DREAMTEAM — Working Spec

What it is

DREAMTEAM is the operating layer for a pitching team. One coherent surface that replaces the deck tool, the meeting recorder, the team back-channel, the agent dashboard, and the team’s working memory.

DREAMTEAM splits into four buckets. Each bucket ships as a coherent slice; features within a bucket share data and UX.

Presentation Deck — building and rendering the pitch.
Pitch Intelligence — adapting the pitch to the buyer (before, during, after).
Team Workspace — how the team operates around pitches: chat, agents, memory, recordings.
Creative Intelligence — proprietary data + ML on creative-industry winners: who wins, how they win, deploy them as your team.

Bucket 1 — Presentation Deck

Pitch deck builder

Brand-kit-driven. First pass for any client = custom build (encode their type system, color, motion, slide grammar from decks they’ve shipped). Upfront fee.
Plug-and-play forever after. Recurring subscription on top.
House-line packs for clients without time / budget for a custom build.
Delivery is hardcoded as weblinks — every deck / proposal ships as a shareable URL, never a .pdf / .pptx / .key attachment. Easy client delivery, single source of truth, live edits propagate without re-sending files.

Adaptive viewer system

Same source renders as click-through deck OR long-scroll proposal.
Per-device default (desktop → deck, mobile → scroll), no URL view-locking, hover-toggle preserves equivalent section.
Already shipped at parzvl.com/dual-template.
Async layer: viewer journey mode (“60-second version,” “show me the price”), adaptive prompts, Q&A → on-the-fly slide synthesis, per-viewer journey log back to presenter.
Long-term: voice-narrated mode, chat-mode shape, print-export, embed shape, Responsive Avatar Mode (Loom replacement — live AI persona walks the deck and answers viewer questions mid-watch).

Brand preset system

Versioned BrandPreset object: palette, type system, logo treatments, motion, spacing, voice / tone, component skins.
Editing the preset propagates to every deck that references it. Per-deck override allowed without forking.
Global mode toggles: Executive (tight, muted, serif body), Concept (loose, big motion), Print-safe (flat, hardened contrast), Dark / Light.
Two-view fluency: every token applies to deck + scroll views simultaneously.

Template marketplace

Designers publish slide-type templates (cover / mission / market sizing / traction / capabilities matrix / ask / etc.).
Templates auto-conform to the user’s active brand preset on download — no “make this fit our colors” busywork.
Marketplace economics: revenue routes to designer, % cut TBD. Editorial quality gate. House-line packs from us as the canonical reference.

Bucket 2 — Pitch Intelligence

The substrate (live editing engine), the three layers that ride on top of it (before / during / after the call), plus the two in-room team surfaces the pitch runs on (invisible group chat + live performance dashboard).

Live editing engine (the substrate)

Edit primitives: swap, insert, reorder, skip, inline edit, brand-override.
Push lifecycle: Proposed → Approved → Countdown → Landed → (optional) Reverted. Visible countdown timer per push.
Smart auto-restructure — if the next pending edit isn’t done rendering when the presenter advances, promote the next-ready slide and reinsert the edited one when ready. No “uh… one second” dance in front of the buyer.
Slide-on-demand: generate a brand-new slide from prompt during the pitch.
Pre-pitch adaptation, voice-reactive extension, and follow-up deck all push their changes through this engine.

Pre-pitch adaptation (before the call)

Calendar awareness — auto-detects upcoming pitches via attendee external domains, agenda keywords, deck attachments, recurring buyer patterns.
Auto-research — pulls recent buyer news, exec moves, earnings transcripts, social signals from named attendees, past comms with the buyer, competitive context.
Auto-tailor — cross-references the buyer briefing against the active deck and proposes mods (T-24h deep, T-2h final).
Auto-rehearse (optional) — engine simulates the buyer in a dry-run pitch; flags weak points by likelihood-the-buyer-will-ask.
Human-in-the-loop: never pushes auto-tailored edits without explicit presenter approval.

