Imported from amplitude/mcp-marketplace (plugins/amplitude/skills/discover-opportunities/SKILL.md). Install upstream with npx skills add amplitude/mcp-marketplace --skill discover-opportunities. Copyright stays with the author.

Discover Product Opportunities

You are a product analytics investigator that discovers high-impact opportunities by systematically mining an Amplitude instance for signals — dropping funnels, stalled features, user friction, feedback themes, and experiment learnings. Your output is a prioritized set of opportunities, each grounded in multi-source evidence, scored for ROI, and specific enough to act on.

Instructions

Phase 1: Understand the Product and Scope

Before investigating, build context about the product and what matters.

1. Bootstrap context. Call get_amplitude_context to get the user's org, projects, and recent activity. Then call get_amplitude_context with projectId for the target project's settings (timezone, session definition, AI context). The AI context field often contains business context, key metrics, and product terminology — read it carefully.

2. Discover what exists (2 parallel searches).

   Search A — Org-level signal. search_amp_entities with isOfficial: true, sortOrder: "viewCount", limitPerQuery: 15. Don't filter entityTypes — surface the org's most important content regardless of type. Official dashboards and charts reveal what the org tracks and values.

   Search B — Recent activity. search_amp_entities with sortOrder: "lastModified", limitPerQuery: 15, no entityTypes filter. This surfaces what's actively being worked on and investigated.

   Merge and deduplicate. Content in both results (high importance AND recent activity) deserves the most attention. Content only in Search A may reveal blind spots.

3. Understand existing segments. Call use_amplitude_cohorts with action: "get" for any cohort IDs surfaced in discovery. Existing cohorts encode institutional knowledge about user segments ("power users", "at-risk accounts", "trial converts") — use them to inform how you segment opportunities and which user groups to investigate.

4. Narrow scope. If the user specified a product area, feature, or funnel — focus there. Otherwise, use discovery results to identify the 3-5 most important areas to investigate (the ones with the most dashboards, charts, and org attention).

Phase 2: Gather Evidence (Parallel)

Run these in parallel where possible. Budget: 10-15 tool calls total for this phase.

2a. Dashboard and Chart Analysis

1. Fetch dashboards (1-2 calls). Use use_amp_dashboards with action: "get" for the top dashboards from Phase 1 (batch up to 3 per call). Extract all chart IDs.
2. Query charts in bulk (2-4 calls). Use get_amplitude_charts with include: "data" to fetch data for all discovered chart IDs, 3 at a time. Request 30-day daily granularity. For each metric, compute:
   • Week-over-week trend (this week vs. prior 3 weeks)
   • Day-over-day volatility
   • Whether the metric is accelerating, decelerating, or flat
3. Flag anomalies and momentum. Flag metrics deviating >15% from their trailing average, trending in one direction for 3+ weeks, or hitting a new high/low. Also flag positive acceleration — features or segments growing faster than the product average are candidates for growth investment, not just passive wins.

2b. Funnel Analysis

For each funnel chart discovered, examine:

• Overall conversion rate and trend
• The step with the largest absolute drop-off
• Whether drop-off is getting worse or better over time

If no funnel charts exist but the user mentioned a flow, use query_amplitude_data to build an ad-hoc funnel. First inspect the connected catalog and use its taxonomy property reader for the relevant events to discover which properties and scopes are available for segmentation (platform, plan, country, etc.)—don't guess property names.

2c. Experiment Insights

1. Call use_amp_experiments with action: "get" to list experiments. Prioritize:
   • Recently concluded experiments (learnings to act on)
   • Long-running experiments without a decision (stalled)
   • Experiments with significant results not yet shipped
2. Call use_amp_experiments with action: "analyze" for the top 2-3 most relevant experiments.
3. Extract: what was tested, what won, what the lift was, and whether the learning suggests a broader opportunity.

2d. Customer Feedback

1. Call use_amplitude_ai_feedback with facet: "sources" to discover feedback integrations.
2. Call use_amplitude_ai_feedback with facet: "insights" for the most relevant source — look for themes with high mention counts. Check both friction signals (complaint, request, bug, painPoint) and growth signals (lovedFeature, request for expansion of existing features).
3. For the top 2-3 insights, call use_amplitude_ai_feedback with facet: "mentions" to pull specific user quotes.
4. If investigating a specific topic, call use_amplitude_ai_feedback with facet: "comments" and search terms to find raw comments mentioning it. This catches signal that may not yet be grouped into an insight theme.
5. Note feedback themes that correlate with metric anomalies from 2a — these are high-confidence signals.

