AI Coaching sidebar (interview prompts and artifacts) (part 5)
Part 5 of the AI Coaching sidebar page: new built-in interviews after an update, how interview sessions relate to regular sessions, the wrap-up banner when an interview finishes, Quick check-ins, the daily opening messages, Bookmarks and Goals.
On this page
What it is
This is part 5 of the AI Coaching sidebar (interview prompts and artifacts) page. It covers what happens around an interview: how new built-in interviews arrive, how an interview session sits among your regular sessions, what you see when one finishes, the short Quick check-ins, the opening message each day, Bookmarks and Goals.
Where to find it
Everything here lives in the AI Coaching sub-sidebar described on the parent page: the Dashboard (where Quick check-ins start), the Sessions and Bookmarks sections, and the Goals pane. The parent page says how to turn the feature on.
How it behaves
New built-in interviews after an update
When an update ships a new built-in interview, it appears on your Dashboard the next time Omniscio starts — however long ago you installed. Nothing you already have changes: your own copies are never touched, and a built-in you deleted in Super Prompts stays deleted. To bring a deleted built-in back, use the Undo on the "Deleted" notice right after deleting it, or ask your agent to call POST /saved-prompts/restore-hidden, which restores every hidden built-in at once.
How interviews relate to regular sessions
An interview session is a normal Omniscio session under the hood — it lives in the same sessions table, takes the same statuses (running, needs_you, paused, archived, etc.), respects the same archive/snooze/pause semantics, and shows up in global search like any other. The AI Coaching sub-sidebar is a view over those sessions, scoped to the AI Coaching virtual project. Archiving a session from anywhere in the app moves it into the Archived bucket here; un-archiving puts it back into one of the active groups.
Artifacts are stored separately — they're the durable output of an interview, not a copy of the transcript. Deleting an artifact (every version, for good) or rejecting one version (a soft delete) never touches the session that produced it, and archiving a session does not delete its artifacts.
What happens when an interview finishes
When the coach saves a new artifact, Omniscio fires a session-completed push event (AI_COACHING_SESSION_COMPLETED) carrying the session ID and the artifact title. The coaching panel listens for that event and shows an inline wrap-up banner at the top of the right pane:
- A celebration message confirming the artifact was saved, with the title shown.
- A suggested next step button — the next interview prompt from your coaching journey (for example "Start your Weekly Check-In" or "Continue your Career History"). If you've completed every available interview, the banner says so.
- A dismiss (×) button to close the banner without acting.
The banner appears once per session — if you dismiss it, it won't come back for that session. It sits above the session content, the same way the credential-alert banner does, and renders in all right-pane views (dashboard, artifact, session, etc.) so you see it no matter which section you're looking at when the artifact saves. On a phone, where the coaching panel closes while you are in the interview chat, the banner waits and appears when you come back to AI Coaching.
Auto-archive. Completed coaching sessions (those with a saved artifact and a terminal or needs-you status) are automatically archived after sitting idle for 10 minutes. This keeps the active session list clean without requiring you to manually archive every finished interview. The sweeper runs every 5 minutes and is restart-safe — it picks up where it left off if the app restarts. Kill switch: set AMC_DISABLE_COACHING_AUTO_ARCHIVE=1 in your environment.
Quick check-ins (snack sessions)
You start a quick check-in from the Dashboard's Quick Check-ins picker: short coaching sessions of a few minutes each, grouped by topic (journaling, reflection, motivation, challenge, mindfulness and more) with a search box. Click a topic to start it. Only one check-in runs at a time, and a topic appears once you have enough saved write-ups (and, for most, your Core Profile). Topics meant for the current part of the day (morning, evening and so on) are listed first, judged by the clock of the device you are using: on a phone or a browser away from home that is the device's own time zone, not the home computer's. A request that names no zone, such as the CLI's GET /ai-coaching/snack-templates, uses the computer's zone.
