Personal assistant triage

I capture by texting myself — ideas, tasks, links, decisions, photos. It all goes into the same place, which is a lot of variety to ingest. The habit works because it's frictionless, but the pile was ever-growing. I needed a system.

So the sorting is automated and the judgment stays mine. And I use deterministic rules. A LOT of rules, so it isn't AI reasoning its way through complex random bits constantly.

The loop

Capture lands via Telegram at any hour — moving off plain self-texting was the one real change, and it's what made nightly backup possible. An hourly job imports it, applies any commands I left inline, and runs a conservative auto-triage. At 2:30am a batch of 150 goes through LLM triage, then a decay pass, then a research pass. At 3:45am a separate run fetches any URLs, transcribes video, and summarizes. At 7:00am I get a morning brief.

Two lanes, and this is the load-bearing decision

Content — links, videos, posts — is never a task. It routes to a wiki that transcribes and summarizes it, and comes back as a reading queue I choose from. Everything else is a commitment, and gets a stage.

Conflating those two is why every read-it-later app becomes an "ohh, I meant to read that" pile.

Some of the rules

Stages are to_do_asap, decide, delegate, schedule, next_up, backlog, in_progress, insight, archive, done — and needs_review.

needs_review is the one that matters. Low-confidence classification doesn't get a guess; it gets surfaced in the morning brief for me to resolve. Same principle as unknown in FoundMoney.

Importance and urgency stay separate fields, not one blended priority score. Blending them is how everything becomes urgent. It's not quite an Eisenhower Matrix, but that's the inspiration.

Action status is independent of item stage. A thing can be fully triaged and still not started. Most systems collapse these and then can't tell the difference between "I haven't decided" and "I decided and haven't moved."

Things decay. To-dos aged past 180 days get deferred or archived instead of sitting there. A list that never forgets is a list you stop reading.

Original text is preserved verbatim and never paraphrased. Anything extracted from a photo is stored separately from what I actually wrote. Once you let the summary overwrite the source, you can't audit a bad classification later.