#Design brief: the live call screen, round 3 (2026-09-25)

#Why a fresh round

Stevan's verdict on variant D: the card on the right "floats aimlessly"; the "careful" notice is stale (it refers to a turn from a while ago and need not be read after); the "covered so far" dots mean nothing without labels; above all, the screen shows a concern box with suggestions but no overall structure for the call at hand. The reference copilot (moritzkremb/jev-sales-copilot) gives the rep a structure (stage rail, facts, checklist that ticks) and we do not. He does not want a fixed list for every call: Jev should inform and guide a structure that fits this call.

#The concept to design around: a living call plan

One object, the call plan, replaces three things that were separate before (call stage, must-say checklist, next-best-move context):

• A plan is a short ordered list of items for THIS call: for an onboarding call after sign-up, e.g. Understand the transfer · Check we can serve them · Documents still needed · How the money moves (who holds it, rate against mid-market, settlement, fund from own account, booking is binding) · First transfer · Next step. For a customer-success call it depends on the purpose: Resolve the issue · Status with specifics · Upcoming payments · Tools that fit (alert, forward, recurring) · Next step.
• Jev adapts the plan to the call: it classifies the purpose (why this call is happening), knows client facts (personal vs corporate, amount, corridor, activated or not, documents outstanding) and marks each item applies / not needed for this call; it detects when an item has been covered (with the utterance that did it), which item is current (where the conversation actually is), and it may add an item when the client raises something (an open concern becomes a plan item until it is handled). Nothing on the screen is a fixed checklist.
• Each item carries 2-3 approved points worth mentioning (from the playbook) and, when a concern is open, the response points for that concern. The current item's points are the only text a rep needs to read.
• States a rep must read at a glance: done (with when), current, next, not needed on this call, added by the client. Labels, never anonymous dots.

#What the live screen must do

• Answer at a glance: where are we in this call, what have we covered, what is worth saying right now, what is still to come.
• One reading focus at a time. The current item's points are the largest text after the headline. Everything else recedes.
• Transcript stays (Stevan likes it), with the rep's and client's turns distinguishable; utterance links from covered items are welcome but small.
• A caution (e.g. "that sounded like a rate prediction") is shown at the moment it happens, near the turn it refers to, for about 15 s, then gone. Never a persistent box.
• Cadence: items change state after a client turn ends; a point ticks when covered and stays visible briefly; at most one new thing to read at a time; nothing blanks; "listening" when Jev has nothing.
• No signals grid, no probabilities, no thresholds, no telemetry, no acknowledge/dismiss, no "since #34", no "rule hidden because". Those live in the review view (out of scope for this round).
• Both scenarios (onboarding mid-call with a concern open; customer success with an issue being resolved), light and dark, 1600x1000, no horizontal scroll at 1440 wide.

#What the engine can provide (design within this, extend only when you say what is new)

• Per utterance: stage guess with confidence; persisted facts (purpose, pair, amount bucket, corporate/personal, docs status, funding source...); must-say items covered with utterance index; open concern with topic; next move with confidence and its approved lines; risk flags; call purpose (new, to be added); item applicability per client facts (rules exist as applies conditions).
• Approved playbook lines, 1-3 per move; concern responses; nothing generated live.

#Anti-patterns (each one was called out; treat as a failing test)

• A card floating in empty space with no anchor to the structure.
• A stale notice that persists after its moment.
• Dots, bars or numbers with no label the rep can read.
• Debug language: thresholds, confidence numbers, rule explanations, utterance-count references.
• Everything shouting: signals, stage, checklist, card, flags all competing.
• A fixed checklist that ignores the purpose of the call.

#Four directions (must feel like four different teams: different layout, type, colour)

• E, Agenda spine: the plan as a vertical stepper on the left; the current item expands in place with its points; transcript in the centre; nothing on the right.
• F, Plan strip + focus: a stateful horizontal plan strip across the top (done / current / next / not needed, with labels); one large focus card below it; transcript beneath or beside, scrolling.
• G, One list: the transcript full height on the left; on the right the plan as a list of collapsible cards where the current one is open in position, done ones collapsed with a tick and time, upcoming greyed; a client concern is inserted as a card at the current position. Plan and suggestions are the same object.
• H, Teleprompter: extreme distillation for the live moment: a one-line breadcrumb (previous · current · next), a large centred card (headline + two points), the last three turns of transcript as a quiet ticker at the bottom; a "plan" drawer for the rest.