win-loss-analysis — Day 4 research skill

The Example 1 Day 4 skill. Reads your won + lost sales-call transcripts and writes a pattern-level analysis to marketing/win-loss/win-loss.md. This is the evidence base under positioning, messaging, and ICP — every claim in those docs should trace back to something a real buyer said here.

---

When to use

• Day 4 of Example 1: before /icp-research + /positioning, so the strategy reads from real buyer language
• You have a fresh batch of 5+ won/lost call transcripts
• A quarter closes and you want to refresh why deals moved
• Churn spikes and you need the pattern, not the anecdote

When NOT to use

• For a single account's interview prep (use /customer-interviews — not in V2 quickstart)
• For behavioural simulation of a buyer (use /icp-behavioural — not in V2 quickstart)
• When you have fewer than ~5 transcripts — patterns need volume; below that you get anecdotes, not signal

Two rules apply before any analysis

• Redact PII first — .claude/rules/pii-redaction.md. Mask end-customer names, emails, and account numbers before processing; keep roles, company, and deal context.
• Bind every claim to evidence — .claude/rules/evidence-bound-outputs.md. Every pattern cites a verbatim quote + speaker, or it lowers its confidence. No invented quotes.

How it works

1. Inputs: sales-call transcripts (Gong, Fireflies, Otter, Granola, Zoom/Avoma VTT, or pasted text), each tagged with the deal outcome (won / lost / churned). Optional: marketing/icp/ICP.md to frame patterns by segment.
2. Normalize each transcript to speaker-attributed turns with timestamps where present.
3. Pick a mode:
   • Single — deep analysis of one transcript
   • Batch (default) — aggregate 5–20 transcripts into patterns with frequency counts
   • Comparison — won vs lost (or retained vs churned) side by side
4. Extract patterns across six dimensions: product, messaging, GTM / sales process, pricing, competition, customer context.
5. Score confidence by frequency: a pattern needs 2+ occurrences across different deals; High = 3+ deals, Medium = 2, Low = single mention. Aim for ≥5 wins and ≥5 losses before trusting a pattern.
6. Writes to marketing/win-loss/win-loss.md (overwrites prior canonical; git history preserves prior versions).

Invoke

/win-loss-analysis

Then paste or point to the transcripts and tag each outcome. Or:

/win-loss-analysis — here are 8 won + 6 lost transcripts: [paste / paths]

Example output

See marketing/win-loss/win-loss.md for the PulseAnalytics example seed. Notice: patterns are grouped by dimension; each carries a frequency + a verbatim quote with speaker; the closing section routes findings to positioning / messaging / ICP.

Dependencies

• Reads from: sales-call transcripts (required); marketing/icp/ICP.md (optional, for segment framing)
• Reads via Granola MCP (optional): meeting transcripts, if wired
• Writes to: marketing/win-loss/win-loss.md (canonical; positioning, messaging, and ICP read from here)

Customization

Split the analysis by segment when your ICP has more than one (enterprise vs mid-market lose for different reasons). Add a competitor column to the competition dimension once you're losing to a named rival repeatedly — that feeds /competitor-research.

Where this fits in the Example 1 chain

Day 1-3: /competitor-research × N → per-competitor files
Day 3:   /competitor-aggregate → competitor canonical
Day 4:   /win-loss-analysis (THIS SKILL) → win-loss canonical
Day 5:   /icp-research reads win-loss + competitors → canonical ICP
Week 2:  /positioning + /product-messaging read win-loss quotes for real buyer language

Refresh cadence

Monthly if deal volume is high; quarterly otherwise. Refresh sooner on a churn spike or a new competitor showing up repeatedly in lost deals.