Sage

Sage is your single point of contact in Zimac — your chief of staff. You talk to her about anything: a note to capture, a question for the team, a meeting to prep, a plan to build. She figures out who should handle it, does the work (herself or through a specialist), and comes back with one clear answer. You never manage the team; you talk to Sage.

Who Sage is#

Sage leads a small staff of AI specialists that support you the way a chief of staff supports an executive. Her job is not to know everything herself — it's to make sure the right expert handles the right problem and that you get back a single, decided answer instead of a menu of options.

She operates on a few principles that shape every reply:

• She triages first. Every message gets sorted: is this something to remember, a question about the past, a person to understand, a plan to build, or work for a specialist? She handles each accordingly.
• She synthesises, she doesn't relay. When a specialist contributes, she tells you who and what — "Cassandra flagged the missing rollback plan" — then gives her own take. You feel like you talked to a team lead, not a switchboard.
• She owns every failure. If a lookup comes up empty or a specialist hits a wall, that's hers to work around — a different search, a smaller ask — not a problem handed back to you.
• She checks before she asks. Before asking you "who is X?", she searches the team's memory and people profiles. The team already knows a lot; she uses it.
• She's calm, warm and decisive. Brief confirmations when she logs something. A nudge when you're about to miss something — she sees the whole board.

Talking to Sage#

There's no syntax to learn. Type the way you'd message a trusted colleague. A few things are worth knowing because they make her noticeably more useful:

Fire off several thoughts at once

If you dump three things in quick succession — a deadline that moved, a question about a teammate, a note to file — Sage triages each one separately and comes back with a single organised reply that addresses all of them in order. Nothing gets dropped.

Paste screenshots

Sage reads images directly — a Slack thread, a dashboard, a diagram, an error message. Paste it and talk about it. When she routes it to a specialist (say, a Slack screenshot to understand a person), the image goes with the hand-off automatically.

How replies show their work#

While a reply forms, Context Bloom can show the recall paths, tool activity, and specialist consultations the host actually observed for that turn. It compacts when the prose starts and then disappears. It never invents a source and is not a display of hidden chain-of-thought.

After a substantive reply settles, an expandable Decision trace can group what was available as Observed records or material read now, Recalled memory admitted to the answer, Inferred synthesis or teammate contributions, and Unknown checks that could not be completed. A long answer with no live or recalled evidence says so. This is a turn-level evidence ledger — not private reasoning and not a claim-by-claim proof that every sentence follows from a listed item.

The response menu keeps two separate tools honest: References in answer lists links and context named in the prose without upgrading them to verified provenance, while Focus sections temporarily narrows a multi-section answer without changing its saved content. Copy as formats the whole reply for plain text, Slack mrkdwn, Jira wiki markup, or an email draft; email opens your mail app and also copies the full text because mail links can clip long bodies.

Your morning brief#

Ask Sage to start your day and she compiles a ranked rundown of the few things that actually need you today — pulled from your deadline board (upcoming and overdue), the recurring rhythms that are due, any projects flagged at-risk, and today's calendar load. It renders as a card with a next-step button on each item, and it's saved to the Brief panel too.

The optional Evening wrap closes the loop at the other end of the day. After the hour you choose, it locally summarizes meetings attended, memories written, open commitments, pending 1:1 debriefs, and due rhythms, then writes one “Tomorrow starts with…” line into the next Morning Brief. It is deterministic, spends no model tokens, runs at most once per local day, and is off by default; enable it or run it now from the Brief panel.

Memory & recall#

Anything you tell Sage worth keeping — meeting notes, a decision, "the deadline moved", action items, names — becomes part of the team's shared memory. This is the backbone of everything else: it's why she can prep a meeting, run a progress report, or answer "what did we decide about X" weeks later.

You just talk; she captures. Recording and reorganising memory is owned by Quill, the team's notetaker, so Sage hands the raw words to Quill (verbatim, so no date or name is lost) — but from your side it's seamless. For quick lookups she reads memory herself.

People & your voice#

Sage keeps a grounded relationship record for people in your work life — sourced identity details, observations, open commitments, meeting history, and clearly labeled hypotheses — plus a profile of your own writing voice. Those records are owned and written by Joy, the people-insight specialist; Sage reads them to do two things especially well:

• Draft replies that sound like you. Ask her to help respond to someone and she writes it in your voice by default, tailored to who it's going to — and tells you the one or two things about that person that shaped the draft.
• Understand someone new. Paste a Slack conversation or tell her about your new skip-level, and that material goes to Joy to build a profile. Joy also learns your voice from the messages you wrote in those threads.

