Movement VIII showed where people are lost: 6.7% of enquiries get no reply, SLA holds at 63%. The ninth movement is about what happens in the messages: how to classify the stream, what to answer with instead of scripts, where the medical question goes, and why a negative review must not be deleted.
Taxonomy, the framework principle, escalation and four breakdowns. Enough to see the approach
Six frameworks with examples, a crisis playbook, the bot spec, the n8n workflow
response_frameworks.md, triage_taxonomy.md, sales_handoff.md, crisis_playbook.md, dify_faq_bot.md, n8n_message_routing.json
Stream volume, SLA, no-show, studio count — from movement VIII's report; economics — from the LUMÉ Bible
A network of four aesthetic-cosmetology studios in Kyiv (Pechersk, Podil, Obolon, Pozniaky), a fifth opening. Fictional. Injectables, device-based facial cosmetology, skincare, facial laser hair removal — that is, medical practice with an MoH licence, not a "beauty salon." This changes the communication: some client questions are, by law, not for an admin.
One classification across all channels. The bot's auto-reply, routing to people, SLA timers and movement VIII's reporting — which channel generates what — all hang off it.
| Code | Type | L1 owner | Reply SLA | Escalation |
|---|---|---|---|---|
| INFO | question about a procedure | admin / bot | 15 min | medical specifics → MED |
| PRICE | price, instalments, promos | admin / bot | 15 min | large package → SALES |
| BOOK | book, reschedule, cancel | admin (EasyWeek) | 15 min | — |
| MED | safety, contraindications, post-procedure condition | doctor | 30 min / now | complication → +manager |
| COMPLAINT | private complaint: service, result, staff | manager | 1 hr | harm / "lawyer" / "MoH" → L3 |
| PUBLIC_NEG | public negativity: review, comment | SMM + manager | 4 hrs | going viral → PR_RISK |
| PR_RISK | viral post, journalist, influencer | marketing lead → founder | 1 hr | crisis playbook |
| B2B | partnership, wholesale, room rental, advertising | marketing lead | 1 day | contract → lawyer |
| SPAM | spam, ads at us, troll | SMM (hide) | — | abuse → screenshot + ban |
| HR | vacancies, "will you hire me" | stub → HR e-mail | 1 day | — |
Priority when ambiguous: MED > COMPLAINT > PR_RISK > everything else. A price question alongside a complaint about a complication is MED+COMPLAINT, not PRICE.
estimate from movement VIII: ~2,398 enquiries + non-lead stream (service questions, spam, vacancies) ×1.8 → ≈ 4,300 contacts, ≈ 48 a day network-wide
1 · baseline
All messages land in a shared Instagram + Telegram inbox; the studio admin replies as best they can, in order of arrival.
2 · what's wrong
In one queue you have "how much is a facial" and "after the injection there's swelling and a fever." Both wait the same. The urgent drowns in routine, the medical reaches someone with no right to answer it, the complaint cools for hours. Movement VIII's SLA (63%) is partly a result of everything being treated as equally important, therefore equally unimportant.
3 · senior level
A classifier on the way in (keywords → LLM fallback, an n8n workflow): every message gets a type code, an owner, an SLA timer and a route to the right Telegram chat. MED and acute conditions rise above the queue. A complaint goes to the manager with a one-hour deadline. Spam is hidden with no human involved. Everything is logged to the same comms_log that movement VIII's report reads.
4 · outcome
Not "reply faster" but reply to different things differently. 10 types, 3 escalation levels, one log. The routine (INFO / PRICE / BOOK, ≈60% of items) goes to the bot, freeing the admin for bookings and hard cases.
A script is read out and sounds like a script. A framework is a principle + a structure + examples, from which the admin adapts a reply to the specific person. Six frameworks — in response_frameworks.md.
| Framework | Principle | Structure |
|---|---|---|
| Info request | give value with no pressure → a soft bridge to a consultation | an honest answer on the substance + what it solves + "the doctor will see whether it suits you" |
| Price | a range + an explanation of the spread, an exam gives the exact figure | range by zone + what it depends on + consultation ₴900, separate from the procedure |
| Complaint (HEARD) | hear · acknowledge · apologise for the experience · resolve · diagnose the cause | "sorry it turned out this way" → "I understand" → a concrete next step → logged in the CRM |
| Medical question | a non-doctor does not advise — a hard pattern, no improvisation | "this is for the doctor, an anamnesis is needed" + [acute → phone + 103] + [planned → booking] |
| Public negativity | calm, short, no client detail, move it to DMs, do not delete | name + "sorry" + "we don't discuss this publicly — medical confidentiality" + "we've written to you in DMs" |
| Handoff to sales | a structured card, not "here's a client, sort it out" | request in the client's own words · timeframe · past experience · objections · price signal · what was promised |
framework 2, channel — Instagram Direct
Why is it so expensive, it's cheaper elsewhere?