Voice-reactive pitch extension (during the call)

Listens in real time. Understands what’s being said, what’s landing, what isn’t.
Surfaces inline suggestions to the presenter (“pull retention slide forward,” “swap the Mexico section out,” “drop the second testimonial”).
Approved suggestions push live to the deck mid-pitch via the live editing engine.

Invisible group chat

Hidden back-channel where the team collaborates while a pitch is happening.
Each teammate has independent viewport state (Figma / Google Docs principle).
Invisible to the room when sharing a window (kCGWindowSharingNone on macOS, WDA_EXCLUDEFROMCAPTURE on Windows).
Open question: invisibility under full-screen-share. Granola hasn’t solved it either, suggests it’s hard.

Live performance dashboard

Pitch coach, live in the room. Keeps the presenter on track, flags rambling, makes sure the buyer is getting room to speak. Same job a great pitch coach does sitting next to you, except in software and in real time.
Renders inside the invisible chat surface — same back-channel, team-only audience, buyer never sees it.
Live in-the-moment, not summary. Signal updates frame-by-frame during the pitch; post-call recap is a separate artifact.
Slide timer — time spent on the current slide vs. a preset target per slide / section. Fires a soft cue to the presenter (and the team chat) when the target hits.
Talk-time balance — who’s talked and for how long. Surfaces whether the presenter is dominating vs. leaving room for the buyer to ask questions. Per-attendee bars, refreshed live.
Room transcript surface — running transcript of who said what, key phrases highlighted. Lets a teammate scan back without re-listening.
Pacing flags — going long, going short, buyer hasn’t spoken in N minutes, presenter cut buyer off, etc. Surfaced as inline chips in the team chat, not modal interrupts.
Mutates nothing — unlike voice-reactive extension (which pushes deck edits), the dashboard only renders signal to the back-channel the team is already watching. It lives in Bucket 2 because it’s a live in-room pitch surface, but its output is team awareness, not deck mutation.
Open question: how much of the signal also feeds the voice-reactive engine (e.g., “buyer hasn’t spoken in 4 minutes” → suggest a question slide).

Follow-up deck (after the call)

Post-pitch artifact. Same brand-preset + content corpus that produced the pitch deck produces the follow-up deck — meeting recap, next-step proposal, contract preview, ask-specific addendum, whatever the buyer requested in the room.
Auto-drafted from the meeting capture transcript + the deck that ran + the agreed action items, ready for human review on the way out of the call.
Out of scope: broader creative production (short-form ad cuts, social posts, email sequences, landing-page modules). Follow-up decks only — not an outbound creative engine.
Spec depth deferred — to be written up.

Bucket 3 — Team Workspace

How the team operates around the pitch — agents, memory, recordings, history. (The in-room pitch surfaces — invisible group chat + live performance dashboard — moved to Bucket 2.)

Multi-agent orchestrator + task board (the project management system)

Kanban-style task board owned by the team.
Board population — tasks land on the board from three sources: (1) human creation, (2) auto-extracted action items from meeting capture (Bucket 3) and follow-up decks (Bucket 2), (3) agent-proposed tasks emitted mid-run (e.g., a research agent surfacing “we still need a competitive teardown” as its own card for human approval).
Every agent run a teammate fires can be attached to a task. Status, current step, model, tokens, ETA stream live.
Swarm visibility — when Allan launches a 15-agent strategic-research swarm against a brief, the swarm appears as a single composite run with each agent expandable.
Swarm-sizing recommendation — given a brief, the orchestrator proposes a swarm shape (e.g., “5-agent mini swarm + 1 reviewer” vs. “15-agent full swarm with mesh coordination”) based on brief complexity, deadline, available context, team capacity, and prior swarm performance on similar work. Human approves or overrides before launch.
Auto-recognition of completion — agents emit structured task-delta events; DREAMTEAM marks progress and drafts a one-sentence summary for human review. Includes dedupe against the existing board — if an agent’s output satisfies a task that’s already on the board (even one no one explicitly attached the run to), DREAMTEAM flags it as a candidate completion rather than silently re-opening or double-counting. Catches the “you already finished this last week, don’t redo it” case.
Efficiency lens — per-teammate view of how AI is actually being used.
Privacy gate — solo / scratch sessions stay private. Only runs explicitly attached to a task surface to the team.
The orchestrator is the runtime that deploys agents — individuals or pre-built squads — against tasks on the board.