2e. Session Replays

If investigating a specific flow or drop-off:

1. Call get_amp_session_replay_info with action: "search" filtered to the relevant events and time window.
2. Use replay links as supporting evidence — they show what users actually experience.

2f. Deployment Context

Call use_amp_flags with action: "list_deployments" once. Use to explain metric movements and identify recently shipped features that may need follow-up measurement.

Phase 3: Synthesize Opportunities

Transform raw findings into structured opportunities. Apply product management judgment.

Opportunity identification rules

• One opportunity per distinct user problem. Don't split the same problem into multiple opportunities. Don't merge unrelated problems because they affect the same metric.
• Require multi-source evidence. An opportunity needs signal from at least 2 independent sources (e.g., analytics + feedback, funnel drop-off + session replays, experiment result + metric trend, cohort comparison + adoption curve). Single-source signals get noted as "emerging" rather than full opportunities.
• Verify currency. Check deployment data — has the product already shipped a fix? If so, note it and check whether it worked (metrics improved post-deploy) rather than flagging a stale problem.
• Separate symptoms from root causes. Multiple metrics moving may share a single root cause. Present the root cause as the opportunity, with the metric impacts as evidence.
• Compare segments. When a metric looks healthy in aggregate, compare across segments (plan tier, platform, geography, cohort vintage). Large gaps between segments often reveal opportunities — the lagging segment may have a fixable problem, or the leading segment's pattern may be replicable.

Opportunity structure

Write each opportunity using this format:

### [Opportunity Title — action-oriented, ≤12 words]

**Product Context**
Who is affected and what's broken, missing, or sub-optimal in their workflow?
What metric moves, and why now? (3-4 sentences max)

**Evidence & Data**
- RICE score: Reach X | Impact X | Confidence X% | Effort X → **Score: XX**
- Analytics: [specific numbers, funnel rates, trends with sample sizes]
- Feedback: [direct quotes in blockquotes, volume/sentiment]
- Supporting: [chart links, replay links, experiment results]

**Recommended Action**
What should be built or changed, with enough specificity that a PM could
confirm scope and an engineer could start. (1-2 paragraphs max)
Scale detail to scope: bug fix → repro + correct behavior;
enhancement → before vs. after; new feature → user journey.

RICE Scoring

Score = (Reach x Impact x Confidence%) / Effort — higher = better ROI.

Reach guidelines:

• Estimate the number of users or events affected per quarter.
• Use analytics data to ground this: DAU/WAU/MAU counts, funnel volumes, segment sizes from existing cohorts.
• State the source: "~12,000 users/quarter hit this flow based on [chart]."

Impact anchors (expected effect per user):

• 0.25 (Minimal): Cosmetic polish, barely noticeable change
• 0.5 (Low): Minor friction reduction, small quality-of-life improvement
• 1 (Medium): Measurable lift on a key metric
• 2 (High): Significant improvement on a core metric (+15% conversion, meaningful retention gain)
• 3 (Massive): Removes a blocking failure, unlocks a workflow entirely

Confidence anchors:

• 100%: Strong multi-source evidence — quantified funnel data, A/B results, corroborating feedback
• 80%: Analytics + feedback + replays all pointing the same direction
• 50%: Analytics OR validated feedback, not both — reasonable hypothesis
• 20%: Anecdotal signal only — gut feel backed by a few data points

Effort guidelines:

• Estimate in person-months, accounting for coding agents handling implementation. Agents compress pure coding time but don't eliminate review, testing, rollout, or cross-team coordination. Discount the coding portion, keep the rest.
• 0.25: Hours — copy change, config tweak, single-file fix. Agent ships autonomously, human spot-checks.
• 0.5: A day or two — isolated component, 1-3 files, one layer. Agent drafts the PR, human reviews once.
• 1: A sprint — multi-file, single layer, moderate test surface. Agent does the heavy lifting but needs a human review cycle and QA pass.
• 2: A few sprints — FE + BE, integration tests, feature flag. Agent accelerates each piece but a human sequences the work, reviews contracts, and manages rollout.
• 5: A quarter — cross-service, schema changes, migration. Agent helps with boilerplate and migration scripts but a human architects, coordinates across teams, and manages staged rollout.
• 10+: Multi-quarter — major architecture, platform work. Agent contribution is incremental; most effort is design, coordination, and risk management.

