matteotitta /revops-incident-response — Decide which leak to fix first
/revops measures the plumbing and stops. This decides what to do about it, in what order, by when.
The unit rule, because it's easy to get wrong: evidence is per-lead, the verdict is systemic. A cohort stat ("483 companies died at intro→trial") names a leak but can't explain it — and an explanation invented from a cohort stat is exactly what gets invalidated on review. Tracing twenty named leads individually tells you why, which tells you what to fix. You ship a systemic decision backed by named leads, not a list of leads to chase.
Workflow at a glance
| Phase | Purpose | Output |
|---|---|---|
| 0. Data-quality gate | Is the CRM trustworthy enough to draw a conclusion from? | PASS / PASS_WITH_CAVEATS / HALT |
| 1. Detection | Which named leads are stalled beyond SLA, per stage? | Named leads — never cohorts |
| 2. Investigation | Why did each one stall? | Per-lead root cause, each citing a quote |
| 3. Triage | Which leak first, how long, in what order? | Branch + fix-time + Day-0/1/14 |
| 4. Hand-off | Who executes? | Named skill + a history.md entry |
Phase 0 can end the run. That is a feature, not a failure mode — see below.
Phase 0 — The data-quality gate (runs first, can halt)
Never trust a leak number before you've checked whether the CRM can support one.
This exists because it already went wrong. ClientCo's 0226-funnel-analysis.md shipped a confident H2-decline story attributed to a named rep's departure. Team review found she worked in CS, not sales, and her attributed deals were likely CRM reassignments. The same doc flags "57% of deals have no assigned rep" — then corrects itself: "this is likely a CRM data quality issue (no automated self-serve deal creation exists), not evidence of self-serve dominance." Its own priority action reads: "Before requesting rep-level data, audit HubSpot deal lifecycle practices first."
That is this phase, written by the client's own analysis, in advance.
Check five things (detail + thresholds → the premium reference):
- Unassigned deals — what share have no owner? Is that real self-serve or missing automation?
- Defunct stages — is any stage still in the schema but dead in practice? (ClientCo: "the MQL stage is effectively defunct since the HubSpot migration".)
- Stage-skipping — deals jumping stages mean stage timestamps can't carry a stall calculation.
- Attribution integrity — are owners assigned by role, or by reassignment artifact?
- Definition drift across systems — the same word meaning different things in two tools. (ClientCo: "'MQL' means different things across HubSpot and Triple Dart".)
Verdicts.
PASS→ proceed to Phase 1.PASS_WITH_CAVEATS→ proceed, but every downstream verdict carries the named caveat inline. Not a footnote.HALT→ stop. Emit no leak verdict. The finding is the data problem, and the deliverable is the fix-list for it. Hand back to/revops.
A HALT is a successful run. Shipping a confident leak verdict on untrustworthy data is the failure this skill exists to prevent — and it has already happened once on this client.
Phase 1 — Detection (named leads, never cohorts)
Read the live CRM. Query stage-entry timestamps and find leads sitting past their stage SLA.
Output shape: a named lead, its stage, days stalled, owner, and deal context. Twenty named rows beat a percentage. If the answer comes back as "59.3% converted," you've reproduced the audit, not triaged it.
Live read via HubSpot MCP (query_crm_data, search_crm_objects, get_crm_objects). Read-only — never write to a client CRM. This is the first skill in the workspace declaring mcps_used: [hubspot]; access is viewer-level per ClientCo/goals/measurement.md.
Redact before you reason: end-client names, emails, phones, account numbers → [CLIENT-n] per pii-redaction.md. ClientCo is FCA-regulated and their CRM carries end-client financial detail. Keep the roles, the company, the deal shape — mask the identity.
Phase 2 — Investigation (root cause, quoted)
For each stalled lead, trace across systems and answer why this one.
Sources: CRM stage history (what moved, when, and what didn't) + call transcripts. Parse transcripts with the existing adapter spec at `research/win-loss/the premium reference — WebVTT / SRT / Gong / Fireflies / Otter / Grain / generic JSON / plaintext. Don't rebuild it. (Granola arrives via MCP as a delivery source, not a sniffed format.)
Every root cause cites a verbatim quote + speaker, per evidence-bound-outputs.md. No quote → mark [INFERRED: from X + Y] or drop the claim. Never invent the reason a deal stalled.
