review-themes

---
name: review-themes
description: Turns every review you have into a short ranked list of things to fix, with the count behind each one and the owner and date attached. It codes every review rather than skimming, separates the themes that are operational from the ones that belong to your environmental health officer, your HR adviser or your listing settings, and refuses to turn a theme into a plan for managing reviews. Use once a quarter, and before you spend money on the thing everyone assumes is the problem.
---

# What three hundred reviews are actually telling you, sorted into things you can fix

You give this every review you have from every platform for a stated period, with ratings and dates. You get back a coded set, a table of themes with the number of reviews and the number of months each appears in, three themes separated out because they are legal or safety matters rather than service problems, the themes that are really a wrong opening time on a listing, and a ranked fix list with a named owner and a date against each line. It will not produce a plan for making the reviews look better.

## What it does

1. **Define the corpus first, in writing, and take all of it.** Every review from every platform between two dates, exported with the rating, the date, the platform and the full text. Not the ones that stung, not the last twenty, not a sample. State the platforms included, the date range, the total count and the count per platform at the top of the output, and state what is missing and why. This matters beyond good practice: the summary you are about to produce is what the Act calls consumer review information, defined at paragraph 13(5)(d) of Schedule 20 to the Digital Markets, Competition and Consumers Act 2024 as "information that is derived from, or is influenced by, consumer reviews". Publishing that in a misleading way is a banned practice, so a summary built from a flattering selection is a regulated document built badly.

2. **Code every review before naming a single theme.** Read the whole set once without writing anything. Then go through again and tag each review with short descriptive codes for what it actually says: "waited 25 minutes for drinks", "asked three times for the bill", "loved the roast", "could not find the entrance". Be generous with codes, because merging them later is easy and splitting them is not. Only when every review carries codes do you group codes into themes. Doing it the other way round, deciding the themes first and counting reviews against them, produces the themes you already believed in, which is the thing this exercise exists to defeat.

3. **Count each theme two ways: how many reviews, and how many distinct months.** Forty reviews across eleven months is a system that is quietly broken. Nine reviews in one week is an incident with a date, probably a shift, a supplier delivery or a person. The two need entirely different responses, and a raw count cannot tell them apart. Record both numbers, plus the earliest and latest date the theme appears, plus its rating spread, because a theme that appears in five star reviews as well as one star reviews is a quirk being tolerated rather than a fault.

4. **Lift out the themes that are safety matters, because they are not service feedback.** Any theme touching allergens goes to the allergen process, not the improvement list. The Food Standards Agency requires that for non-prepacked food, which is what a restaurant serves, "you must supply allergen information for every item that contains any of the 14 allergens", either as "full written allergen information on a menu, chalkboard or in an information pack" or "verbally, with a written notice placed in a clearly visible position explaining how your customers can obtain this information". Its stated preference is that "written allergen information, supported by a conversation, works best for consumers". Themes about illness, injury or discrimination come out in the same move, each with the review references attached, and go to whoever handles incidents, with the environmental health officer or the venue's adviser involved where the reviews allege harm.

5. **Lift out the themes that are employment matters, and handle the tips one properly.** A theme built from reviews naming a member of staff is an employment matter with its own process and its own confidentiality, and it does not belong in a document that circulates. A theme about service charge or tipping has a statutory home: the Code of Practice on Fair and Transparent Distribution of Tips states that "This code and specified provisions of the legislation come into effect on Tuesday 1 October 2024", and that "the employer is required to have a written tipping policy when qualifying tips are paid at or are otherwise attributable to a place of business on more than an occasional and exceptional basis". The Code also settles a question owners ask: "An employer may share their written tipping policy with customers or display it publicly if they wish, but this is not a requirement of the Tipping Act." So a recurring complaint about the service charge is answered by the policy and by how it is explained at the table, not by a line in a marketing plan.

6. **Separate the themes that are listing settings rather than operations.** A surprising share of the reviews that damage a venue are about information, not service. Tripadvisor will remove several of these itself: "Comments solely on business hours or closures" unless "the official information posted by the business created an expectation of service", "Unsuccessful attempts to contact a business", noting that "We do not consider unanswered calls or emails to be a valid description of the service provided by the business", and "Comments solely on a property's policy". If your second largest theme is people arriving to a closed door, the fix is the opening hours on every listing you have, the bank holiday exceptions, and who answers the phone. That is an afternoon's work and it stops the theme.

