Candle Troubleshooting Replies: Tunnelling, Sooting, Weak Scent

Give the AI your own cause list, care wording and replacement rules, then have it draft each reply: questions about burn time and wick length, a short likely cause, care advice, and a replacement only where your rule says so. A person reads every reply, and anything about flames or cracked glass goes to you at once.

The usual trap is asking an assistant to "reply to this candle complaint". It will write something kind, confident and quite possibly wrong, such as a burn-time rule you never tested or a promise of a free replacement you never offered. The fix is to do the thinking once, in writing, and let AI do the repetition.

Follow me on Instagram@sagnikteaches

Replies to the customer, not fixes to the recipe

Everything below is about replying to customers, not about fixing your candle recipe, choosing wicks or testing fragrance load. If a complaint pattern keeps coming back, that is a product problem, and the fix belongs in your testing. The replies below deal with the customer in front of you: what to ask, what to say about the likely cause, what care advice to give and when to replace the candle.

Connect on LinkedInSagnik Bhattacharya

Say a small candle maker sells about 600 candles a month through her own shop and a marketplace. In a typical month 41 customers get in touch about how a candle performs: 19 about tunnelling, 9 about black soot or smoke, 8 about weak scent and 5 about something else. That is about 7 contacts per 100 candles sold. She answers each by hand in about six minutes, which is nearly four hours a month, and her replies vary with her mood and how many she has already written that day. This is the illustration used through the page. The figures are invented, but the shape is common in small candle shops: a few complaint types dominate.

Subscribe on YouTube@sagnikteaches

Step 1: write the cause lists once

For each complaint type, write a list of likely causes in your own words, in the order you would check them, with the question that tells you which one it is. A cause list is not a diagnosis of every case. It is a way to ask the customer the right two questions and avoid a long email thread.

Tunnelling

Tunnelling means the flame burns a narrow hole down the middle and leaves wax at the walls. The causes you can check from a message:

• Short burns. The first burn was stopped before the melted wax reached the edge of the container, so the candle formed a "memory" ring. Ask: how long did the first burn last?
• Draught. The flame leans, and one side melts faster. Ask: is it near a window, vent or fan?
• Wick length. A long, untrimmed wick gives a big flame with soot; a very short one gives a pool that cannot reach the edge. Ask for the wick length.
• A container wider than the wick can melt. This is a product cause and a reason for replacement if your batch notes show it. It is not something a customer can fix.

Sooting

Soot is black deposit on the container, the ceiling or the wall, or smoke from the flame. Likely causes:

• Long wick or mushrooming. A carbon bulb on the wick tip, or a wick left untrimmed. Ask: did they trim the wick before lighting, and how long is it?
• Draught. A flickering flame burns less cleanly.
• Burning too long in one session. The wick grows and the flame gets bigger.
• A wick or fragrance load that does not suit the container. A product cause.

Weak scent

The complaint "I can't smell it" needs more care than the others, because scent is partly a matter of the room and the nose. Ask whether they mean the unlit candle (cold throw) or when burning (hot throw), and consider:

• Cure time. Some makers allow a cure period after pouring, before the scent is at its best. Use your own tested figure; do not let the AI supply one.
• Room size and ventilation. A single candle in a large or draughty room reads as faint.
• Nose fatigue. After a while, people stop noticing a smell they are in. Walking out of the room and back in changes the answer.
• Short burns. Without a full melt pool, less fragrance is released.
• Product cause. Fragrance load or oil choice in a particular batch.

The point of the list is the questions. If you can get two facts, such as first-burn length and wick length, you can usually point to a cause and choose the right reply.

Step 2: write the care card the AI must quote

Customer-facing advice should come from your tested guidance, not from a general idea of what candles need. Put your care instructions on one page and treat it as the single source. Most care cards cover the same ground: burning long enough on the first use for the melt pool to reach the edge, trimming the wick before each burn, keeping the candle away from draughts and flammable items, and never leaving it unattended. The candle trade association's consumer tips, for instance, say to trim the wick to a quarter of an inch (about 6 mm) before every use and link flickering and soot to draughts and vents. Any first-burn time, maximum burn session or wick length on your card should still be yours, based on your own wax, wick and container tests, because a generic figure may not suit your jar.

Your labels and care card must agree with each other. If the card says to trim to 5 mm and the label says 6 mm, a customer will notice. The candle warning labels checklist covers the label side; build both from the same source of truth.

Step 3: write the replacement rules

Do this before you open an AI tool, because the rules are decisions about money and goodwill. A rule list in plain words, kept to a page, is enough. Here is an illustrative set a small maker might choose:

The details are yours to change. What matters is that the AI can quote the rule and does not invent a more generous or a stingier one. If your marketplace has its own returns rules, those apply on top, so note that beside each row.