[CLIENT-7] · Intro → Trial · stalled 34d · owner: [REP]
Root cause: pricing, not fit — "we'd need sign-off above 20k and I can't get that this quarter"
(buyer, [11:04]). Trial never provisioned.
Do not screenshot the journey. Zapier's system reconstructed lead journeys via browser automation because their security team blocked MCP connectors. That was their constraint; HubSpot is a live connector here. MCP reads beat screenshots — importing their workaround would import their handicap.
Phase 3 — Triage (branch → fix-time → Day-0/1/14)
Aggregate the per-lead causes, then decide. Five branches — full trees, thresholds, and fix-time estimates → the premium reference:
| Branch | Fires when the per-lead causes cluster on… | First action |
|---|---|---|
| 1. Qualification leak | Wrong-fit leads reaching a stage they can't clear (pricing shock, size mismatch) | Add the qualification gate one stage earlier |
| 2. No-nurture leak | Right-fit leads going quiet with no follow-up mechanism | Ship the missing sequence for the largest cohort |
| 3. Handoff leak | Stalls at an owner change or an SLA with no tracking | Instrument the SLA before changing it |
| 4. Product/proof leak | A recurring objection engineering or marketing must answer | Route the objection, don't re-sell it |
| 5. Phantom leak | The "leak" is a reporting artifact | Back to /revops — this isn't a funnel problem |
State the volume floor. Per quantitative-evidence-floors.md: no "biggest leak" verdict below the floor. Under it, give a directional read and say so. Name the confounds — seasonality, attribution window, conversion lag — before crediting a cause.
Multi-branch: when causes split across branches, rank by recoverable volume × inverse fix-time, and say what you're not doing.
Phase 4 — Reporting + hand-off
Compose, don't build. Recurrence via /schedule; the weekly roll-up via /dashboard, which ships a pipeline-visualization template. No new cron infrastructure — that layer exists.
Hand off by branch: nurture gap → /email-nurture or /lifecycle · qualification gap → /funnel-strategy · plumbing/reporting → back to /revops. Append a [REVOPS] line to the client's history.md.
Anti-Hallucination Guardrails
- Never fabricate a leak number. Live CRM read or nothing. If the MCP is unavailable, say so and stop — per the
financial-data.mdposture, do not reconstruct pipeline data from memory or from a stale doc. - Never invent a root cause. Quote or lower the confidence.
- Never skip Phase 0. The gate exists because the failure already happened on this client.
- Never write to the CRM. Read-only, always.
- Never crown a leak below the volume floor. Directional read + caveat instead.
- Never ship an unredacted end-client name. FCA-bound data.
Integration with other skills
| Skill | Relationship |
|---|---|
/revops |
The audit half. It measures the plumbing; this decides. A HALT hands straight back to it. |
/win-loss |
Explains closed-lost deals with transcripts. This explains silently-stalled leads. Different populations; its transcript adapters are reused here. |
/lead-scoring |
Prospective — should we pursue this account? This is retrospective — why did this one die? |
/funnel-strategy |
Defines the stages and SLAs this measures stalls against. Hand-off target for Branch 1. |
/email-nurture · /lifecycle |
Hand-off targets for Branch 2. |
Quality checks (pre-output)
Full gate → the premium reference. The load-bearing five:
- Phase 0 emitted an explicit verdict before any leak claim. A HALT emitted no leak verdict at all.
- Detection names actual leads. If the output is a percentage, it's an audit — start over.
- Every root cause carries a verbatim quote + speaker, or is marked
[INFERRED]/ dropped. - The volume floor is stated; confounds and lag are named.
- End-client PII masked. Nothing written to the CRM.
Final ship gate
Run /premortem --output before ship. See /premortem skill for the 5 execution domains and output template.
Then run /scope-guardian-reviewer — the client-deliverable ship gate: scope-creep check on proposals/SOWs (pm-loop.md). Live-CRM findings invite scope sprawl; this is the gate that catches "while we're in here…".
Trivial-case escape: ## Premortem\nNo failure modes — trivial change satisfies the contract for genuinely trivial outputs.
Attribution
The Detection → Investigation → Reporting shape, and specifically the per-lead granularity + recurrence that our cohort-level audit thinking wouldn't have produced, is adapted from Zapier's internal lead-leakage agent system as described in "Zapier's no lead left behind agent", Demand Collective, accessed 2026-07-16. Cite-only — concept port, zero code. Their browser-automation implementation is deliberately not ported (see Phase 2). See .claude/discovery/0726-zapier-lead-leakage-steal-analysis.md.