Todoist

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 a task manager that replaces Todoist: one line of typing becomes a task with a date, a project and a priority, and it is on every device a second later.

STACK
- Node 20+ with Fastify for the server
- SQLite through better-sqlite3, WAL mode, with FTS5
- A PWA that works entirely offline, which is the only sane way to be on every platform without writing four applications
- Caddy in front

THE DATA MODEL
- tasks: id, content, description_md, project_id, section_id, parent_id, position, priority, due_date, due_datetime, due_timezone, due_recurrence, deadline_date, completed_at, created_at, updated_at, deleted_at
- projects: id, parent_id, name, colour, view_kind, is_favourite, is_inbox, position, archived_at
- sections, labels, task_labels
- comments: id, task_id, body_md, author_id, attachment_path, created_at
- reminders: id, task_id, kind, at, offset_minutes, sent_at
- activity: id, entity_kind, entity_id, action, actor_id, payload_json, at — append-only, and what makes an undo and a history possible
- shares: project_id, user_id, role
- Every instant stored in UTC with the intended zone kept beside it. A task due at 09:00 tomorrow is due at 09:00 wherever you happen to be, which means storing the intent and not only the instant

QUICK ADD, WHICH IS THE PRODUCT
- One field. Type a sentence and it becomes a task: 'pay invoice friday 5pm #finance @waiting p1' gives a title, a due date and time, a project, a label and a priority
- Parse dates in the forms people actually type: today, tomorrow, fri, next monday, in 3 days, 15/03, every weekday, every 2 weeks starting monday, end of month
- Ambiguity resolved conservatively and shown before saving — the parsed result is displayed as chips that can be corrected, never applied silently. A tool that guesses wrong and hides it is worse than one that asks
- Parsing runs on the device, offline, instantly. This is the one interaction that has to be under fifty milliseconds or the whole tool feels slow
- Everything typeable is also selectable, because half of people never learn the syntax

RECURRENCE, WHICH IS WHERE THESE GO WRONG
- Two kinds, and they are genuinely different: every Monday, and every 2 weeks after completion. The second reschedules from when you actually did it
- Completing a recurring task creates the next occurrence and keeps a record of the completion. The history must survive
- Daylight saving handled by storing the intent: 'every day at 09:00 in Europe/Rome' is recomputed against the zone, not shifted by an offset
- End conditions: a date, a count, or never
- Skipping an occurrence is different from completing it, and both are recorded

OFFLINE AND SYNC
- Everything works with no network: create, edit, complete, reorder. The device is the source of truth until it syncs
- A change queue on the device, sent when there is a connection, ordered, idempotent by a client-generated identifier
- The server's change feed ordered by a server counter, never by a device clock
- Conflicts resolved per field with last-write-wins, except completion, which is sticky — a task completed on one device must not be resurrected by a stale edit from another
- Sync within a second when there is a connection, which is what makes a task manager feel trustworthy across devices

VIEWS
- Today, upcoming, inbox, per project, and filters as saved queries over date, project, label, priority, assignee and text
- The filter language is the same syntax as quick add, which means one parser and one thing to learn
- Board view per project as an alternative to a list
- Sorting manual, by date, or by priority
- Keyboard for everything

REMINDERS
- Absolute or relative to the due time, per task, with a default per project
- Delivered by push where the platform allows, and by email as the reliable fallback
- Recorded when sent, so nobody is reminded twice for the same occurrence

SHARING
- A project shared with other people, tasks assigned to one of them, comments with mentions
- Notification on assignment and on a comment, batched
- Permissions checked on the server for every read and write

SEARCH AND ESCAPE
- FTS5 over content, descriptions and comments
- Export everything — tasks, completed history, projects, labels and comments — as JSON and CSV in one command
- Import from the common exporters
- Years of completed tasks is a record of what you did with your time and must never be trapped

OPERATIONS
- .env: DATABASE_PATH, BASE_URL, SESSION_SECRET, SMTP_URL, VAPID keys for push
- Migrations on boot, each once
- Nightly backup off the machine, restore script
- Health endpoint

WHAT MATTERS MOST
The parser and offline sync. Build natural-language parsing with its chips-and-correction display first, then the change queue and the completion rule, and use it on a phone with the network off for a week. Everything else is a list.

Give me the repository, the PWA, the parser with its tests, migrations, .env.example, and a README with deploy steps behind Caddy.