A semi-automated discovery system for product teams

I'm setting up a discovery sprint and need your help defining the research frame before I start synthesising sessions. I want to end up with a concise "Discovery context" document I can paste into a Claude Project. Walk me through these one at a time — ask one question, wait for my answer, then move to the next: 1. Product context — what are we building, who is it for, and what core problem does it solve? 2. Discovery objectives — what are the 3–4 key questions we need to answer by the end of this sprint? 3. Participant breakdown — how many prospects vs customers? Any relevant segments or sub-groups I should track separately? 4. Hypotheses being tested — what are the 2–4 assumptions we're specifically testing? For each: what would validate it, and what would invalidate it? 5. Themes to track — what topic areas should appear in every synthesis? (e.g. workflow fit, pricing sensitivity, data pain, competitor awareness, setup friction) Once we've covered all five, output a clean "Discovery context" block I can paste directly into a Claude Project as system instructions. Format it as an active analytical lens — not just background reading. Claude should use it to evaluate every session it synthesises.Copy

You are a product discovery analyst. I'm running a discovery sprint for [product name] with [number] participants split between [prospect type] and [customer type]. [Paste your discovery context output from Step 1 here] When I upload a transcript or notes, synthesise it using this structure: **Session synthesis — [Participant name]** Type: Prospect / Customer Date: [date] Facilitator: [if known] **Top 3 themes from this session:** For each: theme name | sentiment (positive/neutral/negative) | key quote **Hypothesis signals:** For each hypothesis from the discovery context: Validates / Invalidates / Neutral — and the specific evidence **Buy intent / Churn signal:** [Prospects: High / Medium / Low intent with reason] [Customers: High / Medium / Low churn risk with signal] **Standout moment:** The single most surprising or important thing from this session (1–2 sentences) Keep synthesis tight — I'll be running this across [number] sessions and comparing patterns across all of them.Copy

Synthesise the attached transcript. Participant: [name], [company if relevant], [prospect or customer]. Date: [date]. After the synthesis, update the running participant log with this session — I'll ask for the full log at the end of the sprint.Copy

Across all [number] sessions you've synthesised, give me the full project view: 1. Hypothesis scorecard — for each hypothesis from our discovery context: how many validate, how many invalidate, overall verdict, and the 2 strongest pieces of evidence on each side 2. Theme clusters — top 5 themes, each with: mention count, split by prospect vs customer, overall sentiment, and 2 representative quotes 3. Sentiment overview — overall positive/neutral/negative split for prospects vs customers separately 4. Buy intent breakdown (prospects only) — High / Medium / Low with names and the key signal for each 5. Churn risk signals (customers only) — High / Medium / Low with the specific signal that put them there 6. What to build next — recommendation on the top 2–3 priorities based on hypothesis validation, theme frequency, and commercial impact. Be specific and evidence-backed.Copy

I want to build a discovery synthesis dashboard for a product sprint. Before writing any code, help me architect it properly. Work through these with me one at a time — ask the question, wait for my answer, then move to the next: 1. Data model — what entities do we need? Think through: sessions, participants, themes, hypotheses, quotes, signals, synthesis outputs, cohorts. What relationships between them actually matter for the dashboard to work? 2. Synthesis pipeline — does synthesis happen per-session on demand, in batch, or both? Where in the flow does the Claude API call sit — on upload, on user trigger, or scheduled? What gets stored vs re-computed each time? What happens if a session is re-synthesised? 3. Hypothesis tracking — hypotheses need to be evaluated per session (validates / invalidates / neutral) and aggregated across all sessions. How do we structure that cleanly without making it a manual tagging exercise every time a session comes in? 4. Cohort comparison — if we're testing two product directions (Option A vs Option B) with the same or overlapping participant sets, or comparing prospects vs customers on the same questions, how does the data model handle that without duplicating sessions? 5. Cross-session query — when someone asks "which hypothesis has the strongest validation?" or "who appears in both high churn risk and hypothesis invalidation?", that needs to run across synthesised data from all sessions. At 30–60 session scale, what's the right approach — structured JSON query, summarised context window, embeddings search, or something else? What are the tradeoffs? 6. Storage for v1 — what's the simplest storage layer that handles: session state, per-session synthesis output, cross-session theme aggregation, and hypothesis tracking? When does it become a bottleneck, and what's the natural upgrade path? 7. Dashboard views — what are the minimum views to make this system genuinely useful? For each view: what's the core question it answers, what data does it consume, and what interactions does it need? Once we've worked through all seven, output: — A data schema showing each entity as a JSON structure with fields and types — A component tree for the dashboard with the data each component consumes and emits — The Claude API prompt chain: system prompt design, per-session synthesis call, cross-session analysis call — with notes on what context each needs — A recommended build order with the reasoning behind the sequenceCopy