7. **Rank what is left by frequency against cost to fix, and put a name and a date on each line.** Three groups. Fix this month, with the person who owns it. Fix this quarter, with what it costs and who signs it off. Watch, with the number it would have to reach to move up. Every line carries the theme, the review count, the month count, one representative quote, the owner and the date. A fix list with no names on it is a list of observations, and it will read identically next quarter.

8. **Quote anonymously, and never build a file that identifies customers.** A theme is illustrated by one short quote from a review, with the review reference and the date, and nothing that identifies who wrote it: no display name, no booking, no table, no what they ordered. This is partly a privacy duty and partly a practical one. The document will be read by a supplier, a new manager and possibly a buyer, and a spreadsheet naming customers beside complaints is a liability that outlives the problem it was made to solve.

9. **If you publish any of this, publish it honestly, and date the whole thing.** The moment a summary of your reviews appears on your own website it becomes consumer review information in a marketing communication, and CAP Code rule 3.46 applies: marketers must not publish consumer reviews or consumer review information in a misleading way, which includes "giving greater prominence to positive consumer reviews over negative ones" and omitting information relevant to the circumstances in which reviews were written. An average built from one platform and presented as your reviews is exactly that. Put the build date, the corpus definition and the method on the document, and re-run it next quarter against the same rule so the comparison means something.

## Then it checks

1. The corpus is defined in writing with platforms, date range, total count and count per platform, every review in that range is coded, and any exclusion is listed with its reason.
2. Every theme carries a review count, a distinct month count, an earliest and latest date, a rating spread and the review references that make it up, and no theme is asserted without those numbers.
3. No theme was named before coding was complete, and the codes underlying each theme are listed so that somebody else could rebuild the theme from them.
4. Every theme touching an allergen, illness, injury, discrimination or a named member of staff has been removed from the improvement list and routed, with its references, to the process that owns it.
5. Every remaining theme carries a named owner and a date, and every quote used is short, anonymous and carries its review reference.
6. Nothing in the output proposes changing, removing, answering selectively, suppressing or soliciting reviews, and any summary intended for publication carries its corpus definition, its method and its build date.

Any check fails: name it, redo that step once. Failed twice: say what is wrong and stop.

## Rules
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is. A fix list circulated as the owner's decisions goes to the owner before it circulates.
- It will refuse to turn a theme into a plan for managing reviews. Asking fewer unhappy customers, replying selectively, burying a theme under new reviews or timing invitations around bad weeks are all interference with negative reviews, and the Competition and Markets Authority names each of them. The output of this skill is a list of things to change in the building.
- Never name a member of staff anywhere in the output, including in a quote. Redact the name and keep the review reference, so the person who owns the employment process can find it without the name travelling.
- Never treat a theme as resolved because the reviews stopped. Reviews stop for many reasons, including people giving up. A theme is closed when the thing that caused it changed, with a date and a name against the change.
- Never rank a theme by how strongly it was written. The angriest review and the most common problem are rarely the same thing, and ranking by feeling is how a venue spends six thousand pounds on the wrong chairs.
- Never state a benchmark, an industry average rating, a sector norm for complaint frequency, or what a good score is. No such published figure exists for a single independent UK venue, and a borrowed number turns a real finding into a comparison with a fiction.
- Never carry a theme forward from a previous run without recounting it against the current corpus. A theme that was true in March can be fixed, or can have doubled, and a copied table hides both.
- This output is a working document prepared for the owner to check against their own records, and for their environmental health officer, HR adviser, solicitor or accountant to check where a theme touches food safety, employment, tips, discrimination or a legal duty. It counts and groups what customers wrote; it is not an audit, it is not advice, and it does not establish that anything alleged in a review occurred.

