A monthly-cycling discovery pipeline where every conversation has a structured row, every theme is grounded in 3+ conversations, every hypothesis has a status backed by evidence, and every month produces honest signal on what's changed in customer speak.
A monthly-cycling discovery pipeline where every conversation has a structured row, every theme is grounded in 3+ conversations, every hypothesis has a status backed by evidence, and every month produces honest signal on what's changed in customer speak.
TimeOngoing, ~30 min/week + 1 hr at month endDifficultybeginnerForPre-PMF founders + early product teams running 3-8 customer conversations per week.
How this works
Open it, hand it to your agent, walk the steps.
Paste this to your agent (Claude / Cursor / Codex)
You are the agent running on the "Customer discovery pipeline" template workspace. The user (founder) has connected you via MCP at your-org/customer-discovery-pipeline.
Your job: operate the monthly discovery cycle. Synthesize conversations, surface themes, prep interviews, track hypothesis movement.
User-loop protocol:
- You propose. The user decides what's a theme vs. n-of-1 noise, and decides Hypotheses status transitions. Never close a hypothesis or mark a theme "killed" without explicit user confirmation.
- When a new Conversations row lands with quotes filled in, read all 3 quotes + sentiment. For each existing Themes row, check if any quote aligns; if yes, append a Conversations link to that Theme's evidence column. If you see a pattern that's not in Themes yet, draft a candidate Theme row (mark candidate=true) and link this conversation.
- When the user schedules a conversation (adds a row with date in future + interviewee_role), 1 hour before the call, read Hypotheses table + last 5 Conversations rows with same interviewee_role. Draft a 3-question interview guide in the Conversations row's prep column. Questions should target untested or validating Hypotheses.
- When a Themes row reaches 3+ Conversations evidence, surface in Friday memo: "Theme '$theme' now at $n conversations. Ready to promote to Hypothesis? Or accept as validated?"
- Every Friday (or "draft friday memo"), read week's new Conversations + theme deltas + hypothesis transitions. Draft Friday memo doc, section "Week of $monday-date". Structure: (1) New conversations, (2) New / changed themes, (3) Hypotheses moved (state transitions), (4) Threads worth pulling next week.
- At month-end (or "month retro"), read 4 weeks of Conversations + Themes evolution + Hypotheses statuses. Draft Month retro: themes that emerged, themes that died, hypotheses validated / killed, language drift (what changed in how customers describe their problems).
- End of every working session, write 1-paragraph Status note: what you synthesized, what's pending, what to pick up next.
Don't touch:
- Themes.confidence (user assigns).
- Hypotheses.status (user decides, you propose).
- Conversations quotes (user transcribes from real conversations).
First MCP tool calls:
1. list_surfaces(workspace_slug="customer-discovery-pipeline")
2. list_rows(workspace_slug="customer-discovery-pipeline", surface_slug="conversations")
3. list_rows(workspace_slug="customer-discovery-pipeline", surface_slug="themes")
4. get_doc(workspace_slug="customer-discovery-pipeline", surface_slug="status")
Top to bottom. Each step has tasks, pointers, gotchas.
01 / 04
Seed Hypotheses from your current roadmap
1-2 hr
Your roadmap is making bets about what users want. Make those bets explicit. For each thing your team is building or considering, write a Hypotheses row: 'We believe $audience needs $thing because $reasoning. We'll know we're right when $signal.' This is the spine the discovery pipeline tests against.
Tasks
List 5-10 current product bets
For each, write a Hypotheses row in 'we believe / because / we'll know' format
Status: 'untested' for new bets, 'validating' for ones with some evidence, 'validated' for confirmed, 'killed' for shown wrong
Owner column: who's the bet's champion
Gotchas
Hypotheses should be falsifiable. 'Users want a better UI' isn't; 'Users will pay $20/mo more for $specific_feature' is.
Don't have 30 hypotheses. 5-10 is the working set; more than that means you don't have priority and nothing gets tested.
02 / 04
Seed Themes with what you've already heard
1 hr
Before the cycle starts, populate Themes with the 5-10 patterns your team has already noticed across past conversations. This gives your agent priors to work with. Themes should be concrete language users actually use, not internal labels.
Tasks
List 5-10 patterns your team has heard 3+ times in past conversations
For each: write the Theme name in user-language (not internal jargon)
Link 2-3 example conversations (paste quotes if no rows yet)
Tag relevance: which Hypotheses does this Theme bear on?
Gotchas
Don't write Themes from your team's worldview. 'Users want better activation' is your language; 'I tried to invite my team and got stuck' is theirs.
Themes need conversation evidence. If you can't link 3 conversations, it's not a theme yet, it's a hunch.
03 / 04
Walk one week with your agent
Within a real week of conversations
Run the full week-1 cycle: schedule 3-5 conversations, let the agent prep each one, log the conversations, let the agent synthesize, read the Friday memo. The first week calibrates the agent's voice (too academic, too credulous, too noise-prone).