The specialists she routes to#

Sage's superpower is knowing who should handle what. You never have to. When a message lands in a specialist's lane, she hands it over — they do the work with their own tools and expertise — and she folds their answer into her reply. Here's the roster she draws on. Each specialist has a dedicated guide in these docs; this is what Sage routes to them.

You don't need to remember any of this. Ask a database question and Cassandra answers through Sage; ask for design options and Drew can start from the brief, screenshot, or connected Figma file. When a question spans two lanes, Sage gets both views and reconciles them. And if the right specialist happens to be turned off, she still gives you her best answer and offers a one-click chip to switch them on.

Calendar & email#

Sage can read your real calendar and mailbox from the sources you connect. Calendar merges Apple Calendar, Google Calendar, Outlook, and subscribed or imported ICS feeds into one deduplicated schedule. Mail Studio reads Apple Mail on Mac and imported .eml or .mbox files on Mac or Windows; the Microsoft 365 connection gives Sage a separate read-only path to Outlook.

• Calendar — "what's on today", "when's my next 1:1 with Maya", "am I free Thursday afternoon". This is your lived schedule, distinct from your saved-deadline board.
• Mail Studio — "did Maya reply about the migration", "any unread from my VP", "find the thread about the Q3 budget". For Apple Mail and imported-message results, Sage can open the full body and cite a link that jumps into Mail Studio.

On Microsoft 365? Sign in once and Sage can search Outlook, open a message's full body, and pull your Outlook calendar for the week — read-only, as you, seeing only what you can see. Outlook results keep their provider link and are not copied into Mail Studio.

Scheduling & meeting prep#

Prep me for a meeting

Ask Sage to get you ready for a meeting — especially a 1:1 — and in one shot she pulls the event from your calendar, the counterpart's profile (how to work with them, their comms style), what memory recalls about them, and the open commitments involving them. She hands back a tight, grounded pre-meeting brief: who you're meeting and when, the one or two things that matter about this person, the open threads to raise or close, anything time-sensitive.

Put time on the calendar

Ask her to schedule something and she drafts an editable event proposal. She titles it clearly, finds open time against the merged schedule, flags conflicts, and gives you a human-click handoff to Google Calendar, Outlook, or an .ics file that opens in Apple Calendar or another calendar app. Zimac never silently creates an event. On Mac, it can also prepare an exact before-and-after edit for an existing event from a directly connected Apple Calendar; nothing changes until you click Update event. You can target one occurrence or this-and-future, but attendee changes still belong in Calendar.

Capture the one line after a 1:1

When a 1:1 with someone on file has just ended, Zimac can place a one-line debrief prompt in Quick Ask. Answering files that line to the person's profile; dismissing costs nothing, and the same meeting is not offered twice. The check is local and zero-token, and if you rolled straight into another meeting the delivery waits for a gap.

Daily-flow assists#

On macOS, Clipboard Smart hints use local, zero-token recognizers for a fresh stack trace, pull-request or issue URL, tracker key, meeting link, or IP address. If Clipboard history and Smart hints are enabled, the newest confident match from the last 15 minutes appears as ready-made Quick Ask context. It never overrides selected text or a query you typed, and concealed password-manager items are never recorded.

Zimac also governs when deferrable notices interrupt you. By default it holds them while a selected-calendar meeting is running, when the next meeting is too close, or when macOS Focus is detectable and enabled, then releases or batches them in the next real gap. Notices older than six hours expire. A time-bound approval, imminent meeting prep, or something you are actively waiting on bypasses that queue.

Projects & delivery tracking#

Sage doubles as a lightweight TPM. Your initiatives live on a timeline board (the Projects studio), and she manages them from chat. Tell her something is a project — or she'll recognise a coherent, multi-step initiative with deadlines and offer to track it as one.

Beyond just creating projects, she runs the full delivery surface:

• Targets — commit a delivery date for the whole initiative.
• Tasks — create a dated deliverable and pin it to a project so it lands on the timeline.
• Risks (RAID-style) — log a risk with a severity; a high one flips the project off-track. Resolve or drop them as they close.
• Status — set the narrative status line.
• Briefs & reports — a single-project digest (with a paste-ready narrative), or a portfolio read: a red/amber/green roll-up, the at-risk list, upcoming deadlines, open risks, and per-person load.

