InDraft

What it does

You fork the repo, fill in config.yml with your profile paragraph, RSS feeds, posting schedule, and content pillars, and yarn setup deploys it on Vercel. From then on, a daily cron fires at 14:00 UTC, filters down to your configured days and local hour, pulls a fresh item from your sources, drafts a post in your voice, and emails you a magic link.

You open the review page on your phone. You can chat-edit the draft until you’re happy with it, then tap “Post” — Face ID or Touch ID, and it’s live on LinkedIn.

Why I built it

LinkedIn post quality is a function of voice and timeliness. Most automation tools kill both: they generate generic AI prose and they post on a schedule that has nothing to do with what’s actually happening that day. InDraft tries to preserve them — your voice is the model’s primary grounding, the schedule is keyed off fresh sources, and a human is always in the loop before anything publishes.

It’s also a generic, config-driven engine, not a SaaS. Nothing personal — voice, sources, credentials — lives in the repo. Fork it, point your config at your stuff, deploy.

How it works

Architecture

Vercel Cron (daily 14:00 UTC)
   | day/hour filter
Collector: RSS + GitHub
   |
Generator + anti-AI-tell linter
   |
Media: Pexels or owner upload
   |
Vercel KV
   |
Resend: magic-link email to /review
   |
Conversational chat edits
   |
WebAuthn passkey assertion
   |
LinkedIn Posts API

The state machine in src/lib/state/drafts.ts is the single writer of draft status and the load-bearing safety boundary. The only way to reach the LinkedIn Posts API is through a verified WebAuthn assertion — no scheduled or automatic path posts. An e2e test in tests/e2e/publish-guard.test.ts proves this at the state-machine level: a non-PENDING_REVIEW draft can never transition to PUBLISHED, and a captured assertion can’t be replayed against an edited draft because the challenge is bound to the draft version.

Key defaults

• link_placement: "none" — current LinkedIn-algorithm research shows external links suppress reach. InDraft surfaces the source URL inside the review UI instead. Override per post if you want.
• Frontier model for drafting, cheap model for ranking — Claude Opus 4.7 drafts because the anti-AI-tell bar degrades on weaker models; Claude Haiku 4.5 handles source ranking and mechanical tasks. Both routed through OpenRouter BYOK.
• 60-day LinkedIn token, no refresh — self-serve LinkedIn apps issue 60-day tokens with no refresh path. InDraft stores issued_at in KV, emails a reauth ping ~7 days before expiry, and shows a “reconnect LinkedIn” banner in the review UI.

Security model

• The only publish path is a click, then a WebAuthn assertion, then a server transition with a derived publishProof. The assertion challenge is sha256(draft.id || draft.version || draft.body), so a captured assertion can’t be replayed.
• Magic links are HMAC-signed, single-use (nonce stored in KV with a TTL), and expire after 24 hours by default.
• The public /access endpoint always emails the configured NOTIFY_TO_ADDRESS only — never an attacker-supplied target.
• Sessions are HttpOnly, Secure, SameSite=Lax cookies bound to a specific draft id.
• .gitignore excludes config.yml, .env*, and .vercel/. The setup script flows secrets prompt -> vercel env add only — never to disk.

Status

Engine is feature-complete and MIT-licensed. The repo is the deliverable; my personal deployment lives at indraft.dev.