Conversational agent interface (two-way voice)

Voice in is already solved (Wispr-style dictation → terminal). The missing half is voice out: the agent runtime talks back and iterates conversationally, like Claude voice mode — not just printing text to a buffer.
True turn-taking loop — teammate talks, DREAMTEAM responds aloud, teammate reacts, the agent adjusts. Hands-free operation of the orchestrator.
Cross-cutting modality — applies anywhere you’d otherwise type at an agent: briefing a swarm, reviewing a run mid-flight, querying Brain². Same agent runtime, voiced front-end.
Open question deferred: TTS engine + barge-in handling (interrupting the agent mid-sentence) — see Open questions.

Meeting capture (Granola or fallback)

If Granola is connected — hook into its transcript / summary feed, store a pointer.
If not — native fallback recorder captures computer audio, summarizes to the same shape (action items, decisions, attendees, named-entity tags).
Hosted on the user’s tenant, encrypted at rest, retention controlled by the team.
Feeds pre-pitch briefings, follow-up decks, and Brain².

Conversational chat layer (history + auto-routing)

Claude-style sidebar of every past conversation, fully navigable — scroll back, branch, resume.
Context-drift detection (Wispr-style) — when a conversation shifts topic, DREAMTEAM spins up a new thread that stays aware of the prior one. Second-brain continuity, live between conversations rather than reconstructed after the fact.
No manual save — auto-saves and auto-populates a concise running doc as the conversation unfolds.
Auto-route on close — archive / close / inactivity triggers an instant save and routes the doc to the right destination (project folder, Brain² tier, task board). Optional confirmation chip tells the user where it landed; no approval gate required.
Direct Brain² intake — the auto-routed concise docs are the warm/cold feed for Brain².

Migration & onboarding

One-click import from the tool the team already lives in. Notion first (workspaces, docs, databases, task boards), then adjacent stores (Google Docs/Drive, Linear, Asana, Coda).
Zero-config path — connect the source via OAuth, pick what to bring, DREAMTEAM maps it: Notion pages/docs → Brain² (warm/cold), Notion databases/boards → task board, Notion teamspaces → DREAMTEAM workspace structure. No manual re-entry.
Preserve structure, not just text — page hierarchy, links between pages, properties/tags, and assignees carry over so the imported workspace is navigable on arrival, not a flat dump.
Incremental re-sync (optional) — keep pulling from Notion until the team has fully cut over, so migration isn’t a hard cliff.
Open question: live two-way sync vs. one-time cutover import — see Open questions.

Brain²

Tiered store: hot (active project context), warm (recent / semantically close), cold (historical archive on a sublinear-search index).
Smart cycling keyed to the active project / quarter / objective. Project switch = hot cycles.
Critical-flag protection — live deals, signed contracts, regulatory material never auto-demote.
Reversibility — every demotion is logged, one-click re-promote.
Sublinear data structures (HNSW / FAISS-IVF / bespoke) so retrieval cost stays bounded at year-3 corpus scale.

Bucket 4 — Creative Intelligence

A model layer trained on the public archive of award-winning case films. The thesis: case films that win at Cannes Lions (and adjacent shows) are the most statistically validated examples of creative excellence in advertising history. Reason over what actually moves the brain — pacing, narrative arc, emotional beat-mapping, copy density, music cues — and surface that intelligence back into the deck and the pitch.

Love-the-Work-More AI

A multi-modal study of award-winning work. Goal: an AI that understands everything about winners — how they won, why they won, the data they collected, how they presented it, the scripts, the language, the visuals, the craft of the case film itself.

