HG Insights|Phoenix

Skill: CSM Churn Signal Weighting

Watch / Yellow / Red verdicts your CSMs can defend — anchored on the HG signals their book actually shows.

Overview

Teach Claude to read HG signals like a CSM screening a book for churn. The skill carries the 6-signal taxonomy (adoption decline, location snapshot, competitor adds, intent decay, champion coverage, financial stress), the recency windows that govern each, and the Watch/Yellow/Red weighting rubric — so churn-prediction workflows compose verdicts on the same heuristic instead of re-deriving it inline.

Use cases

  • Monday morning screening that beats gut-feel triage

    Your CSMs scan 20 accounts and "feel" 3 are at risk. This skill grounds the verdict on signals every CSM and exec can defend: a 35% adoption drop in the last 6 months, a competitor added 45 days ago, intent decay on your category. Watch/Yellow/Red, with the top 3 signals named.

  • Same heuristic, every workflow

    Six churn-prediction-adjacent workflows (#1, #2, #3, #4, #10, #14) compose this skill so the threshold for Yellow vs Red is identical across every output — no per-workflow drift in the rubric.

View full skill

CSM Churn Signal Weighting

When to use

  • A workflow is composing a Watch/Yellow/Red verdict on a customer's churn risk.
  • A workflow is asked "is this account healthy?" and needs a defensible heuristic.
  • An author is about to inline weighting rules into a churn-adjacent workflow — stop, reference this skill.

Tools you'll touch

  • company_install_time_series — adoption intensity over a closed enum window (timeRange: last_6_months / last_12_months / last_24_months / last_36_months).
  • company_technographic — current-snapshot tech footprint with firstVerifiedDate and productLocations per row (camelCase MCP surface).
  • company_intent — top-N topics per company (default limit: 50, max 200) with trend. Endpoint is extra: "forbid" — filter post-fetch.
  • company_operating_signals — 6-month deltas (cloud_depth_6m_delta, ai_maturity_6m_delta, etc.).
  • contact_search — verify the named buying committee.
  • sec_filing_section — 10-K Item 1A/7A for financial-stress signals (public companies only).
  • list_intent_topics / intent_category — resolve category → topic IDs for the post-fetch filter.

Out of scope: company_contracts. That tool returns ICT outsourced contracts (GSI-mediated — Accenture, IBM, Cognizant), NOT the customer's own product contract or renewal date (company-contracts.ts:233-234). Any workflow that needs renewal proximity must take our_contract_renewal_date as a CRM-supplied parameter.

The 6 signal classes

#SignalTool / fieldRecency
1Adoption intensity declinecompany_install_time_series.intensity_momentum and current_intensity vs lookback-startwindow-driven
2Location-count snapshotcompany_technographic.productLocations (current only — NOT a delta)snapshot
3Competitor detectioncompany_technographic.firstVerifiedDate for net-new products in customer's categorywithin lookback
4Intent shiftcompany_intent.topics[].trend (post-fetch filter by topic_id)over lookback
5Champion coveragecontact_search against {{prior_contacts}} baselinecurrent
6Financial stresssec_filing_section 10-K Item 1A/7Alast 90d

1. Adoption intensity decline

intensity_momentum is negative when adoption is declining. Treat as a decline signal when:

  • intensity_momentum < -10 (strong negative)
  • AND/OR current_intensity is down ≥20% vs the first data_points entry in the lookback.

data_points[] returns { date, intensity } per month — no location count is included. Do not claim a location-count time-series from this tool.

2. Location-count snapshot

productLocations is a current count, NOT a delta. The skill flags a location decline only when the workflow has a baseline:

  • A prior_location_count parameter (CRM-supplied), OR
  • A captured value from a prior run.

Workflows that try to derive a 6-month location delta from install_time_series.data_points are wrong — that field carries intensity, not location count.

3. Competitor detection

Compute the customer's category from their own product ({{your_product}}), then filter company_technographic rows where firstVerifiedDate falls inside the lookback AND the row's category overlaps. Each surfaced row is a competitor-add signal. Cite firstVerifiedDate per row.

4. Intent shift (decay or rise)

company_intent returns top-N topics by score per company (default 50, max 200). Workflow shape:

  1. Resolve the customer's category → canonical topic_ids via list_intent_topics.
  2. Call company_intent with limit: 200 for the domain.
  3. Post-fetch filter topics[] for the resolved topic_ids.
  4. Check trend:
    • trend ≤ -30% on a topic matching the customer's category → decay signal.
    • trend ≥ +30% on a topic matching a competitor's category → rise signal.
  5. If the resolved topic_id is absent from the returned topics[], treat as inconclusive (NOT "no decay"). Fall back to intent_category with the specific topic_name (returns company-list for the topic; filter the result to {{domain}}).

5. Champion coverage

contact_search against {{prior_contacts}} (CRM-supplied buying committee). If prior_contacts is empty, skip absence inference — the workflow cannot reliably distinguish "champion churned" from "the search didn't find them".

6. Financial stress

sec_filing_section 10-K Item 1A (Risk Factors) and Item 7A (Quantitative Disclosures). Layoffs ≥5% workforce, guidance misses, going-concern flags, restructuring — all signal stress. Public companies only.

Recency / decay windows

Defer to the HG Recent Signals skill for the canonical thresholds. CSM-specific additions:

  • 30/60/90-day adoption deltas — install_time_series covers this via the timeRange enum (last_6_months is the shortest available window; for finer-grained periods, work from the data_points series directly).
  • Location-count baselines decay quickly — a prior_location_count older than 6 months is unreliable.

Weighting heuristics — Watch / Yellow / Red

Composing the verdict combines signal strength + signal count. The skill is the source of truth for the rubric.

Signal patternVerdict
0-1 signals firing, or only weak signals (intensity_momentum between -1 and -10)Watch (green)
1 strong signal OR 2-3 weak signalsYellow
2+ strong signals OR (competitor add ≤90d + intent decay ≥30% on our category + adoption intensity_momentum < -10)Red

Combo escalators — any one of these escalates by one band:

  • New competitor with productLocations ≥3 in the customer's geo footprint (broad competitor install).
  • 10-K Risk Factor disclosing layoffs ≥5% workforce in the last 90 days.
  • prior_contacts ≥50% absent from current contact_search (champion erosion).

Renewal-cycle awareness — consumers who supply our_contract_renewal_date (CRM) may apply their own escalation rule on top (e.g., "≤120 days to renewal escalates Yellow → Red"). The skill does not require this parameter; renewal proximity is NOT a churn-prediction signal in the core 6.

Output convention

Verdict line: Verdict: <Watch|Yellow|Red> (color-coded).

Top-3 signals: 3 bullets, each with:

  • Signal name (one of the 6 classes).
  • Evidence (specific number, e.g., "intensity_momentum: -18; current_intensity dropped 35% over last_6_months").
  • Source citation per hg-citation-discipline ([company_install_time_series]).

Recommended actions: 3 actions sequenced 30 / 60 / 90 days, each tied to a named signal.

Common pitfalls

  1. Treating company_contracts as renewal-status. It's GSI-outsourcing data. See "Out of scope" above.
  2. Inferring location decline from install_time_series.data_points. That field carries intensity only. Need a baseline parameter.
  3. Treating "topic absent from company_intent response" as "no intent decay". It's inconclusive — fall back to intent_category.
  4. Inferring champion churn without a prior_contacts baseline. Absence inference is unreliable.

Reference