I understand. The difference is usually in the product and in who does the injecting. We work with original Botox / Dysport / Xeomin (not analogues), the injections are done by a licensed dermatologist-cosmetologist, not a cosmetologist. If price is your priority, that's fine — but honestly: on injectables, saving on the product and the specialist ends up costing more. I can tell you what's in our price, and you decide.
1 · baseline
Write 20–30 ready answers to frequent questions; the admin copies the right one.
2 · what's wrong
A ready answer is recognised instantly — a "support desk" tone, not a clinic you trust. It doesn't cover 40% of the real phrasings, breaks on a follow-up, and can't be applied to a complaint or a medical question, where precision matters, not smoothness. A script teaches reading out, not thinking.
3 · senior level
Teach the principle: each contact type has its own goal (info → value with no sale; complaint → win back trust; medical → do no harm and hand to a doctor) and its own structure. Examples are given as an illustration of the principle, not as copy-paste text. Separately — a "things we never do" list (don't promise a result, don't call it "the most effective" — art. 21, don't understate the range to hook someone).
4 · outcome
Six frameworks instead of thirty scripts. The admin answers in their own words within the principle — and it sounds like a live person from a clinic, not a voicemail. It's judged not by "word for word" but by "builds trust or spends it."
The most common wish of a beauty-business owner is to take down a bad review. For a network whose position is "the economics of trust," that's a strategic mistake, not a debatable tactic.
What a client sees with a deleted negative. Google shows "the owner did not respond." The author writes a second review — angrier, and now about how "they delete criticism." The Streisand effect kicks in: a story that 40 people would have seen gets 4,000. And for a network that sells the removal of fear in a high-risk category, "we hide what's said about us" is a direct hit on the one asset.
What we do instead. A public reply — calm, short, no client detail (medical confidentiality is a real constraint), moving the resolution to DMs. We delete only spam, abuse, doxxing and clear competitor bots — and even then with a screenshot before acting.
framework 5, Google Maps
Booked for biorevitalisation, waited 40 minutes, the doctor was annoyed that I asked questions. Won't come again.
Oksana, I'm sorry the visit left this impression — both the wait and how the conversation with the doctor went. We can't discuss the details of your case here, but we don't want to leave it like this: the studio manager has already written to you in DMs, or call us — [number]. We'll look into both the schedule that day and how the consultation went.
1 · baseline
Hide / delete negative comments and reviews so the feed and rating look clean.
2 · what's wrong
A clean feed without a single critical word reads as bought. Deleting provokes escalation and a second wave. And it contradicts the Bible's positioning: you can't sell honesty in a category where everyone is afraid — and erase the inconvenient reviews.
3 · senior level
Public negativity is a separate type (PUBLIC_NEG) with an owner (SMM answers per the framework, the manager resolves) and a 4-hour SLA. The reply is visible to every future reader: a calm response to a complaint sells better than its absence. A fake review / bot activity — the same calm tone in public + a check (was there a visit) + a report to the platform with evidence, but no public accusations of astroturfing.
4 · outcome
Zero honest reviews deleted. Each answered within 4 hours. After the situation is resolved — a public close: "we've looked into it, here's what we're changing." That's the moment trust comes back with a plus.
Injectables, laser, device procedures — medical practice (MoH order No. 117). An admin's or a bot's answer to a contraindications question is a legal risk and potential harm at the same time.
What counts as a medical question: "can I during pregnancy / breastfeeding," "I have an autoimmune condition / diabetes / herpes," "I take anticoagulants," "is it compatible with…", and any post-procedure condition — swelling, pain, fever, "the papule won't go away."
framework 4
Had biorevitalisation 5 days ago, the papules aren't resolving, I'm panicking. Is this normal?
Thank you for writing straight away. This is a question for the doctor — answering it over messages would be wrong, your anamnesis and an exam are needed. I've already passed it to the doctor who treated you; they'll be in touch. If you have a fever, severe swelling or increasing pain right now — call immediately: [number] or emergency services 103. Do you remember which product was used? — I'll check your record.
1 · baseline
The admin answers frequent medical questions from a cheat sheet ("during breastfeeding it's usually fine," "a bruise is normal, it'll pass").
2 · what's wrong
Even when the cheat sheet is right on the substance — it's a non-clinician answering a medical question: a violation (MoH order No. 117), a risk to the client (the sheet doesn't know about their autoimmune condition) and an undermining of the clinic's position. One wrong piece of advice from "LUMÉ" costs more than all the savings from fast replies.
