Capture
Capture without classification: how AI can file your thoughts in the background
Capture is the floor. Anything that asks the user to classify before writing has already failed. Here is how a background AI replaces the tag picker, the database, and the triage ritual.
The core principle behind Phren is this: capture is the floor. If capture is unavailable for a heartbeat, the product has failed in that moment. Every other surface is downstream of getting the thing in your head out, now, before it is gone.
Every productivity app says this. Very few mean it. The test is simple: how many decisions does the user have to make before the sentence is written down?
The decision count test
Open Notion. To capture a thought, you decide: which database, which template, which view, which parent page, which tags. Five decisions. The thought is gone.
Open Todoist. Decide: which project, which labels, which due date, which priority. Four decisions. The thought is, optimistically, still around.
Open Phren. Open the bar — one keystroke. Type the sentence. Press return. Zero decisions. The thought is captured.
What replaces the decisions
An AI doing the same work in the background, after capture, with the user out of the loop. This is not novel — researchers have been classifying text for decades. What is new is the quality of small, fast language models that can route a sentence to the right place with high enough accuracy to remove the user from the workflow.
The architecture in Phren looks like this. A capture lands. A background job classifies it: thought, task, note, goal candidate, saved resource. A second pass attaches it to a corner — Work, Health, Finance, etc. A third pass tries to attach it to an existing goal, if one fits. All three happen in under a second.
The user sees an outcome. A soft chip, a small move into a list, a quiet "converted to task" note in the daily view. The process is invisible.
When the AI gets it wrong
It will, sometimes. A line about a book might be classified as a task instead of a saved resource. A worry might be misread as a goal candidate. The honest answer is the user can reclassify in one tap, and the system uses that signal to do better next time.
The dishonest answer would have been to keep a tag picker around as a safety net. Phren did not do that, because the cost of the safety net is paid on every capture, while the cost of a misclassification is paid only on the few that go wrong.
The voice of the AI
The AI never narrates. It does not show a thinking indicator. It does not present suggestion chips on every item. It does not summarize at the top of the screen. Its work is visible only as outcomes — never as process.
Most AI productivity apps are loud. They want the user to feel the AI working. Phren chose the opposite. The AI is a quiet hand on the back, never a voice in the room.
Voice capture
The capture path is one keystroke for text. For voice, it is a long-press on the same bar. Transcription happens with a small Whisper-class model, the same classifier runs on the result, and the captured thought lands as if it had been typed.
This matters specifically for thoughts that arrive mid-walk, in the car, mid-conversation. The mouth is faster than the hands. The capture path has to be ready for both.
What this unlocks
- Capture happens regardless of context. Standing, walking, mid-meeting, half-awake.
- The inbox stops being a wall of unfiled items.
- Goals attract relevant captures automatically.
- Tasks appear when a capture is clearly a task — no separate "add task" workflow.
- Notes accumulate in corners without ever being tagged.
- The user's relationship with the app shifts from filer to thinker.
The thesis
AI in productivity software has been mostly about generating things — drafts, summaries, suggestions. Phren bets on a quieter use: classification, routing, attachment. The AI does the work the user was being asked to do, and stops asking.
That is the floor. Everything else is downstream.