Staffing plans & roadmaps#

Sage builds staffing gantts and roadmaps for you — real, editable plans that render as a card in the chat and open in the Gantt studio for dragging around. You feed her resources (the people) and work (the work-streams); the layout engine does the hard part:

• No one is ever double-booked. On-call and vacation are hard commitments that hold their weeks; project work automatically flows around them.
• Auto-scheduling. Leave a work-stream's start open and the engine drops it into the earliest free slot on that person's row.
• Critical path. Give a work-stream dependencies and it's sequenced finish-to-start — the engine computes the critical path so you never hand-sequence a roadmap.
• Analysis, free. Ask "who's the bottleneck", "when does this finish", "who has room" and she reads per-person utilisation, the end date, the critical path, and risk warnings — no tokens, no changes.

A staffing plan can become a tracked project in one step (each work-stream's end-of-week becomes a milestone), and the two stay linked.

Sprint planning#

Tell Sage you're planning a sprint for a team and she pulls the board's last closed sprint from Jira — shipped vs carried, per person, story points where the board estimates — and hands back ready-to-paste chart and carry-over cards plus a grounded draft of the retro and next-sprint plan. Every number and ticket key comes straight from Jira; she never invents them.

She follows your planning flow. Share your "how I run sprint planning" Confluence page and she distills it into a saved runbook for that team, so future plans follow your own steps — and each team can have its own flow. When the draft is ready she can offer, once, to publish it to Confluence (only ever on your explicit yes, and the page isn't live until you click).

Progress reports#

Ask "where are we on X" and Sage compiles a progress report over a time window: your own notes and updates from memory, the projects board, the related Jira movement (via Cody), and — when the work involves code — the relevant GitHub activity. She synthesises what actually moved, not what was hoped.

Contribution deep dives#

When you want to know whether someone's work is good — not just how much of it there is — Sage runs a contribution deep dive. It pairs volume (Joy's DX metrics) with substance (reading the actual work evidence on file — documents, notes, saved artifacts) and returns an honest quantity-plus-quality read, flagging where the evidence is thin.

Documents & commissions#

Sage reads from your document library and drafts anything you need. For most writing — a Slack reply, a short note, an inline answer — she just writes it. For a substantial, rigorous deliverable she offers a heavier gear: a commission.

A commission compiles a document and iterates it to convergence: Sage drafts it, every relevant specialist reviews it round after round pulling real data from live systems (Cody → GitHub/metrics/Jira, Finn → AWS & cost, Quill → the knowledge base), and she revises until their blockers are gone. It runs in the background (5–15 minutes), is capped at about $10 of tokens, and is opt-in — she'll offer it and wait for your yes, never launch a costly job you didn't ask for. The finished document surfaces automatically as a card when it's ready.

The wiki (Confluence)#

Sage works with your team's Confluence wiki (it uses the same Atlassian connection as Jira):

• Search & read — "what does the wiki say about our on-call policy" finds the right pages and reads the best match.
• Scour — have the team absorb a whole body of knowledge: it reads each page, distills the durable facts, and saves them to memory, so the team remembers it going forward.
• Publish & edit — hand over a draft as a confirmable "Publish" card, or open an existing page in a live markdown editor right in the chat. Nothing is written to the wiki until you click — the human click is always the gate.

Your files & drives#

Sage can look through files you own — locally on your Mac and in mounted sources like your Google Drive — all read-only. On Microsoft 365, OneDrive and SharePoint search runs through Quill on the same sign-in.

• Local files — list what's in a folder like ~/Downloads, then read one by name or path (text, code, PDF, Word).
• Mounted folders & Drive — search connected sources for a document and open it to quote accurately. (Native Google Docs match by title and come back as a link, since Drive stores only the link locally; Office and PDF files are full-text searchable.)

Org sites#

When you want to share a list for your org to review or fill in — confirming AWS tag owners, collecting sign-offs, any tracking-sheet replacement — Sage builds an org site: a single page with sections and a data table, published to your whole organisation. Teammates open it in a browser and confirm rows, fix values, or add rows; every action is attributed to their verified identity and syncs back into the app, so your local copy stays the source of truth. She builds the table from real data (yours or a specialist's pull), never invented rows, and you click Publish — she can't publish for you.

Sharing across your org#

Zimac is local-first and private by default, so sharing is deliberate and never leaks your raw notes. Sage gives you a few clean ways to reach your teammates:

• Shareable insight — send a single aggregate conclusion ("Project A is at-risk") to a teammate or the team brain, without exposing the private notes behind it. Only the distilled statement leaves your device; a private "distilled from N of your memories" disclosure is visible to you alone.
• Team broadcast — a directive or announcement to your whole organisation. Every teammate's device gets it (offline ones when they return), each one's Sage synthesises it for them, and they accept or decline — you watch who accepted in the Team hub.
• Ask teammates — pose a question Sage-to-Sage across the team ("who has the Q2 deck?"); each recipient's Sage drafts them an answer from their own memory and they choose whether to reply. Optionally anonymous.
• Trace a connection — "how is this project connected to Priya?" walks the team's knowledge graph and shows the shortest meaningful chain as a visual.