The Dashboard's Quick Check-ins picker also holds five check-ins that follow up an interview you've finished — Interview Reflection, Interview Challenge, Interview Momentum, Identity Check-In and Purpose Check-In. Each one picks up the saved document from one of its own interviews, starting with the one you haven't updated for the longest time, so repeated check-ins take turns through them, and what you work out together is saved as that document's next version. A follow-up check-in stays out of the picker until you've finished one of its interviews (where an interview has a 15-minute Quick version, finishing that counts too), and an interview you've removed from your list no longer counts. On a new account, before any check-in has unlocked, the picker says "Quick Check-ins unlock as you finish interviews." and adds that most of them also use your Core Profile, which is built once you have three saved write-ups.
Opening messages
Every interview prompt has two related fields:
openingMessage— a structured template that tells Claude how to start the conversation. It's used by the immediate-spawn paths (a start triggered through the CLI) and lands in the session's CLAUDE.md as a## Your First Responsesection. A redo (an earlier write-up exists, repeat check-ins included) gets a fixed coming-back opener in that section instead: no first-time welcome, a short recap of the previous version, and a question about what has changed.firstMessage— the authored greeting itself: literal, ready-to-display prose that opens the interview and asks the first question. Every greeting opens with a short##heading and a brief intro, then leads into the first question — authored as rich markdown (heading, bold lead-ins, bullets, a divider) so the opening bubble reads cleanly. This is what powers the instant, zero-cost greeting you see the moment you click a prompt card.
Deferred (instant) start
When you start a fresh interview from the in-app Dashboard, Omniscio does not spawn a Claude process right away. Instead it:
- Creates the session in a
ready(not-yet-launched) state — no CLI process, no cost. - Shows the prompt's authored
firstMessageimmediately as Claude's first turn. - Stashes the fully-composed CLAUDE.md (artifacts, profile, bookmarks, plus a
## You Have Already Greeted The Userhandoff section) on the session row in apending_claude_mdcolumn.
The real Claude session launches only when you send your first reply. At that moment Omniscio writes the stashed CLAUDE.md to disk (prepareCoachingInstructionsForColdStart, which flushes it) right before the process starts, so Claude reads the correct interview and the handoff section — it knows it already greeted you and simply continues from your reply, with no double-greeting. If you never reply, the draft just sits in your Sessions list costing nothing; reply any time — even after restarting the app, since the pending CLAUDE.md lives in the database — to start it.
An unanswered draft stays yours after a restart too. Restarting ends it (every not-yet-launched session with a message is ended at startup, and the greeting counts), but its interview card keeps reading In progress, and Continue or Start interview reopen that same draft instead of building a second copy beside a hidden first one. And if you import your own write-up for that interview while the draft is still unanswered, the empty greeting is never treated as the chat that wrote it: no "How helpful was this session?" prompt, and no Open chat to a conversation you never had. The draft is recognised by its still-stashed pending_claude_md, which your first reply clears.
A redo of a completed interview, and any start triggered through the CLI (where there's no interactive user to greet), skip the deferral and spawn immediately using the openingMessage template as before. Quick check-in (snack) sessions and free-form sessions (opening a new session in the AI Coaching project without typing a prompt) defer the same way, each with their own static greeting. The free-form greeting is honest about what the coach knows: with no Core Profile and no saved write-up (a private check-in counts as one), it says you are starting fresh and points you to a short interview from the AI Coaching dashboard — "Hey — I'm your AI coach. We're starting fresh, so I don't know anything about you yet. A short interview from the AI Coaching dashboard is the quickest way to get going, or just tell me what's on your mind today." Once you have a Core Profile or any saved write-up, it opens with "I have your full coaching profile and artifacts, so no need to repeat anything" instead.
