Campaign copy and SEO briefs still default to Opus. Routing routine drafts to Sonnet would cut spend without hurting quality.
Prospect research runs on frontier models for simple firmographic lookups. A routing rule can keep Opus for complex deals only.
Finance seats still call unapproved consumer endpoints. Enforce the allowlisted provider set for all forecast and close workflows.
Offer and screening chats occasionally hit shadow providers. Lock HR to the governance-approved provider list.
Three MCP servers are enabled org-wide but have had zero tool calls from Engineering in 30 days. Turning them off removes idle context and permission surface.
Sales still has enrichment and calendar MCP servers enabled even though outreach workflows no longer use them.
Okta and asset-inventory MCP servers are enabled for IT but almost never called outside quarterly audits.
Brand asset and ad-platform MCP connectors stay enabled despite campaign work moving to a different stack.
NetSuite query MCP is enabled for all Finance seats but only two analysts used it last quarter.
Engineers paste the same review checklist into chats several times a week. Capturing it as a skill would cut prompt tokens and keep review quality consistent.
SDRs reuse nearly identical prospect research and draft prompts. A shared skill would standardize tone and reduce Opus usage on routine outreach.
These skills have not been invoked in 60+ days and overlap newer shared skills. Removing them reduces skill picker noise and accidental use.
Older competitive-research and release-note skills are unused since the new roadmap briefs skill shipped.
On-call engineers rewrite the same triage prompt for every page. A shared skill would standardize severity scoring and cut repeat context.
Finance analysts restate the same variance narrative structure each close. A skill would keep board-ready language consistent.
Launch-checklist and webinar skills have not been used since the new campaign copy skill shipped.
Older escalation and churn-save skills overlap the new support playbooks and have gone unused.
Legacy interview-score and benefits-FAQ skills are unused after the new hiring screens workflow.
jira.search fails ~18% of the time in PR triage chats, forcing retries that burn tokens and stall reviews.
zendesk.draftReply times out on long tickets, so agents re-prompt and duplicate work.
Several engineers ship multi-thousand-line user MD files that load on every turn. Splitting into task-specific snippets would cut baseline context cost.
Campaign briefs and brand voice live in one large MD file attached to every draft. Smaller per-channel files would keep prompts focused.
Several PMs keep interview archives inside always-on user MD. Moving transcripts to on-demand files would cut baseline tokens.
Reply macros and escalation trees load on every CS chat. Scoping macros to ticket type would shrink default context.
These engineers are among the top spenders but have no user.md, so every session re-learns preferences and project conventions from scratch.
Several SDRs draft outreach without a user.md, so tone and CTA rules get restated in every prompt instead of loading once.
Several overnight debug chats exceeded 200 turns with little progress. Auto-pause after a turn or spend threshold would stop the bleed.
Incident threads sometimes continue long after the SEV is closed. Flag and pause sessions that keep spending with no new alerts.
Call-note synthesis still runs on Opus for most SDRs. Send a one-pager or download it to share with the team.
Debugging sessions often paste full trees instead of the failing package. Share this one-pager with the team.
Product still over-indexes on bulk Drive context for PRDs. Send or download this one-pager for the team.
Forecast chats still include full CSV exports. Share this one-pager before close.
Incidents often restart from blank prompts. Send this one-pager to on-call.
Offer drafts still pull entire policy packs. Download or email this one-pager to recruiting.