Paid engagement (~3 months left as of 2026-06-30) using Ghostteam’s LinkedIn/social signal tracker (clusters LinkedIn posts by category/topic, surfaces top posts) — this is the earlier LinkedIn-signal product, distinct from the ADO discovery platform. Bridge in = roslyn (a Davos-network contact at Noetic).
Champion: ruth-manpower (CMO, ManpowerGroup). The 2026-06-26 pitch meeting (Ruth, Roslyn, Gualin = Head of Europe, Roseanne) landed well — clear value seen. Main steer: make it practical/useful for Ruth first, before any wider rollout.
Agreed deliverable: a weekly digest email. Ruth won’t log in regularly; instead a Thursday email with top categories + top posts goes to her, and Roslyn sits with Ruth on Fridays to write LinkedIn posts off it. The real primary users are the Ghostteam team (we digest + surface insights; Ruth can also log in).
Commercial goal: upsell the MCP app. The signal-tracker engagement is “good cash” but the real prize is selling the MCP app — Elliot to follow up and create urgency.
Engagement risk:roslyn is less responsive now (focused on non-AI work at Noetic, which is selling well) — no hard feelings; she stays a valuable network contact (she + husband John know relevant people; Elliot will lean on her for investor/adviser intros).
The tracker can’t be build-once-forget (operational truth): it needs review ~twice a week. Safer to over-cluster and filter with cluster-level rules (filter by topic, merge clusters; the “not relevant” log informs rule design) than to under-cluster and miss signals. Monitoring should be outcome-based (“did we get clusters today? alert if zero for 2 days”) — a generic “is everything working?” check reports success even when the model silently fails.
Signal-quality problem + the two filters (2026-07-01): clusters are still dominated by low-quality bot posts from India/Pakistan and near-duplicate posts (the same hiring event / text+image reposted by many accounts to fake engagement). Two agreed filters: (1) exclude by author location — there’s no location on the post, so scrape the poster’s profile and drop India/Pakistan accounts (only posts, not commenters — commenters can be anyone); (2) drop near-duplicates — same hiring-event content posted by 4+ accounts. “Filter out bot posts” alone is too vague; needs concrete patterns.
People-coverage gap (2026-07-01): only ~12 individuals tracked (the rest are companies, which post noisy/boring content), so signal leans too heavily on keywords. Want ~200 credible HR / C-level people for better signal — ask roslyn’s contact who supplied the original list to expand it, likely plus a keyword session with ruth-manpower.
Smarter-analysis direction (2026-07-01): rather than only rule-based filtering, connect the cluster data + inputs to Codex / Claude Code as a skill for a human-in-the-loop ~2x/week analysis that iterates on what’s useful and drafts the weekly email to Ruth (top signals + a few most-relevant credible posts with short synopses + clickable links; simple like the Telegram research report — she won’t read the dashboard, and writes a LinkedIn post every Friday). Run manually (~1 hr/week) until stable, then optionally schedule.
Open questions
Will ManpowerGroup buy the MCP app (the real commercial goal)?
Full names/roles for Gualin (Head of Europe) and Roseanne.
Timeline
[2026-07-01] (internal — Ghost Team stand-up, Vincent + Elliot, recorded by Vincent) Worked the tracker’s signal-quality problem: agreed the two filters (location-scrape to exclude India/Pakistan authors; drop 4+-account near-duplicates), the Codex/Claude-Code skill for smarter 2x/week analysis + weekly Ruth email, and the 12→200 credible-people coverage gap (see Current truth). Plan: credentials to be sent, first test today, second tomorrow, then a first weekly-email iteration; Elliot to message Roslyn’s contact to expand the tracked-people list and likely run a keyword session with Ruth.
[2026-06-30] (internal discussion — Vincent + Elliot, recorded by Vincent) Debriefed Elliot’s ManpowerGroup pitch meeting (Ruth/Roslyn/Gualin/Roseanne): value landed; steer = useful-for-Ruth-first. Locked the weekly Thursday digest deliverable (top categories/posts → Roslyn’s Friday post sessions with Ruth; primary users = us). Commercial goal reaffirmed = upsell the MCP app (Elliot to create urgency). Flagged Roslyn’s reduced responsiveness (non-AI focus). On the tracker: a cluster bug (only 4 clusters) traced to Anthropic decommissioning the old Sonnet used for cluster analysis (fix = bump to newer Sonnet); agreed the system needs ~2x/week review + cluster-rules + outcome-based monitoring.
[2026-06-16] (doc) Added as a Q2 re-engagement target (open a ChatGPT-App / MCP discussion) in the quarterly goals — see pipeline.