Discovery Debrief

Discovery Debrief

Structured extraction after a customer/prospect conversation. Every conversation is data. This skill turns conversations into actionable learnings.

Triggers: "talked to a customer", "customer call", "debrief", "discovery call", "had a call with...", "met with..."

Step 0: Context

1. Read MEMORY.md — current wedge, ICP, constraint
2. Read memory/hypotheses.json — active hypotheses
3. Ask: "Who did you talk to? Tell me in free form."

Step 1: Structured Extraction

After the founder's free-form story — extract structured data. Ask clarifying questions ONE AT A TIME (don't batch).

1.1 Who

• Name, role, company, team/company size
• How the contact was established (inbound/outbound/referral)
• Buyer or user? (often different people)

1.2 Demand Reality (is demand real?)

"Would this person be upset if our product disappeared tomorrow?"

Look for behavioral evidence, not words:

• Paying or willing to pay? How much?
• Using the product? How often?
• Building workflow around the product?
• Would scramble if the product vanished?

Red flags: "interesting", "need to think about it", "show it to my colleagues" — this is politeness, not demand.

1.3 Status Quo (what are they doing now?)

"How do they solve this problem today — even poorly?"

Look for the specific workflow:

• What tools/processes do they use?
• How much time/money do they spend?
• Who does it manually?
• What breaks in the current process?

Red flag: "they don't do anything" -> the problem may not be painful enough.

1.4 Narrowest Wedge (minimum product for money)

"What is the smallest version of the product they would pay for right now?"

Not "the full platform", but one workflow, one integration, one use case.

1.5 Surprise (what was unexpected?)

"What in this conversation didn't match your expectations?"

Surprise = the most valuable signal. If there are no surprises — either wasn't listening or confirming bias.

Gold: customer uses the product in an unintended way -> that might be the real product.

Step 2: Signal Assessment

Assess signal strength using Evidence Tiers (CEO Bible):

Classify: "This conversation yielded tier [X] evidence: [specifically what]"

Step 3: Hypothesis Check

For each active hypothesis from memory/hypotheses.json:

• Does this conversation confirm, refute, or remain neutral?
• If confirms/refutes — what specific evidence?

Relay Race check: is there a hypothesis with enough evidence for a decision RIGHT NOW? If yes -> flag: "H00X has sufficient data. Decide now, don't wait for the deadline."

Step 4: Pattern Detection

Compare with previous conversations (from daily logs):

• Same pain recurring? (repeated pain -> strong signal)
• Same trigger? Same workaround?
• Same buyer profile? Or different people with different problems?

Convergence signal: 3+ conversations with same pain, same trigger, same workaround -> wedge is clarifying. Divergence signal: every customer wants something different -> wedge is blurry, need to narrow.

Step 5: Save

1. Log to memory/YYYY-MM-DD.md via structured-log:

## Discovery: [Company] — [Name, Role]
- **Source:** inbound/outbound/referral
- **Demand:** [tier X] — [evidence]
- **Status quo:** [what they do now]
- **Wedge fit:** [high/medium/low] — [why]
- **Surprise:** [what was unexpected]
- **Hypothesis impact:** [H00X: confirms/weakens/neutral]
- **Next step:** [concrete action]

2. Update memory/hypotheses.json if there is hypothesis impact
3. Update MEMORY.md if any of these change: ICP, wedge, positioning, PMF stage

Step 6: Constraint Nudge

Close the debrief with a reminder:

• "Qualified conversations this week: X out of target 8-12"
• If < 4 -> "Priority: more conversations, less code"
• If the conversation was not with target ICP -> note: "This conversation is outside the current wedge. OK for exploration, but don't count as qualified."

Telegram Format

Discovery: [Company] — [Role]

Demand: Tier [X] — [1 line evidence]
Status quo: [what they do now]
Surprise: [what was unexpected]
Hypothesis: H00X [confirms/weakens]

Conversations: X/8-12 this week
Next: [concrete action]

Follow output preferences from USER.md (language, format, platform constraints).

Rules

• One question at a time. Don't batch.
• Quote exact customer words when possible — customer's words > founder's interpretation.
• Don't assess until all 5 points (1.1-1.5) are collected.
• If the founder retells superficially -> push: "What exactly did they say? In what words?"