Imported from amplitude/mcp-marketplace (
plugins/amplitude/skills/user-cohort-forensics/SKILL.md). Install upstream withnpx skills add amplitude/mcp-marketplace --skill user-cohort-forensics. Copyright stays with the author.
User & Cohort Forensics
The arcs
Single user deep-dive: get_amp_user_data include: 'id' (resolve the
user: email, user ID, device ID, or amplitude ID) → include: 'profile'
(lifecycle, acquisition, usage stats) → include: 'timeline' (session-aware
event history). include: 'both' gets profile + timeline in one call when
you know you'll need both. Summarize journeys; don't dump raw events.
Population analysis (batched): build the user set first
(query_amplitude_data with a user-ID group_by to rank by volume, or
use_amplitude_cohorts action: 'find' for a filtered set) →
get_amp_user_data include: 'timeline' in parallel batches — up to 10
identifiers per call, 10–20 calls in flight is normal for population
analysis; hundreds of calls total is fine. Keep each call narrow (event
types, window) so responses stay small.
Cohort spot-check: prefer existing cohorts — use_amplitude_cohorts
action: 'list' or action: 'get' before building ad
hoc definitions. action: 'membership' verifies specific users. Check a
member's timeline for the exact markers (purchase, typing, checkout events)
rather than trusting the cohort definition blindly.
Email → ID resolution: get_amp_user_data include: 'id' per email;
uploaded email lists are resolved in batches of ≤10 identifiers per call.
For bulk exports, batch and note the retry pattern on individual failures.
Parameterization notes
include: 'timeline': always bound the window (last 30 days by default) and pass event-type filters when you know what you're looking for — unfiltered timelines are large and slow.- Rate-limit failures come back flagged
retryablewithretryAfterMs— back off instead of churning. - User identity: a user can match multiple IDs (device, user, email). Say which identity you resolved and flag ambiguous matches instead of picking one silently.
