readable in thirty years by anything; a proprietary store is readable by one program
- The database is an index that can be rebuilt from the files at any time. Deleting it must lose nothing

THE DATA MODEL
- notes: id, path, title, body_md, created_at, updated_at, sha256
- links: from_note_id, to_title, to_note_id, context — **the context sentence is what makes a backlink useful**, because a list of titles tells you nothing about why
- daily: date, note_id
- tags, extracted from the body rather than stored separately

THE WRITING SURFACE
- Markdown, with `[[wiki links]]` autocompleting as you type against existing titles
- **Creating a link to a note that does not exist creates the note lazily**, on first write. That single behaviour is what makes networked notes actually accumulate
- A daily note opened by default, dated, which is where most writing starts
- Everything keyboard reachable, and a command palette

BACKLINKS DONE PROPERLY
- Every note shows what links to it, with the surrounding sentence from each
- Unlinked mentions: other notes containing this title as plain text, offered as suggestions. **This is how a note graph grows without discipline**
- Rebuild the link index on file change, incrementally

SEARCH
- FTS5 with prefix matching and title weighting, fast enough to search as you type
- Search inside the link autocomplete too, so linking and finding are one action

THE ASSISTANT, SCOPED NARROWLY
- Retrieval over your own notes, and the model rewrites what it retrieved. **It must cite the notes it used, always**, or you cannot tell your own thinking from its invention
- Useful prompts: summarise this week's daily notes, find notes related to this one, list the open questions I wrote and never answered
- Run it locally through Ollama so nothing you write leaves the machine. **That is the point of an assistant over a private journal**

SYNC AND SAFETY
- Git, or a file sync tool. Both are solved and both give you history for free
- If you build sync: encrypt on the client, and the server stores ciphertext it cannot read
- Nightly backup off the machine, and an export that is just the folder

WHAT MATTERS MOST
Files on disk, link context on every backlink, and a local assistant. The first protects the notes, the second makes the graph useful, the third makes it private.