Analyze reviews

Analyze reviews

Turn a batch of reviews into a clear, honest analysis. You pick what kind: an overview, what customers love, their complaints, their requests, or a data-quality check. Reviews come from a Feedspace account through the Feedspace MCP, a CSV, or pasted text.

The full method lives in references/analysis-method.md. Follow it exactly. The analysis engine runs entirely in this session and needs no API key; pulling reviews from the Feedspace MCP is a separate, read-only connection.

Steps

1. Get the reviews.

   • Feedspace MCP: first check whether it is already connected - are Feedspace tools (for example list_workspaces) available in this session? If yes, just use them; do not ask the user to connect or authenticate again, and do not paste setup steps. Size the workspace, then fetch the reviews. Only if the tools are not available, point the user to setup in references/mcp-source.md, which also covers the fetching strategy (pagination, large workspaces, dedupe, empties). Ask which workspace if there is more than one.
   • Otherwise read a CSV (see references/csv-format.md) or ask the user to paste reviews. If a CSV's review-text column is not obvious, confirm it - do not guess.
   • Reduce each review to { review: "<text>" }.

2. Snapshot, then ask. Give a 1 to 3 line snapshot of the data (how many usable reviews, one business or several mixed, rating spread, any quality flags), then ask which analysis the user wants: Overview, What customers love, Problems & complaints, Requests & suggestions, or a Data-quality check. Skip the question if they already told you which one.

3. Run that analysis by following references/analysis-method.md: grounded in real review text, honest about what the data supports, clustered by business if the workspace is mixed.

4. Offer one sensible next step, then stop. Do not produce action items (which reviews to feature or reply to). If the reviews came from the Feedspace MCP, offer to go further with the connection - pull more, filter (by rating, review type, form, label, or import source), or run another lens. If they came from a CSV or paste, just offer another lens.

Guardrails

• Ground everything in real review text. Never invent a problem, request, or praise that is not in the reviews.
• If the data is mixed, a demo, or noisy (duplicates, non-reviews, empty entries), report by cluster and flag the issues - do not force a single "customer voice."
• If you analyzed a sample rather than all reviews, say how many of how many.
• The Feedspace MCP path is read-only. Never modify or delete anything.

This version stops at the analysis. Action items - what to do with the reviews, such as which to share or reply to - are intentionally deferred for now.