AI Bot for Customer Review Responses
Automatically analyzes review sentiment and generates a personalized response. We break down the core principle: one complex prompt is a trap, a chain of simple ones is power. We build a workflow with branching for cases when the AI needs to call a human.
IntermediateAutomation25 minn8n, Claude API, Google Sheets
1
Paper first, keyboard second
"Respond to a customer review" is not one task — it's five. Get the review, understand the tone, choose a strategy, generate a response, save it. Try stuffing all that into one prompt and you get a fragile thing where it's unclear what broke. Draw the pipeline on paper in 2 minutes. If you can't — the task isn't understood yet, and building is premature.
Get review
Analyze sentiment
Choose strategy
Generate response
Save result
Decomposition is not "an extra step before the work." It IS the work. Everything else is implementing the decisions you made on paper.
2
Two prompts beat one — here is why
The analyst reads the review and fills out a card: sentiment, key issues, urgency. The copywriter gets the card and writes a response. Why not one prompt? Because you get a checkpoint. You see what the AI understood before it starts responding. Bad answer? Check the card — instantly clear whether the analyst or the copywriter failed. Plus you can swap the copywriter for a different brand voice without reworking the analysis.
❌ One prompt for everything
- Hard to know where it broke
- No intermediate checkpoint
- Changing tone = rewriting everything
✅ Analyst + Copywriter
- See the analysis card — verifiable
- Each step debugged independently
- Copywriter is easy to swap
отзыв → Аналитик → карточка:
тональность: позитив | негатив | нейтрал
проблемы: [список ключевых жалоб]
срочность: низкая | средняя | высокая
нужен_человек: да/нет
если угроза / мат / возврат → нужен_человек = даIf the intermediate result (the analyst card) can't be verified by eye in 10 seconds — the decomposition is wrong.
3
The brief matters more than the copywriter
Why does the second prompt receive a structured card instead of the raw review? For the same reason a good brief produces better results than "write something cool." The urgency field sets the tone: high urgency means an immediate promise of resolution, low means friendly acknowledgment. The key_points field guarantees no complaint gets lost. You control AI behavior through data structure, not a page of instructions. This is more reliable and easier to change: want a different tone for VIP customers? Add a field to the card instead of rewriting the entire prompt. Data passing between steps isn't a technical detail. It's an architectural decision that determines how flexible the system will be in a month.
Копирайтер получает карточку, НЕ сырой отзыв:
ВХОД: { тональность, проблемы, срочность }
ПРАВИЛА:
срочность высокая → обещание решения
срочность низкая → дружелюбное подтверждение
каждая проблема из списка → адресована в ответе
ВЫХОД: ответ ≤120 словA hard word limit on response length (100-150 words) is non-negotiable. Without it, AI writes essays that customers never read.
4
100% automation is a myth. Add the escape hatch
Legal threats, public scandals, subtle sarcasm — there are reviews the AI should never answer on its own. The requires_human field in the analyst card is not a crutch — it is a sign of a mature system. Good automation knows its limits. In n8n this is an IF node in two clicks: requires_human == true routes to a team notification, everything else goes to generation and saving.
Trigger (Google Sheets)
Analyst (Claude)
requires_human?
yes
Notify team
no
Copywriter (Claude)
Save to Sheets
In the analyst prompt, explicitly list escalation cases: legal threats, abusive language, refunds above X amount. Without concrete examples, the AI will guess.
5
Test with a checklist, not gut feeling
Two successful tests don't prove quality. Prepare 10 reviews: 3 positive, 3 negative, 2 neutral, 1 with a threat, 1 with sarcasm. Check not "do I like the answer" but specific criteria from the checklist. If something is off — fix the prompts, not the architecture. 90% of the time the problem is in instruction wording.
Testing checklist
Sentiment correct for all 10 test reviews
Response addresses each key_point from the card
Escalation triggers on threats and sarcasm
Response stays under 150 words in every test
No edge cases covered — only obvious scenarios
No baseline for % correct escalations
Save all analyst cards alongside responses in Google Sheets. After 2 weeks you will have an audit dataset — error patterns invisible in 10 tests.
Result
An n8n workflow that automatically responds to customer reviews based on sentiment and escalates complex cases to a human agent. Two prompts instead of one provide a checkpoint and simplify debugging.