3 · senior level
MED is a separate type with the doctor as owner; the admin and the bot only route it by a word-for-word pattern. The n8n classifier catches acute markers (fever, swelling, bleeding, allergy) separately from planned ones — the acute is duplicated to the "LUME-911" chat marked "call, don't wait for messages." The bot never answers a medical question on the substance, even if the answer is in the knowledge base.
4 · outcome
100% of medical questions to the doctor, within 30 minutes for planned ones and immediately for an acute condition. It's not slower for the client: the admin replies instantly ("passing to the doctor"), and the one with the right to answer does the answering.
LUMÉ has no sales department. "Sales" = the admin (drives to booking and payment) + the doctor at the consultation (prescribes the programme). Between them a structured handoff is needed — otherwise the doctor starts from zero, and marketing optimises the wrong metric. Spec — sales_handoff.md.
| Card field | What to write | Example |
|---|---|---|
| Request in the client's words | verbatim, not "translated" into a service | "I want my jawline not to sag, but not like celebrities" |
| Timeframe / occasion | event, deadline, season | sister's wedding in 6 weeks |
| Past experience | where, what, why they didn't like it | fillers 2 years ago, "overfilled" |
| Objections and fears | stated and read between the lines | afraid of unnaturalness, swelling at work |
| Price signal | reaction to the range | at ₴22–30K — "that's about what I expected" |
| What was promised and booked | so the doctor doesn't contradict it | consultation ₴900, Thu 15:00, Pechersk |
1 · baseline
The admin booked the client for a consultation — the doctor will sort it out on the spot.
2 · what's wrong
The doctor spends the first 10 minutes of the consultation on what the admin already knew (request, timeframe, past experience, fears). The client repeats their story a second time — that's irritating. And crucially — after the consultation no one records whether the lead was on-target, and marketing sees only CPL, not quality.
3 · senior level
A structured card (6 fields) goes to the doctor 2 hours before the consultation, the admin's pickup SLA — the same 15 minutes as in movement VIII. After the consultation the doctor (not the admin) tags it: on-target · on-target·thinking · off-target·budget · off-target·request · premature. The tags aggregate by channel in movement VIII's report.
4 · outcome
The stand-up conversation shifts from "Meta's CPL went up" to "Meta gives 22% off-target-on-budget vs 9% on Google — fix the audience, not the bid". That's the loop closing between communication and media buying.
What the admin and SMM decide themselves, and what they must raise. The full "can decide / always escalates" table — in triage_taxonomy.md.
INFO, PRICE, BOOK, a public reply per the framework, hiding spam. The default for all incoming.
reaction: 15 min · public field: 4 hrsCOMPLAINT, a decision on public negativity, B2B, any content deletion, any deviation from the framework, a query sum > ₴20K.
entry: L1 doesn't resolve it in one reply · client unhappy with L1's answer · reaction to escalation: 30 minPR_RISK, a complication threatening health, a mention of court / MoH / police, a media request, a crisis.
entry: harm to health or a legal threat · incident reach > ~10K · a journalist · reaction: immediatelyA Dify Chatflow on Instagram Direct, site chat and Telegram. It takes the routine off the admin after hours and at peak times, cleanly handing a human everything outside its scope. Spec and system prompt — dify_faq_bot.md.
"I'm an assistant, I'll help with information and booking; for medical questions and complaints I'll pass you straight to a live specialist." Stated up front, not after an error.
INFO / PRICE / BOOK / MED / COMPLAINT / SALES / B2B / OTHER. Temperature 0.
From the knowledge base: price ranges (from the Bible), procedure cards, addresses, hours. Booking — collects studio, service, day, name, phone → into Google Sheets + Telegram to the admin.
A word-for-word script + a [HANDOFF:*] tag → into the right Telegram chat (doctor, manager, marketing lead). Does not try to resolve.
Every contact → a row in comms_log: channel, type, whether there was a handoff, time to pickup. Feeds movement VIII's report.
1 · baseline
A knowledge-base bot answers anything it finds in the documents, including "can I have the procedure with my diagnosis" — the answer is in the procedure card, after all.
2 · what's wrong
A bot answering a medical question is the same mistake as a non-clinician admin, only worse-scaling: it will answer with equal confidence the one who can and the one who can't. Plus a hallucinated price or address undermines trust in the clinic for no reason.
