Later

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 social scheduling that replaces Later — and read the obstacle first, because it is not code.

**Publishing on somebody's behalf needs an approved application on each platform.** Several require a review, some restrict it to business accounts, some charge for the tier that allows it, and permissions get revoked when policies change. Check, today, for the platform you actually care about, whether an independent developer can still get it. If the answer is no, no amount of good software succeeds. That is the whole verdict.

STACK
- Node 20+ with Fastify, server-rendered HTML
- SQLite through better-sqlite3, WAL mode
- A worker for publishing, separate from the web process
- Caddy in front

THE DATA MODEL
- channels: id, platform, handle, display_name, avatar_path, credentials_encrypted, token_expires_at, refreshed_at, is_active, limits_json
- posts: id, body, media_ids_json, link, first_comment, created_by, created_at, status, approval_state
- targets: id, post_id, channel_id, scheduled_for, status, attempts, client_key, external_id, external_url, published_at, error, platform_body — one row per channel, because the same words are not the same post everywhere
- media: id, path, sha256, kind, width, height, duration_ms, alt_text, bytes
- slots: id, channel_id, weekday, time_local, position
- metrics: id, target_id, kind, value, collected_at — append-only, never overwritten
- Every instant UTC, with each channel's own zone kept for display

THE GRID PREVIEW, WHICH IS THE FEATURE PEOPLE CHOOSE IT FOR
- A profile rendered as it will look: the next nine or twelve posts in a grid, in order, including what is already published
- Drag to reorder within the grid, which reorders the schedule
- Crops shown as the platform will crop them, per platform, because a square preview of a portrait image is a lie
- This is design work more than engineering, and it is what makes the tool feel like a planning surface rather than a queue

PUBLISHING WITHOUT DUPLICATES
- The worker claims a target with a lease, publishes, and writes the external identifier in the same transaction as the status change
- Idempotency: a platform that times out may still have published. Before retrying, look for the post by a client key where supported, otherwise by searching the recent timeline. A duplicate post is a public embarrassment
- Retries with backoff for transient errors only, never for a validation failure
- Rate limits per channel, honouring the platform's own headers
- A failed target alerts by email with the platform's message verbatim

LIMITS AS DATA
- Character count and how it is counted, media count, accepted types, aspect ratios, file size, video length and codec, whether threads or a first comment are supported, whether alt text is accepted
- All declared per platform in one table and validated at scheduling time, not at publish time. Finding out at nine in the morning that a post is four characters too long is exactly what this must prevent
- Alt text on every image, with an option to refuse scheduling without it

THE CALENDAR
- Slots per channel per weekday; adding to the queue fills the next free one
- Month calendar across channels, a list of the next fortnight, drag to move, pause a channel, holiday mode
- Best-time suggestions computed from your own past engagement and labelled as what they are: a weak signal from a small sample, not a science

APPROVALS AND METRICS
- Draft, needs approval, approved, scheduled, published, failed. Nothing publishes from needs-approval by any path
- Metrics collected on a schedule after publishing and appended, absent where a platform does not provide them and shown as absent rather than zero
- Link clicks through your own short domain, which is the one number nobody can take away

OPERATIONS
- .env: DATABASE_PATH, STORAGE_PATH, BASE_URL, ENCRYPTION_KEY, SESSION_SECRET, SMTP_URL, per-platform client ids and secrets
- Migrations on boot, each once; tokens refreshed under a lock
- Nightly backup off the machine, restore script
- Health endpoint reporting the oldest due target and any expiring token

WHAT MATTERS MOST
The application review, then idempotent publishing. Confirm you can actually get permission before building the scheduler, then pull the plug on the worker mid-publish and prove you get one post rather than none or two.