Creating your first AI-powered block

What an AI block is

Where a manual block needs you to list every distractor by hand, an AI block takes a single English-language task context — "writing my essay", "coding the checkout endpoint", "reviewing the Q2 deck" — and the classifier regularly checks whether what's on your screen still matches. If it doesn't, the AI intervenes. You skip the busywork of curating a 40-domain blocklist and trust the model to make the call.

Step-by-step

Open FocusDragon and click "New Block" in the sidebar.
Give it a name ("Essay sprint" works) and skip the Blocked Items section entirely — you don't need to list sites or apps for an AI block.
Find the "Task context" field and type one to two sentences describing what on-task looks like right now. Specific beats vague: "writing the methods section of my thesis in Google Docs, with one tab of Scholar for citations" classifies better than "studying".
Pick an intervention mode (see below).
Pick a lock type if you want to commit to the session (Timer is the usual starting point; No Lock works too).
Click Create Block, then hit Start.

Nudge or Hard

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 whitelists the surface for the rest of the session. The right choice for almost every block.
Hard — force-quit on sight. Every drift kills the offending app immediately and registers a 5-minute cooldown. Relaunching during the cooldown re-kills and extends. No override. Use only when you genuinely cannot afford to drift.

Nudge is the default because it's the balanced point: real intervention when you actually drift, with one warning and a one-click "I was on-task" escape hatch for false positives. Move to Hard only after you trust the classifier's calls on your typical work.

The pre-flight check

When you hit Start, FocusDragon runs a pre-flight modal that refuses to launch the block if the AI prerequisites aren't met. The two it cares about are Screen Recording (so the classifier can see your display) and Accessibility (so it can read window titles on native apps and force-quit them on drift). If either is missing, the modal deep-links you straight to the correct System Settings pane — grant access, come back, hit Start again.

Don't try to work around the pre-flight modal. An AI block that starts without Screen Recording is silently blind — it'll never trigger an intervention, and you'll waste a session thinking the AI is broken when it just couldn't see anything.

What you'll see while it runs

On each check the classifier returns one of three verdicts: focused, distracted, or unclear. Focused is silent. Distracted triggers your chosen mode. Unclear (the model isn't confident either way — common with split screens or ambiguous content) is logged for the session report.

Iterating on your task context

If the AI keeps flagging legitimate work as distracted, your task context is probably too narrow. If it lets through obvious drift, it's too broad. Edit the block, refine the wording, restart. Most people find their sweet spot inside three sessions.

For a deeper guide to writing task contexts that the classifier reads well, see the AI task context doc.