AI Focus — FocusDragon Docs

What AI Focus does

When you start a block with a stated task ("writing my essay", "coding the API"), AI Focus watches your main display and intervenes the moment you drift off-task. It's not a content filter — it's a classifier reading your actual screen, so it works on sites and apps you never thought to add to a blocklist.

The watch loop

At regular intervals, FocusDragon captures a downscaled, compressed screenshot of your main display and pairs it with a short text snippet (browser DOM excerpt or app window title).
That bundle plus your task context goes to FocusDragon AI, which returns one of three states: focused, distracted, or unclear.
If the verdict is a high-confidence distracted, the intervention you picked for that block fires.

The model is explicitly instructed to prefer "unclear" over a guess. False-positive force-quits destroy trust faster than any other bug, so borderline calls cost you nothing.

Two intervention modes (set per-block)

Nudge — the default for new blocks. One 5-second warning the first time you drift, then a force-quit of the offending app or close of the offending tab if you keep drifting. An "I was on-task" button lets you whitelist the surface for the rest of the session. Best for most blocks.
Hard — force-quit on sight. Every drift kills the offending app immediately and registers a 5-minute cooldown. Reopening the app during the cooldown re-kills it and extends the cooldown further. No override. Best for deep work when you genuinely cannot afford to drift.

AI blocks always actively enforce — there's no log-only mode.

Task context

Your task is set per-block on the block editor (up to 500 characters) — a block's task is block-scoped, not user-scoped, so each AI block carries its own intent. Empty task = no classifier runs and the block falls back to pure rule-based blocking.

Privacy

The screenshot and text snippet are sent to FocusDragon AI via our server proxy, classified in a single call, and immediately discarded. They are never written to disk on your Mac. FocusDragon never stores them on its own servers. Under the hood, FocusDragon AI routes through third-party model providers — they do not train on these inputs and retain them for up to 30 days for abuse monitoring before permanently deleting them. The provider may rotate over time as we improve the classifier; the no-training, short-retention contract is the load-bearing privacy guarantee, not the specific provider.

Capture only runs while a block with a non-empty task is active — the loop is torn down the moment the block ends. Banking, email, identity, healthcare, and incognito tabs are hard-coded to exclude their text snippet; only the URL and tab title flow alongside the screenshot.

Graceful degradation (the 100h/month allowance for AI Focus, 150h for Max)

AI Focus includes up to 100 hours of AI-watched focus per month (AI Focus Max raises this to 150). The AI assistant automatically slows its polling rate as you approach the monthly allowance, and continues at a reduced cadence past it — the classifier never stops working when you exceed the cap.

Most users never see degradation. The curve exists so the worst-case heavy user is still served — just with slower reaction time — instead of getting cut off mid-month.

A Tier-1 rule layer (TikTok FYP, IG Reels, YouTube Shorts) hard-blocks before the classifier runs, costs no tokens, and is on by default. You can opt out of specific surfaces if you want them routed through the classifier instead.