## Built from
- Digital Markets, Competition and Consumers Act 2024, Schedule 20, paragraph 13, https://www.legislation.gov.uk/ukpga/2024/13/schedule/20/paragraph/13/enacted, Act of Parliament given Royal Assent in 2024, read 16 September 2026: the definition of consumer review information at 13(5)(d) and the misleading publication examples at 13(5)(i), in steps 1 and 9.
- Competition and Markets Authority, "Fake reviews" (CMA208), https://assets.publishing.service.gov.uk/media/67eeb64fe9c76fa33048c790/CMA208_-_Fake_reviews_guidance.pdf, published 4 April 2025, read 16 September 2026: paragraph 2.11 on what counts as consumer review information, and paragraphs 4.3 to 4.5 on suppressing, cherry picking and interfering, which set the refusal in the rules.
- Food Standards Agency, "Allergen guidance for food businesses", https://www.gov.uk/government/publications/allergen-guidance-for-food-businesses/allergen-guidance-for-food-businesses, updated 17 July 2026, read 16 September 2026: the non-prepacked food duty, the fourteen allergens and the ways information may be provided, which decide what step 4 removes from the improvement list.
- Department for Business and Trade, "Code of Practice on Fair and Transparent Distribution of Tips", https://www.gov.uk/government/publications/distributing-tips-fairly-statutory-code-of-practice/code-of-practice-on-fair-and-transparent-distribution-of-tips-html-version, published 29 July 2024, read 16 September 2026: the 1 October 2024 commencement, the written tipping policy requirement at paragraph 34 and the position on showing it to customers at paragraph 40, in step 5.
- Tripadvisor, "Trust and safety review posting guidelines", https://www.tripadvisor.co.uk/Trust-lvBd3L1aU38Y.html, no publication date shown on the page, read 16 September 2026: the ineligible experience categories quoted in step 6, which is how a listing settings theme is told apart from an operational one.
- Committee of Advertising Practice, "Fake consumer reviews", AdviceOnline, https://www.asa.org.uk/advice-online/fake-consumer-reviews.html, 11 December 2025, read 16 September 2026: rule 3.46 and its examples of publishing in a misleading way, in step 9, together with the confirmation that the unfair commercial practice on reviews "came into force on 6 April 2025".

# review-themes

## You are given
Every review the venue holds for one stated period, exported from every platform it appears on, one row per review, carrying the review identifier, the platform, the date, the rating, the full review text and the language it is written in. The corpus definition already written by Claude: the platforms included, the start date, the end date, the expected total, and any platform or date range deliberately excluded with the reason. The codes applied to each review, one review to many codes, each code a short phrase in the reviewer's own terms. The themes already formed by Claude, each with the list of codes that make it up and a one sentence statement of what the theme is. The routing decision for each theme: `operational`, `safety or legal`, `employment`, `tips`, or `listing settings`. For every theme routed away from the improvement list, the process it goes to and the person who owns that process. The owner's fix list with a named owner and a date against each operational theme. The previous run's theme table, where one exists.

## Produce
Write into a `./review-themes-output/` folder:

1. `corpus.csv` with these columns in this order: `platform`, `date_from`, `date_to`, `reviews_supplied`, `reviews_coded`, `reviews_excluded`, `exclusion_reason`, `earliest_review_date`, `latest_review_date`, `rating_1`, `rating_2`, `rating_3`, `rating_4`, `rating_5`. One row per platform plus a `TOTAL` row. Any platform where `reviews_coded` is less than `reviews_supplied` is listed in `gaps.md`.
2. `coded-reviews.csv` with columns: `review_ref`, `platform`, `review_date`, `rating`, `review_text_verbatim`, `language`, `codes_applied`, `code_count`, `themes_assigned`, `contains_named_individual`, `contains_allergen_word`, `contains_illness_or_injury_word`, `contains_discrimination_word`. `review_ref` is T0001 upward. `codes_applied` and `themes_assigned` are semicolon separated. The four `contains_` columns are `yes` or `no`, and every `yes` records the exact words found in `trigger-words.csv`.
3. `trigger-words.csv` with columns: `review_ref`, `category`, `word_or_phrase_found_verbatim`, `position`, `theme_assigned`, `routed_to`. `category` is exactly one of `allergen`, `illness`, `injury`, `discrimination`, `named individual`. The allergen search covers the fourteen named by the Food Standards Agency: celery, cereals containing gluten, crustaceans, eggs, fish, lupin, milk, molluscs, mustard, peanuts, sesame, soybeans, sulphur dioxide and sulphites, tree nuts, together with their common names.
4. `themes.csv` with columns: `theme_ref`, `theme_statement_verbatim`, `codes_in_theme`, `review_count`, `share_of_corpus`, `distinct_months`, `earliest_date`, `latest_date`, `rating_1_count`, `rating_2_count`, `rating_3_count`, `rating_4_count`, `rating_5_count`, `platforms_present`, `routing`, `pattern`. `theme_ref` is TH01 upward. `routing` is exactly one of `operational`, `safety or legal`, `employment`, `tips`, `listing settings`. `pattern` is `spread` where the theme appears in three or more distinct months, `concentrated` where it appears in one, and `emerging` where the earliest date falls in the final third of the period.
5. `theme-quotes.csv` with columns: `theme_ref`, `review_ref`, `quote_verbatim`, `quote_word_count`, `redactions_made`, `contains_identifier`. One quote per theme, at most forty words. `redactions_made` lists any personal name replaced. `contains_identifier` is `yes` or `no` and must be `no`; any `yes` is listed in `gaps.md`.
6. `fix-list.csv` with columns: `theme_ref`, `group`, `action_verbatim`, `owner`, `due_date`, `estimated_cost_gbp`, `review_count`, `distinct_months`, `closed_date`, `what_changed`. `group` is exactly one of `fix this month`, `fix this quarter`, `watch`. Only themes whose `routing` is `operational` or `listing settings` appear here. `closed_date` is set only where `what_changed` is non-empty.
7. `comparison.csv` with columns: `theme_ref`, `theme_statement_verbatim`, `review_count_this_run`, `review_count_previous_run`, `share_this_run`, `share_previous_run`, `direction`, `recounted_against_current_corpus`. `direction` is exactly one of `up`, `down`, `unchanged`, `new`, `gone`. `recounted_against_current_corpus` must be `yes` for every row. Written empty where there is no previous run.
8. `gaps.md` - a numbered list of: any platform with uncoded reviews; any review with no codes; any theme with fewer review references than its stated `review_count`; any theme routed as safety, employment or tips that also appears in `fix-list.csv`; any quote containing an identifier; any operational theme with no owner or no due date; any theme carried from a previous run without being recounted; any review in a language nobody on the team reads; and any allergen, illness, injury or discrimination trigger whose theme has no recorded process owner.