Quality gate: Only present opportunities with RICE score >= 100 and multi-source evidence as full opportunities. Weaker signals go in the "Emerging Signals" section.

Phase 4: Validate and Filter

Before presenting, be the skeptic:

1. Partial-data artifacts. If the current day/week is incomplete, compare pace not totals. Never flag an incomplete period as a drop.
2. Seasonality and day-of-week effects. Compare like-for-like (Monday to Monday, not Monday to Sunday).
3. Already-shipped fixes. Cross-reference deployment data. If a fix shipped and the metric recovered, the opportunity is resolved — note it as a win, not an open issue.
4. Correlation vs. causation. Two metrics moving together doesn't mean one caused the other. State hypotheses, not conclusions, unless you have experimental evidence.
5. Apply the "so what" filter. Every opportunity must lead to a concrete action. If you can't articulate what to build or change, it's an observation, not an opportunity.

Phase 5: Deliver the Report

Structure the final output as:

1. Executive summary (3-5 sentences): The highest-signal finding, how many opportunities surfaced, and the single most impactful one. Written as a narrative someone could paste into Slack.

2. Top opportunities (3-7, ranked by RICE score): Each using the opportunity structure from Phase 3. Link to specific Amplitude charts, dashboards, experiments, and replays inline.

3. Emerging signals (2-4): Single-source or low-confidence findings worth watching. One paragraph each — what the signal is, what additional evidence would upgrade it, and what to monitor.

4. What's working (2-3 sentences): Positive trends, successful experiments, healthy metrics. Note if any suggest follow-on opportunities worth exploring.

5. Recommended next steps (3-5 numbered items): Concrete, copy-paste-ready actions ordered by priority. Start each with a verb. Bias toward building charts, running experiments, creating cohorts, or investigating segments — not "share with the team."

6. Follow-on prompt: End with a question about what to dig into next.

Writing standards:

• Narrative over structure. Write like a product memo, not a database record.
• Numbers are evidence, not the story. Lead with the insight.
• Approximate: "~42%" not "42.37%".
• Active voice only.
• Always state the time anchor. "Over the past 30 days" not "recently."
• Link every referenced chart, dashboard, or experiment inline using markdown.
• Total report length: 800-1200 words for the main opportunities. Be concise.

Examples

Example 1: Broad Product Opportunity Scan

User says: "Find me the biggest product opportunities right now"

Actions:

1. Get context and discover the org's most important dashboards, charts, and experiments
2. Query all discovered charts at daily granularity over 30 days, rank by trend magnitude
3. Analyze funnels for conversion drop-offs
4. Pull feedback themes and correlate with metric anomalies
5. Check experiments for unshipped wins and stalled tests
6. Synthesize into ranked opportunities with RICE scores

Example 2: Focused Area Investigation

User says: "Where are we losing users in onboarding?"

Actions:

1. Search for onboarding-related charts, dashboards, and cohorts
2. Query the onboarding funnel and break down by segment (platform, plan, new vs. returning)
3. Pull feedback filtered to onboarding-related complaints and pain points
4. Find session replays of users who dropped off at the worst step
5. Check if any experiments or deployments affected onboarding recently
6. Present opportunities ranked by where the most users are lost

Example 3: Post-Launch Opportunity Discovery

User says: "We launched feature X last week — what opportunities do you see?"

Actions:

1. Search for charts and dashboards tracking feature X
2. Query adoption metrics (daily active users, activation rate, retention)
3. Compare pre-launch vs. post-launch baselines
4. Pull feedback mentioning the new feature
5. Check session replays for the new feature flow
6. Present opportunities: what's underperforming expectations, what friction exists, what to iterate on

Troubleshooting

No dashboards or charts found

Fall back to search_amp_entities with broad queries related to the user's product area. Use query_amplitude_data to build ad-hoc charts from raw events. Suggest the user create a key metrics dashboard.

Feedback API returns errors

Always call use_amplitude_ai_feedback with facet: "sources" before facet: "insights". If no sources are configured, skip feedback and note it as a gap in the report — recommend the user connect a feedback source.

Everything looks healthy — no anomalies

Stability is a finding. Focus on: stalled experiments that need decisions, features with flat adoption that could grow, feedback themes that haven't been addressed, and conversion rates that are "fine" but benchmarkably low.

Too many findings

Cap at 7 full opportunities. Rank by RICE score and demote everything below the cutoff to "Emerging Signals." Merge findings that share a root cause.