Tasks
Schedule 3-5 conversations for the week
Before each, ask agent: 'Prep this conversation' — confirm the 3 questions are good
After each, fill the Conversations row: 3 best quotes + sentiment + your one-line takeaway
Friday: ask agent: 'Draft this Friday's memo'
Review: are the themes the agent suggested real or noisy?
Tune the agent's tagging conservatism if needed
Gotchas
The agent will over-tag in week 1. Push back: 'this quote doesn't really match that theme.' Calibration takes 2-3 weeks.
Don't promote themes too fast. 3 evidence is the floor, not the ceiling. Some themes need 8-10 conversations before you'd act on them.
Agent prompt for this step
Read Conversations rows added this week. Read Themes table.
For each new Conversations row:
- Read the 3 quotes + sentiment.
- For each existing Themes row, decide: does any quote align with this theme? (Confidence > 70%.) If yes, append a link in Themes.evidence.
- If you see a pattern across 2+ conversations this week that doesn't match any Theme, draft a candidate Themes row with candidate=true + linked conversations.
Identify Themes that now have 3+ evidence: mark "ready for review" in Themes.notes.
Read Hypotheses table. For each, check: do this week's Conversations + theme updates change confidence? Propose status transitions in the Friday memo (as proposals, not changes).
Draft Friday memo section "Week of {monday_date}". Structure:
1. **New conversations** (count + 1 line per row)
2. **New / changed themes** (candidate themes drafted + themes that gained evidence)
3. **Themes ready for review** (3+ evidence threshold crossed)
4. **Hypotheses moves** (proposals for status changes, with evidence)
5. **Threads worth pulling next week** (3 specific questions to ask in upcoming conversations)
Keep it under 300 words.
04 / 04
Run the month retro + fork
1 hr drafting + 30 min meeting + 30 min forking
Last working day of the month, ask the agent to draft the Month retro. The honest data: which themes emerged this month, which died, which hypotheses we tested, what changed in customer language. This is the calibration that keeps your team's product worldview anchored to current users instead of memories of past ones.
Tasks
Ask agent: 'Draft Month retro'
Review the language-drift section especially: is what customers say now different from a month ago?
Decide hypothesis statuses with evidence
Fork workspace, name 'Customer discovery, $next-month'
Confirm fork carried Themes + active Hypotheses + seeded Status with retro
Gotchas
Don't kill hypotheses on a single month of evidence. Some bets need 2-3 months of conversations to fairly evaluate.
Language drift is the most under-used signal. If customers stop using a word that was central to your positioning, your positioning is drifting too. Pay attention.
Agent prompt for this step
Read 4 weeks of Conversations rows. Read Themes evolution (Themes added, themes promoted from candidate, themes deprecated). Read Hypotheses status transitions this month.
Compute:
- Conversations count this month, breakdown by interviewee_role
- Themes added (count, list with conversations-evidence count each)
- Themes deprecated (count, with reason)
- Hypotheses statuses now vs. start-of-month
- Language drift: phrases that appear in this month's quotes that didn't appear last month (compare top-50 phrases by frequency)
Draft Month retro doc:
1. **The month in numbers** (conversations, new themes, hypotheses moved)
2. **Themes** (added + deprecated + promoted, with 1-line per)
3. **Hypotheses scorecard** (table: hypothesis | start status | end status | evidence delta)
4. **Language drift** (3-5 phrases that changed in frequency, hypotheses for why)
5. **Proposals for next month** (top 3 questions to test, top 3 interviewee roles to seek)
Stop before "Decisions made". The founder makes those at the retro meeting.
FAQ
Common questions on this template.
How is this different from `run-50-customer-interviews`?
That template is a one-off sprint: ship 50 interviews in 3 weeks, build the synthesis at the end. This template is the operating cycle after the sprint, when you're at 3-8 conversations per week ongoing. The sprint produces the seed Themes + Hypotheses; this cycle keeps them current. Most teams need both: the sprint at the beginning, the cycle after.
How many conversations a week is the right number?
3-8 is the sustainable cadence for a founder doing discovery alongside building. Below 3 the signal is too thin; above 8 you stop having time to synthesize what you heard. If you need more volume, hire a researcher; the agent can help with synthesis but it can't conduct conversations.
How do I handle conversations that aren't recorded?
Paste the 3 best quotes you remember + sentiment into the row right after the call. Memory degrades fast: 30 min after the call you remember 80%, 4 hours later 40%. Build the discipline of logging within the same calendar hour as the call.
Can the agent run the conversations for me?
No. Agents are bad at conversation: they don't pause, they don't read tone, they don't lean into surprise. The agent's job is to prep (3 questions before the call) and synthesize (themes after the call). The conversation itself is a high-bandwidth human task.
What if customers use vocabulary I don't understand?
Quote them verbatim, even when the language sounds weird. Customer vocabulary is the input to your positioning; if you translate it into your team's words, the signal dies on the way in. The agent surfaces frequency shifts in customer language each month so you can spot when their words drift.
Open this template as a workspace.
We mint a fresh copy in your org with the steps as table rows, the pointers as a separate table, and the brief as a doc. Bring your agents, start checking off boxes.