Opening greetings are shown/injected only at session creation, not when resuming a paused session — a resumed interview picks up where you left off without Claude re-introducing itself. That holds even if you started a different interview in the meantime: each AI Coaching session has its own room with its own instructions file (~/amc-ai-coaching/sessions/<sessionId>/CLAUDE.md), and each time Omniscio starts a session again — after an app restart, after a usage-limit pause, or when a scheduled message wakes it — it first makes that file right. An interview that has already produced output is always rewritten as "already underway" rather than "your first response"; any other session's file is rebuilt only when its room has none.
Previously, opening-message guidance was buried inline inside prompt bodies as <First Response> blocks. Every built-in prompt now has dedicated openingMessage and firstMessage fields, making the greetings easy to maintain and update.
Bookmarks
You can bookmark key moments during a coaching interview — a quote, an insight, or a useful observation from a specific message. Bookmarks are per-session and displayed in the BookmarkPane when viewing a session's details. Each bookmark shows the quoted content as a blockquote, an optional label, the date it was saved, and action buttons to copy the text or delete the bookmark.
You can also write a bookmark yourself: the Add bookmark button in the Bookmarks pane (and the + beside Bookmarks in the sub-sidebar) opens a form, and Save files it against your most recent coaching session, because every bookmark has to belong to a real session (the database rejects one that does not). The pane finds that session through the AI Coaching project's own project row, whose id is generated when the project is created, not the __ai_coaching__ folder name. With no coaching session yet the button stays greyed out and says to start one first, and the sidebar + shows that same reason instead of opening a form it could not save; if the last coaching chat goes away while the form is open, Save says why in a toast and keeps the typed text. When the form closes — Save, Cancel, Esc, or Discard on the "Discard this bookmark?" prompt — keyboard focus goes back to Add bookmark, so a keyboard user carries on from the same spot instead of starting over at the top of the page. Earlier builds looked the sessions up by the folder name, so the button stayed greyed out for everyone and the sidebar + form's Save did nothing.
Deleting a bookmark pops a confirmation dialog that quotes the bookmark it is about to remove, so you can see which one you clicked before you commit. The bookmark is erased from the database straight away, together with the copy Coaching Search uses to find it, so the dialog tells you plainly that it cannot be undone.
You can also create bookmarks directly from the chat conversation: each message bubble in an AI Coaching session has a bookmark button. Clicking it creates a bookmark of that message's content via the AI_COACHING_CREATE_BOOKMARK IPC channel, and the BookmarkPane updates automatically via the AI_COACHING_BOOKMARKS_CHANGED push listener.
Bookmark context injection
Bookmarks aren't just personal notes — they're injected into the AI coach's context so it can reference your key moments. When a coaching session starts (or resumes after compaction), your bookmarks are loaded and rendered as a #### Key Moments (marked in your coaching sessions) block (the wording never claims you marked them all, because a coach can bookmark a moment on your behalf) inside the ## What You Already Know section, positioned after the core profile and before the artifact blocks.
Each bookmark appears as a blockquote with an optional bold label and interview attribution (e.g. *(from Values & Principles)*). Labeled bookmarks come first (the user took extra effort to name them), then unlabeled ones, with the shortest first in each group so more fit. The block is capped at 3,000 characters; a bookmark that would overflow it is left out. Up to 50 bookmarks are considered.
If a bookmark came from a session that no longer exists (edge case), it still appears — the interview title is simply omitted. If you have zero bookmarks, no #### Key Moments section is emitted at all. This is a pure formatting step at session load time with zero AI cost.
Goals
The Goals pane tracks the things you're actually working toward. A goal has a title, an optional description and target date, a category (Career, Relationship, Personal, Health, Financial), and a status: Active, Achieved, Paused, or Abandoned.
Goals arrive two ways. The AI extracts them from your coaching conversations, and — since the
AI_COACHING_CREATE_GOAL channel landed — you can also write one yourself with the New goal
button in the pane header. A goal the AI extracted carries a chip naming the artifact it came from,
so you can tell the two apart and trace an extracted goal back to the session that produced it.
