Analyse a Batch of Reviews for Themes, Sentiment and Actions

You are a customer-feedback analyst for an Australian small business. Analyse the batch of reviews/comments below with discipline: themes drawn from the data, counts of what's actually there, quotes kept exact, and caveats stated before actions.

<context>
BUSINESS: [TYPE — so themes make sense]
REVIEWS: [PASTE THEM, ONE PER LINE OR BLOCK, WITH SOURCE AND DATE IF AVAILABLE]
DECISIONS PENDING: [WHAT YOU'RE WEIGHING UP — e.g. whether to change suppliers, retrain, renovate]
PERIOD COVERED: [IF KNOWN]
</context>

Before classifying, read the whole batch once and build the theme taxonomy FROM the data — do not impose predefined categories. Merge themes with fewer than two mentions into 'other'.

<task>
1. Per item: sentiment (positive / negative / mixed / neutral) and theme tags. Sarcasm or ambiguity gets flagged 'uncertain' rather than forced into a bucket.
2. Summary stats: counts by sentiment, counts by theme — always as 'X of the N items provided', never as claims about customers in general.
3. Theme table: theme | count | sentiment lean | 1-2 EXACT verbatims (unedited, in quotes) | which pending decision it bears on, if any.
4. Separate product themes from service themes from price themes — they route to different owners.
5. Trend note ONLY if dates were provided ('delivery complaints cluster after March'); otherwise state that no trend read is possible.
6. Distortion check: name any theme driven by one prolific or extreme voice, and weight it accordingly.
7. Sample-bias caveat: who leaves reviews for a business like this (the delighted and the furious) and what that means for reading the middle.
8. Three actions, each traced to a theme and its count, each matched to a pending decision where possible — plus one 'what NOT to change', where the data shows something working that intuition might fiddle with.
</task>

<output_format>
Summary stats → theme table → trend/distortion/bias notes → three actions + one protect.
</output_format>

Rules: no percentages on tiny bases without saying the base; no invented quotes or paraphrases presented as quotes; if the batch is under ~10 items, say plainly that this is anecdote-reading, not analysis. En-AU spelling.

Copy the block above straight into Claude — anything in [BRACKETS] is yours to fill in.