## Rules
- Codex counts, groups, matches and compares. It never writes a code, never names a theme, never decides a routing, never writes an action and never judges what matters. All of those arrive with the input.
- Never summarise a review. Quotes are copied exactly and truncated only at a word boundary, with the truncation marked.
- Never write a person's name into any file except `trigger-words.csv`, where it exists to be found and removed, and `redactions_made` in `theme-quotes.csv`, where it records what was taken out. Staff names never reach `themes.csv`, `fix-list.csv` or any quote.
- Never place a theme routed as `safety or legal`, `employment` or `tips` into `fix-list.csv`, whatever its review count. Record it in `themes.csv` with its routing and its process owner and stop there.
- Never write, suggest or record any action that changes, removes, answers, filters or solicits reviews. This output is a list of changes to the business. An action about the reviews themselves is recorded verbatim in `gaps.md` as out of scope and is not carried into `fix-list.csv`.
- Never write an industry average, a sector benchmark, a good score, a target rating or a comparison with any business other than this one. No published figure exists for a single independent UK venue.
- Never carry a count forward from a previous run. Every comparison figure for this run is recounted from `coded-reviews.csv`, and `recounted_against_current_corpus` records that it was.
- Counts, shares, month counts and date ranges are computed, never estimated. `distinct_months` counts calendar months containing at least one review in the theme. `share_of_corpus` is the theme's review count divided by the coded total, given to one decimal place.
- Every quoted review, theme statement and action is copied exactly, including punctuation, capitalisation and any emoji.
- Never connect to a review platform, never fetch reviews, and never publish, post or send anything.
- Use British English, GBP and DD Month YYYY dates. No em dashes in any file you write, and any supplied text containing one is recorded verbatim and flagged in `gaps.md`.
- Every file ends with this line: this is a working document prepared for the owner to check against their own records, and for their environmental health officer, HR adviser, solicitor or accountant to check any theme touching food safety, employment, tips, discrimination or a legal duty. It counts and groups what customers wrote and does not establish that anything alleged occurred.

## Return
The absolute path of each file written and the row count of each CSV. The corpus definition as supplied, with the total supplied, total coded and total excluded, and the per platform breakdown. The number of distinct codes and the number of themes. The full `themes.csv` ordered by review count, showing for each theme its review count, its share of the corpus, its distinct month count, its date range, its rating spread, its routing and its pattern. Every allergen, illness, injury, discrimination and named individual trigger found, with the review reference, the exact words and where the theme was routed. The three groups of `fix-list.csv` with their theme references, owners and due dates, and any operational theme missing an owner or a date. Every quote flagged as containing an identifier. The comparison table with the direction of each theme against the previous run. And the `gaps.md` item count.