Step 4: turn it into a prompt that cannot wander

Now combine the three sources into a single instruction. Keep it in one saved prompt that your team uses, rather than rewriting it each time. If you use ChatGPT or Claude, a Project keeps the instruction and your documents together (both vendors' help pages describe project instructions plus uploaded files), and a business plan keeps that content out of model training by default. Custom GPTs are scheduled to stop running on 11 December 2026, so do not build the routine on one.

You draft customer replies for a small candle company. Use ONLY the cause lists, care card and replacement rules below. Never invent burn times, wick lengths, cure times or promises. If the message is not covered, write NOT COVERED and say what the maker needs to decide.

Steps:
1. Classify the message: tunnelling, sooting, weak scent, safety, other.
2. If safety (flames, cracked or shattered glass, burns, fire, a child or pet involved), output only: SAFETY, then use the safety script word for word.
3. Otherwise list which facts are still missing (first-burn length, wick length, draughts, batch code, photo) and ask for at most two.
4. If enough facts are given, name the most likely cause from the list in one sentence without blaming the customer, give the care advice, and apply the replacement rule.
5. Tone: warm, brief, under 130 words, plain English, no exclamation marks.

[Paste cause lists, care card, replacement rules, safety script]

Customer message:
[paste]

A realistic first run on the message "My lavender candle has dug a tunnel and there's loads of wax left on the sides. Very disappointed" returns a reply asking two things: how long the first burn lasted and where the candle sits, and explaining in a sentence that a first burn that stops early leaves a ring the candle then follows. It does not promise a replacement, because the rules need those facts first. That is the correct behaviour, and it is the main reason to write the rules down.

Step 5: the reply templates, filled in

Below are three illustrative replies. Treat them as the sort of output you want from the assistant, after you have read and edited them. The first-name placeholder stays until you send.

Tunnelling, facts missing

"Hi [first name], thank you for telling us, and I'm sorry the candle hasn't burned the way you hoped. A tunnel usually comes down to how the first burn went, so two quick questions: roughly how long did you leave it lit the first time, and does it sit near a window or vent? A photo from above would also help. Once I know, I can tell you whether it can be rescued or whether we should send you another."

Tunnelling, short first burns, care advice

"Thanks, [first name]. That helps. Candles tend to follow the pool of wax they made on the first burn, so if the first couple of burns stopped before the wax reached the edge, the flame keeps to the narrow channel. The good news is you can usually even it out: let the candle burn for as long as our care card suggests for this size, in a still spot, until the surface is fully melted across, and trim the wick to about 5 mm before each burn. If it hasn't improved after two long burns, reply with a photo and we'll look again."

That figure of 5 mm has to come from your care card. If you do not give the assistant a number, it must write "to the length on your care card" instead.

Sooting, with the batch code

"Hi [first name], thank you for the photo. The black on the glass looks like soot, which can happen when the wick is long or the flame flickers. Could you tell me the wick length when you lit it and whether there is a draught nearby? If the wick was trimmed and the air was still, I'd like to replace the candle and look at the batch, so please send the code from the base."

Step 6: the safety script is not for editing

Some messages are not troubleshooting. A cracked jar, a flame that rose far above the container, a candle that caught something alight or a burn to a person is a safety matter. Write one safety reply yourself, with a lawyer's or insurer's eye if you have one, and make it a fixed script the assistant must use word for word. The script should ask the customer to stop using the candle, extinguish it only if it is safe to do so, keep the candle and packaging, send photos and the batch code, and say that a person will be in touch. It should not discuss fault, compensation or cause. Route the ticket to a person immediately, and log it. Your insurer may have a view on how to record these, so ask.

Make the classification step in the prompt conservative: if there is any doubt, mark it SAFETY. A false alarm costs you a minute; a missed one can cost far more. Check the weekly sample specifically for safety messages that were not classified as such. The weekly spot-check routine works well for this.

A walkthrough of one week of complaints

Say six messages arrive on a Monday. Here is how they go through the routine, with the AI drafting and a person deciding.

1. "Wick keeps smoking." The assistant classifies as sooting, asks for wick length and draughts. Draft goes out in two minutes after a read.
2. "Scent is barely there, second candle I've bought." Weak scent, repeat customer. The assistant drafts a reply asking which scent and room size, and flags "repeat purchase" for the maker. The maker decides to include a small sample of a stronger scent.
3. "Candle cracked the glass last night." Marked SAFETY. The fixed script goes out, the maker is alerted, and she calls the customer the same day.
4. "Tunnelled after two burns, photos attached." Facts supplied: first burn about 45 minutes. The rule says care advice. The assistant drafts it; the maker adds a line of her own.
5. "Smells different from the one I bought last year." The assistant says NOT COVERED. The maker replies herself, because it may be a formulation change.
6. "Can I return this? I don't like it." Not a performance complaint. The assistant points to the returns policy and drafts a reply from it.

