Setup & Customization Checklist

1. Data access

The audit is only as good as Step 1. Before the first run, confirm the agent can actually pull each source:

• {{USAGE_DATA_SOURCE}} — workflows/automations with execution counts and enabled state, connected senders and integrations, registered users with last-login dates, and feature-level usage/spend for the past {{LOOKBACK_DAYS}} days. If you can't get execution counts, Section 1's dead-workflow check degrades to guesswork — fix access first.
• {{CRM}} — company activity feed: notes, calls, emails, deal stage history.
• {{CALL_LIBRARY}} — transcript search by account domain. Optional but strongly recommended; it anchors "what are they solving for."
• Account notes/memory — wherever your team records goals and context.

2. Placeholders

• {{USAGE_DATA_SOURCE}}, {{CRM}}, {{CALL_LIBRARY}}, {{TIER_LABELS}}, {{ALERT_DESTINATIONS}}, {{SCORING_WORKFLOW}}, {{VOICE_GOVERNANCE}}, {{STALE_DAYS}}, {{LOOKBACK_DAYS}} — all filled.
• Map {{TIER_LABELS}} to your actual tier names. The 3-tier minimum in Section 2 is the judgment rule; the labels are cosmetic.
• Set {{STALE_DAYS}} to match your product's activity rhythm. Default 5 assumes daily-cadence workflows; weekly-cadence products should use 10–14 or every audit will cry wolf.
• Point {{SCORING_WORKFLOW}} and {{VOICE_GOVERNANCE}} at the account's central workflows, not your org's — the audit inspects the customer's configuration.

3. Policy decisions

• Character caps (750 / 1,000 / 500 per section) — these are the skill's core discipline, forcing signal over narrative. Adjust only if your team consistently needs more; never remove them.
• The 30-account cross-reference before flagging spend waste — decide your sample size (30+ is the tested floor). Flagging waste from raw spend percentages is the most common false alarm this rule exists to prevent.
• Interventions in Section 6 — review the failure-to-intervention map and swap in your own playbook actions (e.g., your onboarding session format, your champion check-in motion) while keeping the one-intervention-per-flag rule.
• Decide who receives the report and where it lands (doc, channel, CRM note). The skill produces the report; routing is yours.

4. Behavioral invariants (do not remove)

• All Step 1 data gathered before any section is written — sections produced from partial data are worthless.
• No "likely," "probably," or "seems" anywhere in the output.
• Tier persistence checked by comparing two counts (tagged vs. sequenced), never by eyeballing tags alone — partial persistence is the failure mode this catches.
• Spend flagged as waste only after the per-account cross-reference, or when no qualification gate exists at all.
• Messaging judged from workflow instructions read word-for-word — never from titles, descriptions, or usage counts.
• Section 3 stops at notification coverage; sequence/outreach state belongs to Section 4. Scope bleed is how audits balloon.
• Zero saved message examples for a connected sender is a hard flag.
• Every flagged finding in Section 6 pairs with exactly one concrete intervention.