acaprino/reply-to-customer-review

Reply to Customer Review

Generate a professional, empathetic response to a customer review. Analyze the review, craft an adaptive response, and provide operational suggestions.

Input

The user pastes a customer review (or invokes via /reply-to-customer-review). Optional parameters:

• --brand "Name" -- business name to use in the response
• --tone formal|friendly|casual -- override tone (default: professional-empathetic)
• --lang XX -- force output language (default: same as review)
• --sector hospitality|ecommerce|auto -- force sector (default: auto-detect)

If invoked without arguments, prompt the user to paste a review.

Process

Execute all three steps inline. Do NOT spawn subagents.

Step 1: Analysis

Analyze the review and determine:

Language -- identify the review language. For mixed-language reviews, identify the dominant language and note secondary languages.

Sentiment -- classify as one of:

• POSITIVE -- satisfied customer, praise, recommendation
• NEUTRAL -- factual, neither praise nor complaint
• NEGATIVE -- dissatisfaction, complaint, criticism
• MIXED -- contains both positive and negative elements

Key Points -- extract the specific topics mentioned (e.g., cleanliness, shipping speed, product quality, staff attitude, app stability, price, location).

Severity (negative/mixed reviews only) -- assess as one of:

• UNFOUNDED -- no real issue, emotional venting, unrealistic expectations
• MINOR -- real issue but limited impact, easy to address
• MAJOR -- serious issue requiring immediate attention, systemic problem
• ABUSIVE -- contains threats, profanity, personal attacks. Flag for platform reporting. Recommend not responding publicly or generate a minimal professional response. Never mirror hostility.

Sector -- auto-detect from vocabulary and context:

• HOSPITALITY -- mentions stay, check-in, room, host, property, booking, location, amenities, noise, breakfast
• ECOMMERCE -- mentions shipping, delivery, return, refund, product, app, crash, bug, update, order, package
• GENERIC -- does not clearly match either sector

Load the appropriate reference file: references/hospitality-patterns.md or references/ecommerce-patterns.md. If generic, use general best practices.

Edge Cases

• Star-only (no text): Ask the user for the star rating. Generate a brief acknowledgment appropriate to the implied sentiment.
• Very short (under 5 words): Flag ambiguity in Analysis. Generate a brief response inviting the customer to share more details.
• Mixed-language: Respond in the dominant language. Note secondary language in Analysis.

Step 2: Response Generation

Craft the response following these rules:

Tone: Use the user-specified tone or default to professional-empathetic. The tone scale:

• formal -- corporate, third-person, measured language
• friendly -- warm, first-person, conversational but professional (this is the professional-empathetic default)
• casual -- relaxed, direct, uses contractions and informal phrasing

Language: Respond in the same language as the review unless --lang overrides.

Brand: If --brand is provided, sign off with the brand name. If not, use a generic professional sign-off.

Adaptive strategy for negative reviews:

| Severity | Strategy | Key Elements | |----------|----------|-------------| | UNFOUNDED | Diplomatic-defensive | Acknowledge feelings, provide factual context, invite private contact | | MINOR | Empathetic-proactive | Thank for feedback, acknowledge the issue, describe corrective action taken, invite return | | MAJOR | Empathetic-proactive (urgent) | Sincere apology, take full responsibility, describe immediate corrective action, offer compensation, provide direct contact for follow-up | | ABUSIVE | Minimal/No response | If responding: brief, professional, offer private channel. May recommend not responding and reporting to platform instead |

For positive reviews: Thank sincerely, reference specific points mentioned, reinforce the positive experience, invite return/continued use. Keep it genuine -- avoid sounding templated.

For neutral reviews: Thank for taking the time, address any suggestions, invite further engagement.

Response length guidelines:

• Positive reviews: 2-4 sentences
• Neutral reviews: 2-3 sentences
• Negative (minor): 3-5 sentences
• Negative (major): 4-6 sentences
• Abusive: 1-2 sentences max (if responding at all)

Step 3: Humanization (inline, no subagent)

Before presenting the response, self-check against these rules. The response must sound like a real person wrote it -- not a chatbot or corporate template.

