Sentiment Analysis (feedback corpus over a period)

Read the mood of the feedback corpus — for a topic, a segment, or a time window — and turn it into scored themes backed by real quotes. Not a vague "customers seem happy," but "onboarding sentiment is –0.4 this quarter, driven by setup friction, and here are the six people who said so."

Step 0 — Relevance & availability

Confirm this is a feedback/satisfaction question spanning multiple conversations and Evermuse is connected (see using-evermuse Step 0). If disconnected, produce the framework labeled ⚠ ungrounded and tell the user to authorize the MCP. For a single conversation's sentiment, use /evermuse:customer-research instead.

Evermuse Grounding (required)

Follow ${CLAUDE_PLUGIN_ROOT}/skills/using-evermuse/SKILL.md.

Required — one parallel batch of searches. Verify the product (Rule 1), then fire in a single parallel batch:

• 3–4 evidence searches, each worded from a different angle (limit up to 50). Evidence comes back rich and varied — expect large, useful result sets.
• one guidance search and one context search (limit up to 50). These are usually sparse or empty; run them anyway and note when they're thin.

Read each response's digest — it reports how many more results exist. Use judgment on whether a query is worth pulling deeper (raise limit toward the 100 max and/or page with next_offset to avoid repeats), weighing payload size, remaining context, task complexity, and the value of the data. For deep pulls, consider spawning sub-agents — instruct them to return every citation with the same metadata the tools return (url, who_said_it, meeting_name, created_at) so you can still cite.

Optional — considered use. Once grounded, reach for the other tools only when they add value: a quote-angled search for verbatim voice, read_source for a single deep dive, inline citations (references/citations.md), and add_source to save the deliverable. Sources, citations, and saving are optional — not required.

Instructions

Analyze sentiment for $ARGUMENTS over the specified period (or all-time if none given — state which).

1. Inventory the feedback pulled — count of notes, date span, distinct accounts. If the window is thin, say so before scoring.
2. Cluster into themes — recurring topics of praise and complaint. Aim for a handful, not one bucket.
3. Score each theme — a sentiment score from –1 to +1, derived from the sentiment_analysis labels on the underlying quotes, not vibes. Note the split (e.g. "mostly negative, 5 of 7 mentions").
4. Separate drivers — for each theme, the satisfaction drivers vs. detractors, each with a real example.
5. Track shift over time — if the window allows, whether sentiment on a theme is rising or falling, and against what (a release, a change).

Output

Lead with a one-line overall read (net sentiment + the single biggest driver), then per theme:

### [Theme] — sentiment [score] ([N mentions / M accounts], mostly [pos/neg])
**Loves:** [driver]
> "[verbatim]" — [Name], [Meeting], [Date] · sentiment: [positive] [`1`](URL)
**Frustrations:** [detractor]
> "[verbatim]" — [Name], [Meeting], [Date] · sentiment: [negative] [`2`](URL)

Then top pain points ranked by frequency × severity and 2-3 highest-impact recommendations. Represent minority/dissenting sentiment — don't flatten a split into a false consensus. Flag any theme resting on a small sample. Offer /evermuse:customer-research to drill into any single theme.

Further reading

• Sentiment-synthesis method adapted from phuryn/pm-skills (MIT); grounding is Evermuse-native.