That is four drafts used, one safety call and one handled by the maker alone. Time spent: about 35 minutes against roughly 40 in the old way, less impressive than hoped for a single day. The saving comes from consistency and from the batch log, and it grows as volumes rise. The store conversations to copy show how similar flows look in other shops.

Keeping the batch log

Every complaint that qualifies for a replacement should record the batch code, scent, container size, date and what the customer described. This is the log that turns replies into product knowledge. Ask the assistant to produce one row per ticket from the message and your classification: it is good at extracting facts into columns. Once a month, ask it to count complaints by batch and cause. If three tunnelling cases trace to one batch of a 300 ml jar, you have found a product problem, and replies cannot fix that. Pass it to whoever tests wicks and fragrance load.

A tidy log also supports good decisions about replacement costs. In our illustration, 41 contacts a month with 12 replacements at a landed cost of, say, $6 each is $72 a month. Seeing that number next to the 4 hours of reply time shows where the effort should go.

What can go wrong with AI-written candle replies

The invented figure

Say the assistant tells a customer to burn the candle for "at least four hours at a time" because that is a figure it has seen often. Your own tested guidance says three. The reply is now inconsistent with the label. The prevention is the instruction to quote only the care card, and a weekly check for any number in a draft that does not appear in your card.

The blame

Drafts can drift into suggesting the customer did something wrong: "you should have trimmed the wick". That is accurate more often than it is useful, and it loses repeat customers. The prompt asks for a neutral cause sentence for that reason. Read for tone, and keep a short list of phrases you never want to see.

The over-generous promise

An unguarded assistant offers refunds and free products as a courtesy. Your rule table prevents that; the weekly sample checks it.

Adapting this to a marketplace or a trade customer

If you sell through a marketplace, its messaging tool and returns rules sit on top of yours. Keep customer data in the marketplace where required and paste only what you need into the assistant, such as the message text and product name, with names and addresses removed. If you supply shops, the replies change tone: a retailer wants a batch code, a count of affected units and a clear route to credit, and a short template for that is worth writing separately. The returns and refunds wording tutorial has policy language you can adapt, and the five jobs for AI in a candle business puts customer replies in context with the rest.

If you want help turning your complaints into a working reply routine, that is what my AI implementation consultation covers. Start with the cause lists and the rule table; they are worth having even if you never use an AI tool.

Questions candle sellers ask about AI-written replies

Can AI diagnose why a customer's candle is tunnelling?

It can sort likely causes from what the customer tells you, such as burn time, draughts and wick length, and ask the right follow-up questions. It cannot see the candle, so treat its answer as a short list of possibilities, not a verdict.

Should every complaint get a replacement?

No. Set rules by cause and evidence: a candle with a clear production fault gets a replacement, while one that tunnelled after repeated short burns gets care advice. A written rule keeps replies consistent and stops the AI inventing promises.

What should I do if a customer mentions flames or a cracked jar?

Tell them to stop using the candle and extinguish it safely, ask for photos and the batch code, and escalate to a person at once. Keep that reply a fixed script and do not let the AI improvise on safety.

How do I stop the AI inventing burn times or wick advice?

Give it your own care card and cause lists, and tell it to use only that material and to say NOT COVERED otherwise. Then check a sample of drafts each week against your tested figures. A quick test: find every number in a draft and confirm it appears on your card.

Should customers receive the same reply for a different scent or wax?

Not always. Different waxes and fragrance loads behave differently, so keep a short note per product line. Give the assistant that note along with the ticket so the care advice matches what the customer bought. A wide three-wick jar and a small tin rarely share the same first-burn advice.

Further reads

• AI Ticket Triage: Tag, Route, and Prioritise Support Requests — Tag, route and prioritise requests before anyone replies.
• Your First 30 Days of AI Customer Support for an Online Shop — A 30-day plan for AI support in an online shop.
• Peak Season Ecommerce Checklist: Stock, Support and Shipping — Stock, support and shipping before the busiest weeks.
• Ecommerce Return Fraud: Find Repeat Abusers in Your Returns Data — Spot repeat claimants in your returns data.
• AI for WooCommerce Stores: Descriptions and Support Replies — Support replies and descriptions for a WooCommerce store.
• Ecommerce AI Chatbot Design: Order Status, Returns and Sizing Flows — Design order, returns and sizing flows for a store chatbot.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the seller's own tested burn guidance (to be supplied by the maker); the candle trade association's consumer safety and expert tips pages at candles.org; OpenAI help articles on Projects and on the custom GPT retirement; Claude help article on Projects (all checked 5 November 2026).