# Review Agent Insights > > - **Type:** Skill - **Install:** `agentstack add skill-amplitude-mcp-marketplace-review-agent-insights` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [amplitude](https://agentstack.voostack.com/s/amplitude) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [amplitude](https://github.com/amplitude) - **Source:** https://github.com/amplitude/mcp-marketplace/tree/main/plugins/amplitude/skills/review-agent-insights ## Install ```sh agentstack add skill-amplitude-mcp-marketplace-review-agent-insights ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # Review Agent Insights Surface everything Amplitude's AI agents have found recently. Query **every available agent type** in `get_agent_results`, validate for staleness, and synthesize into a unified narrative ranked by impact with concrete follow-up actions. --- ## CRITICAL: Tool Reference **Primary tool:** - **`Amplitude:get_agent_results`** — Retrieve pre-computed analyses from Amplitude's AI agents. Supports multiple agent types (check the tool's `agent_type` enum for the current list). Each agent type is queried separately. All support filtering by `created_after`, `created_before`, `query`, `agent_params`, and `limit`. **Supporting tools:** - **`Amplitude:get_context`** / **`Amplitude:get_project_context`** — Bootstrap user, org, and project info. - **`Amplitude:get_deployments`** — Check whether fixes have shipped for flagged issues (staleness validation). --- ## Instructions ### Step 1: Bootstrap Context (1-2 calls) 1. Call `Amplitude:get_context` to get the user's org, projects, recent activity, and key dashboards. If multiple projects, ask which to review — or review all if the user wants a broad scan. 2. Call `Amplitude:get_project_context` for the target project's settings and AI context. Determine the **review window** from the user's request: - Default: **last 7 days** (good balance of recency and coverage). - "What's new today?" → last 1-2 days. - "Catch me up on this month" → last 14-30 days. - Always compute the `created_after` ISO 8601 timestamp for the review window. ### Step 2: Query All Agent Types (parallel) Check the `get_agent_results` tool descriptor to discover every available `agent_type` in the enum. **Make one call per agent type, in parallel.** For each: - `agent_type`: the agent type from the enum - `created_after`: the review window timestamp from Step 1 - `limit: 10` If the user asked about a specific area (e.g., "onboarding insights"), add a `query` matching that area to every call. If an agent type supports additional filtering via `agent_params` (e.g., impact ratings, categories, dashboard IDs), use them to focus results when the user's request suggests a narrower scope — otherwise omit `agent_params` to get the broadest view. For each result returned, note: - Which agent type produced it - The key findings or summary - When the analysis was run (creation date) - Any metadata specific to that agent type (impact ratings, categories, dashboard IDs, etc.) If exactly 1 result is returned for an agent type, artifacts auto-expand. If multiple results, note the previews and fetch full artifacts for the 2-3 most relevant ones (by recency or by matching the user's focus area) using `session_id`. If any agent type returns fewer than 3 results and supports `agent_params` filtering, consider a second call with relaxed filters to broaden coverage. ### Step 3: Validate Freshness Agent insights go stale within days. Before synthesizing, filter out or flag anything unreliable. 1. **Check creation dates.** For each finding, note how old it is relative to today: - ** 14 days old**: Stale — exclude from the main narrative. Mention in passing only if nothing newer exists for that area. 2. **Cross-reference with deployments (1 call).** Call `get_deployments` once. For each AI-detected issue, check if a deployment shipped a fix or change to the affected area after the analysis was run. If so, note the finding as "potentially resolved by [deployment]" rather than presenting it as an active issue. 3. **Deduplicate across agent types.** The same problem may surface from multiple agent types. Merge these into a single finding with multi-agent evidence — don't present the same issue multiple times. ### Step 4: Synthesize and Rank 1. **Rank by impact and evidence strength.** - Multi-agent findings (flagged by more than one agent type) rank highest. - High-impact or high-confidence findings from a single agent type rank next. - Low-impact or older findings rank lowest. 2. **Group by theme, not by agent type.** Organize findings by product theme or problem area ("Checkout flow," "Onboarding," "Search feature"), not by which agent produced them. Within each theme, weave together evidence from all contributing agent types. 3. **Identify gaps.** Note agent types that returned no recent results, or product areas with no coverage. ### Step 5: Present the Review Structure the output as a narrative digest that a PM could forward to their team. **Required sections:** 1. **Summary** (3-4 sentences): Which agent types were queried, the review window, how many results total, the single most important finding, and overall assessment. 2. **Key Findings** (3-7 items, ranked by impact): For each finding: ``` ### [Finding Title — action-oriented, ≤10 words] **Impact:** [Critical/High/Medium/Low] | **Agents:** [list agent types that contributed] | **Freshness:** [X days old] **What the AI found:** Describe the insight — what anomaly, friction, or issue was detected. Be specific about the product area and the evidence from each agent. **Staleness check:** Note if deployments shipped after the analysis, or if the finding needs fresh validation. Omit this line if the finding is 14 days old).** Present a brief summary with clear staleness warnings and recommend re-running analyses. - **Overwhelming number of findings.** Cap at 7 key findings. Rank by impact × freshness × evidence breadth. Mention the total count. - **User asks about a specific area.** Add a `query` parameter to every agent type call. Present only relevant findings. - **Unrecognized agent type results.** If a new agent type returns results in a format you haven't seen before, present the raw findings with the agent type name and any available metadata. Don't skip results just because the agent type is unfamiliar. ## Examples ### Example 1: Broad Review User says: "What has the AI found recently?" Actions: 1. Get context — identify key project and dashboards 2. Check `get_agent_results` for all available agent types, query each in parallel with `created_after` set to 7 days ago 3. Validate freshness — cross-reference against deployments, filter out stale findings 4. Synthesize and group by product theme, noting which agent types contributed to each finding 5. Present unified findings ranked by impact, note coverage gaps for agent types with no results ### Example 2: Focused Area Review User says: "Any AI insights about onboarding?" Actions: 1. Get context 2. Query all agent types with `query: "onboarding"`, `created_after` set to 7 days ago 3. Filter to only onboarding-related findings 4. Present findings + gap recommendations for agent types that returned nothing for onboarding ### Example 3: Nothing Found User says: "Show me all AI agent insights" Actions: 1. Get context 2. All agent type queries return empty within the review window 3. Present: "No AI agent results found in the last 7 days. Here's how to generate them:" — list each agent type that was queried, what it does, and suggest specific content to analyze ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [amplitude](https://github.com/amplitude) - **Source:** [amplitude/mcp-marketplace](https://github.com/amplitude/mcp-marketplace) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-amplitude-mcp-marketplace-review-agent-insights - Seller: https://agentstack.voostack.com/s/amplitude - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.