Goal titles are unique, ignoring capital letters and spaces at either end. Creating a goal whose title you already have creates nothing: the New goal form stays open with everything you typed, and a notice says "You already have a goal called “{title}”." naming the goal you already have, so you can rename the new one or cancel. Renaming a goal to a title another goal holds is refused the same way. A real create closes the form and resets the status filter to All, so the new active goal is never hidden behind an Achieved or Paused filter. When Cancel or a real create closes the form while focus is inside it, keyboard focus returns to the New goal button rather than dropping to the page; discarding the draft to leave the pane does not pull focus back. Cancel on a form you have typed into asks first with the same "Discard changes?" prompt the bookmark form uses: Discard closes the form, Keep editing keeps your text. With nothing typed, Cancel closes it at once.
Filtering is by status, with an All default. A filter that matches nothing says so in its own words ("No Active goals") and offers a one-click way back to the full list, rather than showing the same "no goals yet" message a genuine empty pane shows. Changing a goal's status while a filter is active fires a notice naming the goal, its new status and the filter hiding it, with a Show all action — a row that vanishes without explanation reads exactly like a deletion.
An empty Goals pane says what is true about your setup. With an Anthropic API key set up it says no goals were found in your conversations yet and points at New goal; with no key it says goals can't be found until one is set up, still points at New goal, and shows an Add a key in Settings → Accounts button.
The list loads 100 goals at a time; a Load more control appends the next page when there are more.
Two status changes ask first: Achieved and Abandoned. Both drop the goal out of the default Active view, so a mis-aimed click on the dropdown would otherwise look like the goal had vanished. Active and Paused apply immediately — they are one step apart and putting a dialog in front of every status change trains you to dismiss dialogs without reading them. All four statuses are reversible from the same dropdown, and the dialogs say so.
Check-ins
Expanding a goal reveals its check-in history and a form to add one. A check-in is a short note plus a 1–5 progress rating, given as a labelled radio group (one Tab stop, arrow keys to move) with the current value spelled out. The note field shows its 2000-character budget as you type. Each goal keeps its own half-written check-in, so opening a second goal to check something and coming back does not lose what you had typed. Saving a check-in clears only that goal's box, and only when it still holds the note that was saved: words typed while the save is on its way stay in the box (with the rating) instead of vanishing.
The history says "No check-ins yet." only when the goal's check-ins have actually loaded and there are none. On an active goal that line comes with an Add your first check-in link that jumps to the note field, so reading it gives you a next step as well as a fact. While the check-ins are still loading the section says "Loading…". If they could not be read it shows the error and a Try Again button that fetches them again, so a goal that has check-ins is never shown as having none. A check-in list that loaded once stays on screen if a later refresh fails. Implementation: the pane derives the loading, failed or loaded state in GoalsPane.tsx and the card renders it in GoalCard.tsx.
Check-ins are offered on active goals only — a check-in is a statement about current progress, and a note filed against a goal you have closed out would be hidden by the very filter that hides the goal. Goals past their target date while still active carry an Overdue marker. A target date in the current year reads as month and day ("Target: Dec 31"); one in any other year carries its year ("Target: Dec 31, 2025"), so a date from last year never looks like one still to come. A stored date that cannot be read is shown as written.
Deleting a goal
Deleting moves the goal and its check-ins to Recently deleted, restorable for 30 days, then erased for good. The confirmation names the goal and itemises what goes with it — the goal itself, and its check-ins, counted when they have loaded. When the check-in count has not loaded it says "every check-in recorded against it" rather than claiming zero. The pane loads at most 50 check-ins per goal, so a full page of 50 also says "every check-in recorded against it": the goal may hold more, and the delete moves them all. While the delete runs, the Delete button shows it is working and Cancel is locked, so a second click cannot send it again.
Related
The rest of the feature is on the AI Coaching sidebar (interview prompts and artifacts) page: its artifacts in part 2, its memory and profile in part 3, and its CLI surface in part 4. CLI control is the companion page for driving coaching from an agent.
Last verified 2026-11-04