Build a discovery synthesis dashboard as a single self-contained HTML/JS file. No backend, no build step, no external dependencies beyond Satoshi font (Fontshare CDN). Architecture to implement: [Paste your architecture document from Step 1 here] Layout and design system: - Two-page tab structure at top: "Data Feeds" | "[Your project name]" - Top bar: product wordmark left, "Discovery System" label centre, sync status badge right - Colours: #fafafa page background, #ffffff card surfaces, #1a1a1a dark elements, #e5e5e5 borders, #6b6b6b muted text - Warm gradient mesh behind the page (orange/pink/purple radial gradients fading to white at top) - Pill buttons (border-radius: 50px), fadeSlideUp entrance animation on load elements Data Feeds page — build these components: - Stats bar: Total inputs, Prospects count, Customers count, Pending synthesis count, Last sync timestamp - Source tiles (3): Transcripts folder with file count, Slack feedback channel with live indicator dot, CS notes with entry count - Session feed: 6 entries minimum, each showing filename or source, timestamp, participant type tag (Prospect / Customer), project tag, synthesis status badge (Pending / Synthesised) - Session slide-in panel (click any entry to open): participant name + date + facilitator, metadata grid, transcript excerpt that blurs at the bottom edge, "Synthesise now" primary button → 2-second spinner → synthesis output (hardcoded), muted "Demo mode — would call Claude API" note beneath the button Project page — build these components: - Project header: project name, session count pills (total sessions, prospect count, customer count, last updated date), active sprint badge - Option A / Option B toggle buttons with a description line beneath that updates per option - Two-column layout per option (Prospects left, Customers right), each column containing: — Sentiment slider: gradient colour track, position marker dot, percentage label below — Hypothesis alignment section: 3 hypothesis cards per column, each with the hypothesis text, a "Validates" row with coloured avatar initials, an "Invalidates" row — clicking a card opens a modal with full quotes and per-person verdict — Key themes section: 3–4 theme cards per column, each with theme name, sentiment badge, avatar row — clicking opens a modal with all supporting quotes and participant list — Buy intent section (Prospects column only): three buckets (High / Medium / Low), each with named participants + one-line signal — Churn risk section (Customers column only): three buckets (High / Medium / Low), each with named participants + specific risk signal - Chat panel fixed to the right of the project layout: "Query signals" header, sub-label, pre-populated conversation with 3 exchanges (user question then Claude response), disabled input field with "Demo mode" label Hardcoded data — make it feel like a real sprint: - 6 sessions: mix of synthesis states, named fictional people, specific fictional companies, realistic timestamps - 3 theme clusters per cohort, each with 3+ verbatim-style quotes attributed to named participants - 3 hypotheses tracked per option with a realistic validates/invalidates distribution (not all validating) - Buy intent: 2 High, 3 Medium, 2 Low prospects with specific signals - Churn risk: 2 High, 3 Medium, 2 Low customers with specific risk signals - Chat: 3 pre-written exchanges — one on hypothesis validation strength, one on why cohorts differ, one on overlap between churn risk and hypothesis invalidation - No lorem ipsum. No "Company A." Named people, specific companies, specific quotes. Product tour (5 steps, auto-starts on load): - Step 1: Centred intro card, no spotlight — what the system does in one sentence - Step 2: Spotlight on stats bar — inputs, sources, synthesis on demand - Step 3: Switch to project page, spotlight on project header — session count and sentiment overview - Step 4: Spotlight on A/B toggle — testing two directions across the same participants - Step 5: Spotlight on chat panel — querying the full dataset No live API calls anywhere. All synthesis outputs, theme data, and chat responses are hardcoded. Actions that would trigger a real Claude call show a clearly labelled demo state.Copy