3 · senior level
A hard red line in the system prompt: never answer a medical question on the substance — only a word-for-word script + handoff, even if the answer is in the base. Don't promise a result, don't call it "the most effective" (art. 21), don't diagnose "from a photo." Don't know the address — "I'll check with the admin," don't invent it. Control metric: the [HANDOFF:MED] share — if it drops, the bot is "sticking" to medicine and the prompt gets fixed.
4 · outcome
The bot closes ≈50% of the INFO / PRICE / BOOK stream and 100% of night-time contacts on this topic, hands a human everything else with the context preserved. The speed gain is real; the risk of a wrong medical answer is zero by design.
The same approach as in movement VIII: every node does one clear thing. n8n_message_routing.json — 25 nodes, import-ready.
| Node | What it does |
|---|---|
| Webhook | receives incoming from IG DM / comment / site chat / Telegram; acknowledges immediately |
| Normalise | an idempotency key by source (one message isn't processed twice), "is this a lead?" detection and public-channel detection |
| "This is a lead" branch | forwards to movement VIII's lead pipeline — a DM enquiry isn't lost between two workflows |
| Pre-classification | keywords catch acute medical markers and legal threats before calling the model — so it doesn't depend on the model for the critical part |
| LLM classifier (fallback) | Claude, temperature 0 — only for messages pre-classification didn't parse; if the model fails it doesn't block the stream |
| Route resolution | class → owner → SLA minutes (with Europe/Kyiv business hours) → Telegram chat; the priority rule MED > COMPLAINT > PR_RISK |
| Log + notify | a row in comms_log + a contact card to the owner with a deadline in Kyiv time |
| "Acute" branch | a duplicate to the "LUME-911" chat: "doctor call now, don't wait for messages" |
| SLA timer | waits until the deadline → checks whether it was answered → escalates to the manager if not |
| "Answered" webhook | closes the timer: time to reply, who closed it, lead-quality tag — into comms_log |
| Weekly summary (Mon 20:00) | p50/p90 reply time, SLA hit rate, bot deflection share, channels and campaign sources, lead-quality distribution → to the marketing lead |
What needs the owner's access: connecting real Instagram / Telegram webhooks, Telegram bot tokens for the studio chats, an EasyWeek key for booking from the bot, an Anthropic key for the classifier. The logic is written and imports; "live" keys are flagged and not passed off as done — as in movement VIII.
For PR_RISK and complications with a public trail. The goal is not to "hush it up" but to pass through the incident so that trust afterwards is higher than before. Full protocol — crisis_playbook.md.
Screenshot, link, reach, author, substance.
Founder + chief doctor + marketing lead in the "LUME-911" group. A lawyer — if there's a legal threat.
Was there a visit, what procedure, who treated them, what's in the informed consent. Separate the facts from the author's version.
Is a public position needed or is a private contact enough; who speaks for LUMÉ (by default — the founder or the medical director, not SMM); the first step for the client.
A call, acknowledgement, a concrete offer (an exam with an independent doctor, a review, compensation — as fits). A public reply — only if the story is already public.
Principles of a public reply: fast (silence > 24 hrs reads as "hiding"), acknowledge the person's experience without arguing the facts, show action not intention ("the medical director has been in touch, an exam is set for…"), don't hide behind legal wording, one voice for the company, after resolution — close it publicly with an account of what's changing in the process.
Things we never do: don't delete the original post or pay for its removal, don't publish the correspondence with the client, don't "expose" their version, don't leave ads on those accounts running during the crisis, don't reply at night on emotion — there's a trio and a protocol.
Communications don't create a separate ledger; they fill comms_log, which movement VIII's report reads.
Every metric is tied to a decision: speed → admin headcount and shifts; bot share → prompt and knowledge-base tuning; [HANDOFF:MED] share → control that the bot doesn't advise; lead-quality tags → targeting by channel (Section V).
Stream volume is an estimate, not a measurement. ≈4,300 contacts/quarter is derived from movement VIII's 2,398 enquiries with a ×1.8 coefficient for the non-lead stream. The real ratio depends on account activity and ad pressure; with live channels connected it's refined against actuals.
The frameworks haven't been field-tested. The example replies are built on the Bible's positioning and industry practice, but real tone calibration is reviewing actual dialogues with admins, which can't exist on a model client.
The n8n classifier is a working skeleton. The pre-classification regexes cover frequent phrasings; the LLM fallback fires on rare ones and slang. Accuracy on a real stream needs measurement and fine-tuning on the first 200–300 labelled messages.
The legal frame is from open sources. MoH order No. 117, Advertising Law art. 21 — read, not advised on by a specialist lawyer. For a real business the crisis playbook and moderation rules are agreed with a lawyer.
research/source_register.md, block Hfigures.json