Timely

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 automatic time tracking that replaces Timely — and know why this is ALMOST.

The idea is **a memory of what you actually did**: the software records everything, and at the end of the day you turn that memory into timesheet entries rather than trying to remember. What makes it hard is agents on every platform, a categorisation layer that stays current, and the fact that this watches everything you do — so read the privacy section first.

STACK
- A small native agent per platform, plus a browser extension for per-site detail
- Node 20+ with Fastify, SQLite through better-sqlite3
- Caddy in front, on a private network

THE PRIVACY POSITION, FIRST
- This records your whole working day. Store the application and the domain by default; window titles only where you have opted in per application, because a title often contains a client's name or a document's
- The memory is private to the person until they draft an entry from it. Nobody else — not a manager, not an administrator — sees the raw timeline. That single rule is what makes people willing to run it
- No team dashboard of individuals' raw activity. The moment this is used to watch employees it stops being a memory aid and becomes surveillance
- A pause that is obvious and immediate, and private browsing never recorded

THE DATA MODEL
- devices: id, user_id, platform, token_hash, last_seen_at
- spans: id, device_id, at, duration_ms, app_name, domain, window_title, is_idle, is_private — one row per focused span, not per second
- calendar_events: id, user_id, external_id, starts_at, ends_at, title, attendees_json — meetings are hours too, and they are the easiest to forget
- documents: id, span_id, kind, ref — a file or a ticket identifier, where the agent can see one
- draft_entries: id, user_id, local_date, project_id, task_id, description, started_at, ended_at, duration_seconds, confidence, source_span_ids_json, status — the proposal
- entries: id, user_id, project_id, task_id, description, started_at, ended_at, duration_seconds, is_billable, is_locked, invoice_ref — the confirmed record
- entry_revisions, rates with effective dates, projects, clients, budgets — the ordinary tracking model underneath

FROM MEMORY TO TIMESHEET, WHICH IS THE PRODUCT
- At the end of the day, the timeline is grouped into blocks: contiguous spans on related applications, documents and domains, plus the calendar's meetings
- Each block is proposed as a draft entry with a project, a description and a duration. The project comes from rules and from what you chose last time for similar activity
- **Nothing is recorded as time worked until a person confirms it.** The proposal is a draft; the entry is a decision. That distinction is the whole ethical and practical basis of the tool
- Confirming or correcting a draft trains the matching for next time — the same domain, the same document, the same meeting title
- The description is generated from what was actually in the block and is editable. A generated description nobody edits is fine; one nobody can edit is not
- Show the confidence, and put low-confidence drafts last

THE AGENT
- Poll the focused window every few seconds and emit a span on change. Spans, not samples, or the storage is enormous
- Idle detection from the operating system's own idle timer, recorded as idle rather than attributed to whatever was on screen
- Queue locally when there is no network, so a laptop on a train loses nothing
- Small and quiet: it runs all day and must be invisible in memory and processor use

RATES AND BILLING
- Rates with an effective date, so an entry bills at the rate in force when the work happened. Changing a rate must never restate last quarter
- Billable and cost rates both, because the gap is margin
- Rounding applied at report or invoice time only. The raw seconds are the record
- Approval locks a period; a locked entry cannot change without a logged unlock
- Everything computed from rows at query time. No stored totals

WHAT IT SHOWS
- The day as a timeline, with the drafts beside it
- Hours by project, client and person; planned against actual if you also plan
- Unbilled work, which is the report that finds the money you forgot to invoice
- Time not accounted for, honestly shown rather than quietly filled

OPERATIONS
- .env: DATABASE_PATH, BASE_URL, SESSION_SECRET, RETENTION_DAYS, ENCRYPTION_KEY
- Migrations on boot, each once; raw spans on a stated retention with a sweeper, shorter than the entries they produced
- Nightly backup off the machine, encrypted
- Health endpoint reporting each device's last contact

WHAT MATTERS MOST
Spans rather than samples, and confirmation before anything counts. The first keeps it cheap; the second is what makes a memory of your day acceptable to have.