Win Loss

Win/loss analysis

Analyze sales call transcripts to extract actionable insights on why deals are won, lost, retained, or churned. Cross-reference findings with ICP, firmographics, and competitive context to produce strategic recommendations.

Claude Code triggers

Invoke when user says:

• "Win/loss analysis"
• "Analyze sales calls"
• "Why did we win/lose"
• "Churn analysis"
• "Retention analysis"
• "Sales call insights"
• "Deal outcome patterns"
• "Customer feedback synthesis"
• "Analyze these transcripts"
• "What patterns in our sales calls"

Do NOT invoke when:

• User wants general transcript analysis → use transcript-analysis
• User wants competitor research → use competitor-research
• User wants single customer interview analysis → use transcript-analysis
• User wants sales enablement assets → use sales-enablement

Input requirements

Required

Optional (improve quality)

Validation

Before proceeding: at least one transcript provided; outcome known or inferable from transcript; customer name identifiable.

If inputs are missing: ask the user for transcripts. Clarify if outcome should be inferred from transcript signals.

Transcript intake — normalize any recorder format

Transcripts arrive in many shapes: Gong, Fireflies, Otter, Grain exports, Zoom/Avoma VTT, SRT, recorder JSON, or plain pasted text. Before Phase 1, normalize whatever you're handed into one shape — speaker-attributed turns, timestamps where present. See the premium reference for the sniff-and-parse table per format and the normalized target shape. /transcripts inherits the same reference.

Two rules apply to every transcript before analysis:

• Redact PII first — .claude/rules/pii-redaction.md. Mask end-client names, emails, account numbers before processing; keep roles, company, deal context. (Load-bearing for ClientCo and any regulated client.)
• Bind every claim to evidence — .claude/rules/evidence-bound-outputs.md. Every extracted pattern cites a verbatim quote + speaker; normalized turns make speaker attribution reliable.

Process

The analysis runs in 3 phases. Read the premium reference for the full step-by-step (4 transcript-processing steps, 4 aggregation steps, 4 synthesis steps, plus per-phase checkpoints and the process flowchart).

Phase summary:

1. Transcript processing — classify outcome, identify speakers, extract customer context, pull verbatim quotes for the 6 dimensions
2. Pattern aggregation — group by outcome, count frequency, rank patterns (3+ mentions), cross-reference by ICP/competitor/persona
3. Insight synthesis — state pattern, provide evidence with frequency + confidence, identify opportunity, generate executive summary
4. Canonical-doc drift → propose diff (optional — active client, cadenced batch only) — check recurring patterns against the client's current positioning/messaging docs and emit gated diff proposals. See the premium reference.

Cadence

For an active client, run the batch analysis on a monthly cadence — the review cycle that keeps positioning/messaging current between quarterly refreshes. Phase 4 rides that same monthly run; it adds no separate schedule. Single-call and ad-hoc runs skip Phase 4.

Phase 4 — when it runs

Phase 4 fires only on a cadenced batch for an active client with canonical docs (projects/consulting/active/{client}/positioning/ or messaging/). It proposes diffs; it never applies them — the merge is a separate, two-layer human gate (Genesys PMM, then client signoff). Win-loss stays additive; positioning/messaging stay human-locked. Skip Phase 4 silently for single-call runs, comparison-only runs, prospects, or clients with no canonical docs. Full protocol, eligibility (recurrence gate), voice gate, and merge bridge: the premium reference.

Core frameworks

Analysis modes

Default to batch analysis mode when multiple transcripts are provided.

6 analysis dimensions

Confidence scoring

Outcome classification

Anti-hallucination guardrails

1. Quote verbatim. All insights must trace to specific transcript quotes.
2. Never invent patterns. If a pattern appears in only one call, label it "Single mention — pattern unconfirmed."
3. State frequency. Always note how many calls support each finding (e.g., "4 of 7 calls").
4. Acknowledge gaps. If a dimension has no data, mark "Not discussed in transcripts."
5. Distinguish roles. Tag who said what — prospect vs. sales rep vs. champion.

Gotchas

• Correlation as causation. Reports "deals with longer sales cycles were lost" as if cycle length caused the loss → always distinguish patterns from causes. Use "associated with" not "caused by".
• Small sample bias. Draws conclusions from 2-3 deals instead of waiting for sufficient data → flag sample size prominently. Minimum 5 wins and 5 losses for reliable patterns.
• Missing verbatim quotes. Summarizes what buyers said instead of extracting exact quotes → verbatim quotes are the primary deliverable. Summaries are secondary.
• Single-dimension analysis. Only looks at win/loss by competitor, missing dimensions like deal size, ICP segment, or sales cycle stage → cross-tabulate across at least 3 dimensions.
• Conflates product feedback with sales insights. Mixes "they wanted feature X" with "they didn't trust our team" → separate product gaps from sales execution issues. They feed into different downstream skills.

Integration with other skills

MCP data integration

Level: 0 — Context (heavy pulls)

Pulls fresh

Fallback (no MCP)

• User-provided call transcripts or recordings
• Manual deal review notes