Night Shift watches#

Tell Sage to keep an eye on something and she puts it on your Night Shift watchlist — the team re-checks it on a schedule and alerts you only when it materially changes: a ticket's status moves, a metric crosses a threshold, an item is added or removed. Watchable things include Jira tickets and searches, sprint progress, and Chronosphere metrics. Adding a watch runs it once immediately as its baseline.

Slack DMs#

With the Slack integration connected, Sage can catch you up on your direct messages — list recent DMs, read one conversation in full, and draft a reply. As everywhere, sending is a confirmable card you click; she drafts, you send.

Microsoft Teams chats#

The Teams sibling of her Slack powers, on the same Microsoft 365 sign-in: Sage lists your recent 1:1 and group chats and reads one conversation in full — read-only. There's no send path yet; when you want to reply, she drafts the text for you to paste (a click-to-send card, like Slack's, is on the roadmap). Chat text is treated strictly as data — she'll summarize what someone wrote, but nothing in a message can steer what she does.

The Frame & Sight#

On macOS, Stick the Frame lets Zimac recede to a thin, click-through border over each display while you work in other apps. A separate edge tab stays interactive and follows you across Spaces. Its tray brings together unread Volition thoughts, Night Shift findings, and Cadence nudges; open a card to read its full text, jump back to the main window, mark a thought helpful or unhelpful, or snooze and dismiss supported nudges. The neutral Frame does not observe the screen.

Sight is the Frame's optional screen-reading leg and is macOS-only in the current release. It is off by default, must be enabled once, and must then be armed with an explicit Observe click for every session in addition to macOS Screen Recording permission. A session ends when you pause, unstick, quit, or capture fails. The amber border and “observing” label are driven by the same controller that gates capture.

The current Sight observes the primary display only, excludes Zimac's own windows and the cursor, waits for visible changes to settle, and runs Apple Vision OCR on device at a restrained cadence. Raw pixels are never persisted and never leave the Mac. Secret-redacted recognized text can pass through your configured model route for distillation, and accepted facts enter local memory with receipts and undo.

The exact-answer toolkit#

Behind the scenes Sage carries a registry of about 1,000 exact, pure-compute tools — hashing, base64/hex encoding, unit and timezone-free date math, statistics and percentiles, colour spaces and contrast ratios, CIDR/IP math, checksums, ID and token generation, JSON/CSV shaping, finance, geo, and more. When a request needs a precise computation, she runs the real tool and gives you the computed answer instead of estimating it. You don't invoke these directly — you just ask, and the number she returns is exact.

Skills & self-improvement#

Sage can teach herself new, reusable capabilities — and prove they're correct before she trusts them. This runs on Zi, Zimac's own proof-carrying language: a capability is only installed if a verifier proves it meets its written specification, every obligation discharged.

• Acquire a capability — when you need something exact and repeatable that no skill covers, she writes the spec and an implementation, and installs it only if the proof checks out. She then shows you the verified spec to confirm it matches what you meant.
• Use a capability — for anything a proven skill covers, she runs it for an exact, guaranteed-correct result rather than doing the math herself. Out-of-domain inputs are refused rather than answered wrong.
• Gem a skill — turn a complete, provable Zi module into a permanent skill in your Skill Vault.

Tone & lessons#

You can shape how the whole team communicates, and correct it when it gets something wrong — both stick across every conversation:

• Tone — "keep everybody brief", "I need snap decisions", "assume I'm new to this domain" sets a standing directive for Sage and every specialist, until you change it. Say "back to normal" to clear it.
• Lessons — "you forgot to…", "next time do X", "I prefer Z" are filed as durable feedback that's woven into the team's instructions from then on, so the behaviour actually changes. Sage acknowledges in one line and, when she can, redoes the thing correctly right away.

Jumping into studios#

Sage doesn't just describe where to go — she takes you there. When something is better handled in one of the app's studios (Review, Gantt plans, People, Sites, the Team hub, Settings to connect an integration), she surfaces a one-click "Open" chip, or jumps you straight there when you ask. So a plan she just built is one click from the timeline, and a draft is one click from the Review studio.