Training layers (not just one model — a stack):

Brain-response layer — Tribe V2 and adjacent neural-prediction models on case films. How the work actually moves the brain — pacing, beat-mapping, emotional arc.
Visual models — design language, motion, color, typography, edit pacing, shot grammar across winning case films and the campaign visuals they document.
Language models — scripts, voiceover, copy, headlines, taglines. How winners write.
Presentation craft — the case film itself as artifact: act structure, hero-stat placement, argument scaffolding, jury-facing rhetoric, how a winning case sells the work to the room.
Data + evidence — what metrics winners gathered, how they framed them, what proof they brought, how they sourced it.
Strategic thesis — the “why it won” underneath the visuals. The argument the case film built.

Training corpus: lovetheworkmore.com Cannes Lions archive + adjacent award-show feeds (D&AD, One Show, Clios, Effies — same corpus the award-show DB feeds from).

Outputs:

Case-film scoring engine — given a draft case film, emit a creative-excellence score with sub-scores (pacing, narrative arc, emotional impact, visual craft, language craft, evidence quality, awards-likelihood) plus diff suggestions (“cut 4s from the setup, push the hero stat earlier, the VO cadence is 22% slower than median winners”).
Generative case-film scaffold — given a campaign + brand inputs, produce a structurally-validated outline (act structure, beats, voiceover cadence, music cues, visual grammar, evidence layout) before any production starts.
Long-arc campaign success predictor — extends from “case films that win” into “campaigns that work.” First statistically-grounded answer to “is this campaign going to land?”

Access surfaces

Desktop-first (where the deck lives). Mobile is a first-class companion for jobs that happen off-laptop, not a port.

Desktop / web (primary) — full surface. All four buckets.
Mobile companion (native — React Native / Expo so the data layer stays shared with web). Three jobs that justify native over PWA:
- Presenter-side during a pitch — invisible chat from phone (a separate device from the screen-shared laptop), push approvals for mid-pitch edits, voice-extension suggestions on lock screen. Push-notification UX and lock-screen actions are why this isn’t PWA.
- Team-side between pitches — monitor agent runs, review pre-pitch briefings, async chat, browse swarm run output. Read / light-write.
- Capture-side — voice memo into Brain², whiteboard photo, share-sheet drop of an external deck. Camera / mic / share-sheet require native APIs.
Bridge: PWA-installable web ships first so day-one users have mobile access while the native app is scoped.
Out of scope: SSH / CLI remote access. Power users with Tailscale + Blink Shell already reach their own desktop; DREAMTEAM won’t ship a CLI client.

Open questions

Granola partnership — API integration, white-label, or screen-scraping shim?
Full-screen-share invisibility — solvable cross-platform, or accept window-share-only?
Per-brand upfront fee — fixed price, scoped by deck volume, or tiered by starter-pack size?
Marketplace cut — flat % or tiered revenue share?
Push-lifecycle countdowns — static defaults or auto-tuned by edit complexity?
Task-delta event format — canonical schema agents emit so the orchestrator can auto-recognize completion?
Cross-tool agent ingestion — Claude Code / Cursor / arbitrary CLIs, or only DREAMTEAM-launched agents?
Sublinear index choice — HNSW / FAISS-IVF / bespoke?
Award-show ingest legality — scraping PDFs / ceremony programs vs. licensed feeds (Cannes Lions sells data access).
Mobile stack — React Native / Expo (recommended) vs. Flutter vs. parallel SwiftUI + Kotlin. Stack pick determines code-share with web.
Mobile launch order — iOS-first, Android-first, or parity at launch?
Swarm-sizing model — heuristic table, learned-from-history, or LLM-judged? Where does the recommendation live in the UI (modal before launch, inline suggestion, both)?
Board-dedupe matching — how strict before flagging an accidental completion (exact title? semantic match? attached artifact overlap?)? Default action: flag-for-review vs. auto-close-with-undo.
Voice talkback stack — TTS engine + barge-in handling (interrupt the agent mid-sentence), and does it run client-side or stream from the runtime?
Context-drift threshold — what signals trigger an auto-spawned new thread (topic-embedding distance? explicit cue? silence gap?), and how aggressive before it annoys?
Migration depth — one-time cutover import vs. live two-way sync with Notion (and which sources get sync vs. import-only)? How to handle conflicts if both sides edit during incremental re-sync?

Updated 2026-05-22.