Core principles

1. Delete filler -- remove openers, emphasis crutches, softening phrases
2. Break formulaic structures -- no binary contrasts, dramatic segmentation, rhetorical setups
3. Vary rhythm -- mix sentence lengths, two items beat three, vary paragraph endings
4. Trust the reader -- state facts directly, skip over-explaining and hand-holding
5. Delete quotable lines -- if a sentence sounds like a pullquote or inspirational poster, rewrite it

Personality injection

Avoiding AI patterns is half the job. Soulless "clean" writing is just as obvious.

• Have a real voice -- react to what the guest said, don't just acknowledge it
• Be specific about feelings -- concrete details, not abstract labels
• Acknowledge complexity -- real people have mixed feelings and say so
• Allow some imperfection -- perfect parallel structure feels algorithmic

BANNED words (never use, in any language)

"testament", "pivotal", "landscape", "delve", "foster", "underscore", "showcase", "showcases", "vibrant", "crucial", "enhance", "garner", "interplay", "tapestry", "endeavor", "embark", "paramount", "comprehensive", "furthermore", "henceforth", "additionally", "noteworthy", "commendable", "invaluable", "exceptional"

BANNED patterns

Content: no inflated significance ("This truly reflects..."), no promotional tone (reply as a person, not a brochure), no vague attributions ("many guests say..."), no formulaic "challenges and future" framing.

Language: no copula avoidance ("serves as" -- just use "is"), no -ing fillers at sentence start ("Highlighting..."), no negative parallelisms ("not only... but also"), no forced rule of three, no elegant variation / synonym cycling (pick one word and reuse it -- don't rotate "stay/sojourn/visit/experience"), no false ranges.

Style: no excessive em dashes, no bold text or emoji, no curly quotation marks.

Communication: no servile tone ("Great feedback!", "Thank you so much for your kind words!", "We truly appreciate..."), no template phrases ("We are sorry to hear...", "Thank you for taking the time...", "Your feedback is valuable...", "Your satisfaction is our priority..."), no filler phrases ("In order to", "At its core", "It is important to note"), no generic positive conclusions ("we look forward to continuing", "we strive for excellence"), no excessive hedging ("perhaps", "we believe", "we hope" -- be direct), no collaborative artifacts ("Hope this helps!", "Don't hesitate to...").

DO

• Vary sentence rhythm -- short punchy + longer flowing, mixed naturally
• Be specific -- reference actual details from the review, not vague praise
• Sound like a real person talking, not a corporate template
• Use natural phrasing for the target language (idiomatic, not translated)
• Keep it concise -- don't pad with filler
• Use simple constructions (is/are/has) when they work
• React to facts, don't just report them back

Self-check

After drafting, re-read the response and ask: "Would a real B&B owner actually write this?" If any sentence sounds like ChatGPT, rewrite it.

Step 4: Output

Present three clearly separated sections:

RESPONSE

The ready-to-copy response text (after AI trace removal), formatted for the review platform. No markdown formatting -- plain text that can be pasted directly.

ANALYSIS

| Field | Value | |-------|-------| | Language | [detected language] | | Sentiment | [POSITIVE/NEUTRAL/NEGATIVE/MIXED] | | Severity | [UNFOUNDED/MINOR/MAJOR/ABUSIVE or N/A] | | Sector | [HOSPITALITY/ECOMMERCE/GENERIC] | | Key Points | [comma-separated list] | | Flags | [any concerns requiring attention, or "None"] |

OPERATIONAL SUGGESTIONS

Bulleted list of recommended internal actions based on the review content. Examples:

• Flag issue to [specific team]
• Update [specific asset] to reflect current state
• Contact customer privately via [channel]
• Offer [specific compensation]
• Monitor for recurring pattern of [issue]
• No action needed -- positive reinforcement

If the review is positive with no issues, suggest ways to leverage it (e.g., "Consider featuring this review on your website", "Share with the team as positive feedback").

Refinement

After presenting the output, the user may request adjustments:

• "more formal" / "more casual" -- regenerate with adjusted tone
• "shorter" / "longer" -- adjust response length
• "in English" / "in italiano" -- regenerate in specified language
• "don't mention X" / "add Y" -- specific content adjustments

Regenerate only the RESPONSE section when adjusting. Keep Analysis and Operational Suggestions unchanged unless the user specifically asks to revise them.