Who is this assistant for?

Every bad review ruins your day and a month later you remember none of them. You are about to fix the wait because it is what people write most. You know service is failing and not on which shift or at which location. You changed three things two months ago and cannot tell whether any of them worked.

An owner remembers the last bad review; the pattern costing them money sits in the thirty before it and nobody adds them up. This assistant reads the lot as data: it sorts every review by the problem it names rather than by its star, applies a threshold so two appearances out of forty never pass as a pattern, and weighs them, because the wait shows up often and stings little while a cold plate shows up rarely and nobody comes back. Each pattern lands on a station with its shift and location, and leaves with an owner, a date and a number to be read six weeks later in new reviews about the same problem. A count with no plan is a curiosity and a plan with no count is a hunch; here they travel together or not at all.

AI assistant · N4 Executable plan

◆ Included in the annual plan⭐ Reviews, Reputation and Automated Messagingv1.1.1 · 2026-08-26

Recurring Problem Detector for Restaurant Reviews

Turns a pile of reviews into a short list of problems ranked by weight -frequency times harm-, each tied to a station, a shift and a location, with a three-action corrective plan and the six-week test that says whether it worked.

Depth

12/23

Your data

Compatible AIs

Steps to use it

Depth of the resulttap a level to see what you get

◀ Put out today's fireRun the year ▶

🗺️

N4Executable plan

A plan with owners, deadlines and the metric you review. Not advice: an agenda someone can execute on Monday.

◆ This assistant 23 of the 164 published assistants are at this level

Is this for you?

What it does and what you get

An owner remembers the last bad review; the pattern costing them money sits in the thirty before it and nobody adds them up.
This assistant reads the lot as data: it sorts every review by the problem it names rather than by its star, applies a threshold so two appearances out of forty never pass as a pattern, and weighs them, because the wait shows up often and stings little while a cold plate shows up rarely and nobody comes back.
Each pattern lands on a station with its shift and location, and leaves with an owner, a date and a number to be read six weeks later in new reviews about the same problem.
A count with no plan is a curiosity and a plan with no count is a hunch; here they travel together or not at all.

What you end up with

Five chained deliverables: the cleaned lot with what was discarded and why, the count table with appearances, harm and weight, the three patterns that enter with their finding, the corrective plan with owner and date, and the six-week test with the deciding number.

EXAMPLE This is what you getAn example built with Sazón de Origen, the house's fictional restaurant.Not your data: your result is built with YOUR restaurant's.See the full example →

Sample result

Sazón de OrigenFictional contemporary Colombian restaurant in Medellín, 2 locations, 6 years in business. Every sample result in the library comes from this same case. Fictional restaurant

The owner asked this: «I have 58 reviews from the last three months across Google and the delivery app. I know something is off with weekend service, but every time I read them I am left with the last one that stung.»

Block A — the lot: 58 read · a 90-day period · Google 39 and delivery app 19 · 14 with no problem named (they join the mute pile) · 4 discarded with their reason: two wanted a dish the house does not sell, one had no visit, and one from the Sunday when lunch broke on a local public holiday in the market's calendar.

Block B — the count, sorted by weight and not by appearances:

Problem	Appearances	% of those naming	Harm	Weight	Station	Slot and location
Dish arrives cold	5	12.5%	3	15	Packaging	Evening · delivery · Centro
Wait at the table with no warning	9	22.5%	1	9	Floor	Saturday evening · Centro
Portion smaller than expected	4	10%	2	8	Kitchen	All day · both locations
Unexplained service charge	2	5%	2	4	Price	— · Norte

The bottom one does not enter: the service charge stays on the watch list with its figure —2 appearances out of 40 do not reach the threshold— and gets re-read at the second count. Not the bin: the list.

Illustrative example generated with a fictional restaurant. Not a promise of results, not a client case.

This is 2 of 3 parts. Behind the wall are **the corrective plan with owner and date** —what changes, who does it, what it costs per month and what the portion-weight adjustment does to food cost— and **the 6-week test**: the deciding number and what happens if it did not drop.

See the full example — free account

Works with these AIs

Claude — the best for this assistant The best fit here: it holds the threshold on a long lot and does not promote a two-appearance problem to a pattern just because the review is well written.

How to use it

Gather the last 90 days of reviews from every channel: export what the platform lets you and copy the rest by hand, with the date and the channel on each one.

Copy the assistant into your preferred AI, paste or attach the full lot and answer what it asks about slots, locations and dishes.

Check the count before looking at the plan: if a problem sounds odd, ask for the reviews holding it up and read them.

Put a name and a date on every correction and book the second count six weeks out on the same day the first one starts.

Which of your data it uses

of 23

This assistant works with 12 of the 23 fields in «My Restaurant», among them Restaurant name, Business type and City and country. It does not ask out of curiosity: these are what make the answer speak about YOUR scale instead of an industry average — the same calculation on a business of another size returns a number you cannot decide with. You fill them once and they apply across the library, so the second assistant you open already starts with them in place. Whatever you leave blank, the assistant asks for one thing at a time instead of inventing it, and says so before answering, so you know what it is working from.

Restaurant nameBusiness typeCity and countryNumber of locationsWhat makes you differentActive sales channelsStrong and slow daypartsBest-selling dishesMenu price rangeTeam sizeAverage check (with currency)Main goal this year

Always up to date

Every assistant carries a visible version and date. When an AI changes how it works, the assistant's version goes up and you see what changed.

v1.1.12026-08-26The body stops contradicting the answer ladder. It used to say «do not move on until it arrives», the very order the ladder was born to replace: the assistant asked the owner for what the owner does not have and stalled. It now says what the house doctrine says — ask one at a time, and if they do not know, continue with a marked assumption. FORMAT OF THE RESULT was not touched, so the example still corresponds and its seal moves up with the card.

v1.1.02026-08-25Body rewritten to pay down duel debt: it scored 76.5 out of 90.0, the best in the batch and still under the threshold. The `lexico` signal read ZERO: the card sorted and weighed complaints without naming the restaurant trade even once, which made it a pattern counter that would serve a hair salon just as well. The vocabulary now enters where it actually decides: the wait is crossed against occupancy and the order time before blaming the kitchen —much of the time it is table turnover—, the count is set on the shift's covers as denominator, the sales channel splits dining room from delivery because they do not fail for the same reason, a taste or size complaint is checked against the standard recipe and the portion weight at the pass, a pattern landing on a star dish moves up in priority, and a fix that raises portion weight is counted against food cost and contribution margin before it is signed. `cuidados` read 3 of 6 and `verificacion` 4 of 5: the card's own «TRAPS OF THE TRADE» section fell outside the seven-label skeleton that arraigo and the renderer read, so its warnings are now spread across ROLE, METHOD and BEFORE YOU DELIVER, where the reader needs them. `cifras` read 3 of 12 despite this being a counting card: the threshold, the window and the deadlines are now measured —8%, a lot of 40, a 90-day window, a second count at 6 weeks, a watch-list review at 30 days, the pattern that lives in the ones from 2 months ago—. The SUPUESTO: line is added (recipe §0.1) on the harm scale, which is the assumption that genuinely reorders the plan. Level N4 is preserved: the FORMAT stays in lettered, unnumbered blocks. Lead-magnet kit regenerated against the new body.

v1.0.02026-08-25The assistant is born.

Same method, another variant

Inside the app this is a single control: you press it and the assistant reframes itself, including the variants that don't appear here because they don't have their own page yet.

Which moment

ask for reviewsread what they already saywin back whoever left upset