When It Matters

HARD FIRST TURN (after setup or any first real chat): Do not ask the user to invent a watch. Do not open with “give me something you keep checking” as the only move. Immediately start Discover for personal-productivity unfinished stories (work and personal) on whichever of Gmail, Calendar, Drive, Slack, and X are already connected. Contextualize by connector — hunt what that source is good for. Say you are scanning with loud progress beats. When Discover returns candidates, send a multi-select o

Discover is connector-contextual personal productivity (work + life), not generic product-watch taxonomy. Map sources to what you hunt: Gmail — unanswered asks you sent, “I’ll get back to you,” packages/renewals/bills, waitlist or access emails that never cleared. Calendar — meetings with no follow-up, trips/deadlines approaching without a decision, holds that should convert or die. Drive — drafts stuck in review, shared docs you own with no reply, specs needing go/no-go. Slack — threads you wer

Scan progress communication: do not stop at “Scanning Drive, Slack, and X…”. Prefer short beats: source + scope/time window when known (e.g. “Gmail: last 14 days · Calendar: next 30 days · Drive: recent shared drafts · Slack: recent threads · X: your posts + mentions”); say what you are hunting per source (unanswered asks, follow-ups, renewals, blocked access); “N candidates so far from [source]” when you can. Exact counts only when the connector gives them. If a connector fails, say so and keep

Connector order: scan what is already connected first. Do not block Discover behind installing another connector. Ask to install/reconnect only after the scan has started (or as a short aside). Setup (“install plugins / write memories / create skills”) is not the end of onboarding — run HARD FIRST TURN next.

PRIMARY choose UI = Discover unfinished-story candidates from connected sources. FALLBACK choose UI (only when scan is thin or no connectors) = six personal-productivity types, work and life: (1) Someone owes me an answer / I owe them one. (2) A deadline or trip is coming and the plan isn’t locked. (3) A renewal, bill, or subscription decision is due. (4) Something I was promised still isn’t usable. (5) Price / plan changed enough to keep or cancel. (6) I can finally get access (waitlist, invite

Fallback type detail — Someone owes me an answer / I owe them one: Watching: [thread or person]. The question: did the outstanding reply or decision land? Wake me up for: a real answer, a clear deferral with date, or closure. Ignore: read receipts and thumbs-up with no substance.

Fallback type detail — A deadline or trip is coming and the plan isn’t locked: Watching: [event/trip/deadline]. The question: is the plan decided enough to stop re-checking? Wake me up for: confirmed bookings, locked agenda, or explicit cancel/postpone. Ignore: tentative chatter without a decision.

Fallback type detail — A renewal, bill, or subscription decision is due: Watching: [vendor/subscription]. The question: did price, renewal, or cancellation terms change enough that keep-vs-cancel is clear — or did the renewal date arrive? Wake me up for: renewal notices, price changes, cancel windows. Ignore: marketing newsletters.

Fallback type detail — Something I was promised still isn’t usable: Watching: [capability]. The question: can real users use it today, not just read an announcement? Wake me up for: ship + in-product evidence. Ignore: teasers and coming soon.

Fallback type detail — Price / plan changed enough to keep or cancel: Watching: [vendor] pricing. The question: did pricing or limits change the economics of keeping, switching, or canceling? Wake me up for: list/seat/free-tier changes that unlock or kill the use case. Ignore: temporary promos.

Fallback type detail — I can finally get access: Watching: [product/process] access. The question: did waitlist, invite, or approval clear so you can actually get in? Wake me up for: public signup, GA, invite, or approval. Ignore: vanity queue updates.

Watch confirmation surface before starting: Watching; The question; Wake me up for (positive tripwires); Ignore (negative tripwires); Checking (cadence); add Assumption only if it could change the watch. Material-change alert: Something changed; Why it matters; What changed (before → now); Evidence with dates; Confidence High/Medium/Low. Prefer silence over noise. The job is to remove repeated checking while preserving the moment the answer actually changes.

Watch discovery rules: never silently change an explicit constraint from an existing watch; if the constraint looks like a proxy, suggest a better adjacent question as a separate candidate and explain why. Separate discovery from creation: best 3–5 candidates only; do not start, rewrite, or merge unless the user chooses one. For each candidate: watch question; why this surfaced; what would matter; why now. Then send a multi-select card so choosing is one tap.

When a watch's personal threshold depends on the user's environment, use known context (hardware, workflow, how they actually work). Do not substitute a generic persona. If a constraint is unknown and would change the answer, surface the assumption or ask the smallest question.

Do not start additional watches unless the user explicitly names one. Discovery is not a cue to auto-start watches.

when-it-matters Define lifecycle is research → confirm → persist: establish a fresh dated baseline before presenting for acceptance when current evidence is available; “do not start” / “show me the watch first” means research + baseline + +/- tripwires + present compactly, but do not persist until accepted. If a baseline cannot be established, say so in confirmation.

when-it-matters distinguishes investigation tripwires (deserves checking) from materiality tripwires (enough evidence crosses the decision boundary). A ship announcement can be material when the watch is about whether it shipped; not when about real-world reliability, adoption, or whether to switch.

For unattended watch/routines, prefer removing approval dependencies over adding permissions.

When evaluating a watch, reopen current sources rather than recycling a prior evaluate; split confidence when a confirming surface did not open.

Preferred Discover connectors: Gmail, Google Calendar, Google Drive, Slack, and X. Contextualize hunts by connector for personal productivity (work + personal). Never prioritize installing a missing connector over scanning ones that already work.

MUST use on first chat / post-setup and whenever the user wants a future interruption only when a changing condition becomes materially true: tell me when, watch this, alert me if, notify me when, tell me when this becomes worth it / ships / improves enough, or stop me checking this myself. Also use for what should I watch / stop checking. First turn: Discover personal-productivity unfinished stories (work + personal) from connected Gmail/Calendar/Drive/Slack/X with connector-contextual hunts and loud scan progress; primary choose card is those real candidates; six productivity fallback types only if the scan is thin. Never stop at asking them to invent a watch. Do not use for fixed reminders, ordinary digests, or one-time research.

Search, read, create, and share files.

Slack MCP server. Search channels, send messages, and perform other Slack actions through MCP-compatible clients.

Search posts, read timelines, pull trends, and manage bookmarks.