Hey, it's Cindy 🌱 You commented closer, so here is the part every "connect AI to your DMs" tutorial skips. The thing that decides whether a lead buys is the system prompt.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
Read this before you build anything. An assistant in your DMs is answering real people on your behalf, so this guide builds it the safe way: it replies only to people who message you first, it says out loud that it is an assistant, it hands off to you the moment a conversation gets sensitive, and for the first two weeks it drafts replies for you to approve instead of sending on its own. Every rule below is baked into the prompts. Do not strip them out to make it faster.
why this is the whole job
The brain versus the wiring
Connecting a DM tool to Claude is plumbing. You paste an API key, you add a step to a flow, you turn it on. Anyone can do that in an afternoon, which is why every tutorial stops there.
What you are actually shipping is a system prompt: the standing instructions your assistant reads before every single reply. It holds your offer, your price, what you refuse to promise, the questions that separate a real lead from a browser, and the exact moment it should stop talking and tag you in. Most DM bots sound like a broken robot because that document is three sad sentences long. A good one runs to a page and reads like an onboarding doc for a new sales hire, because that is exactly what it is.
Good news: a page-long system prompt is boring to write and Claude is very good at boring structured writing. You supply the facts about your business, Claude supplies the structure, the tone rules and the guardrails you would have forgotten.
start here
Build it from zero
🗒️ Write your offer facts down first. Before you open anything, get five things on paper: what you sell, the real price, who it is genuinely for, who it is not for, and the one action you want a qualified person to take next. Vague inputs are the number one reason these assistants sound generic.
💬 Open Claude and run the writer prompt. Go to Claude (the free plan handles this fine) and paste the writer prompt below. It interviews you first, then writes the full system prompt.
🔍 Read every line of what it produces. You are checking three things: are the facts right, are the promises ones you can keep, and does the escalation rule fire early enough. Fix the facts yourself, do not ask Claude to guess them.
🧪 Stress test it before it touches a real person. Run the red-team prompt further down. It throws fifteen hard DMs at your assistant so you find the embarrassing answers in a chat window instead of in someone's inbox.
🔌 Wire it into your DM tool. Paste the finished prompt into the assistant field of whatever tool handles your DMs. The wiring section below covers the ManyChat route, which is the one most creators use.
👀 Run it in review mode for two weeks. Have it draft, you approve. You will edit roughly the first thirty replies, and every edit is a line you add back into the system prompt. After that it earns the right to send on its own.
copy this
Prompt 1: the one that writes your assistant
This is the prompt from the reel. Paste it into a new Claude chat and answer its questions properly.
✍️ write my DM assistant prompt
I want you to write the complete system prompt for an AI assistant that replies to inbound Instagram DMs about my offer. Before you write anything, interview me: ask me what I sell, the exact price and what is included, who it is genuinely right for, who it is wrong for, the three objections I hear most often, the single next step I want a qualified person to take, and three examples of how I actually talk so you can match my tone.
Then write the system prompt as a structured document with these sections: role and identity, the facts about the offer, the qualifying questions to work through conversationally, how to handle each objection honestly, tone rules, hard guardrails, and an escalation rule. It must open by identifying itself as an assistant for my account rather than pretending to be me, it must never invent a price, a discount, a result or a guarantee that is not in the facts section, it must never claim a specific income outcome, and it must reply only to people who have messaged me first.
The escalation rule matters most: write it so the assistant stops and tags me in whenever someone asks something not covered by the facts, sounds distressed or vulnerable, mentions money they cannot afford, asks for a refund or a legal or medical opinion, or asks the same question twice. When it escalates it should say plainly that it is passing the conversation to a human and roughly when I will reply.
Keep every message it writes under sixty words, one question at a time, no stacked questions, no hype language, no emoji spam. Finish by giving me a short checklist of the facts I need to keep updated in this prompt whenever my offer changes.
steal this shape
What a finished one looks like
Claude writes yours from your answers, but here is a complete example so you know what you are aiming for. This is the structure that makes the difference, so swap the details and keep the bones.
📄 example system prompt
ROLE
You are the assistant for @examplestudio on Instagram. You reply to people who have messaged this account first. In your first message, say you are the studio's assistant and that a human reads every conversation. Never claim to be the owner and never use her name as a signature.
THE FACTS (only these are true, never add to them)
- Offer: a 6-week 1:1 brand design programme.
- Price: 2,400 AUD, payable in full or in three monthly instalments of 850 AUD.
- Included: weekly 60-minute calls, a full logo and brand kit, two rounds of revisions, and a Notion handover doc.
- Right for: service businesses with an existing website that are rebranding.
- Not right for: pre-launch businesses with no offer yet, and anyone needing it delivered in under 4 weeks.
- Next step: a 20-minute fit call booked through the link in bio.
QUALIFYING (conversational, one question at a time, never a form)
1. What does your business do right now, and who buys from you?
2. What is making you look at rebranding this month specifically?
3. Do you have a timeline you are working to?
4. Is a 2,400 AUD investment workable for you right now?
If the answers show a good fit, invite them to the fit call. If they are pre-launch or need it in under 4 weeks, say so honestly and point them to the free brand checklist instead of pushing the programme.
OBJECTIONS (answer honestly, never pressure)
- "That is more than I expected." Acknowledge it, restate exactly what is included, mention the three-instalment option once, and stop. Do not offer a discount, there is no discount.
- "Can you guarantee more sales?" No. Say clearly that branding is not a sales guarantee, describe what the programme does deliver, and offer two past examples if asked.
- "I need to think about it." Agree that it is worth thinking about, ask if there is a specific unknown you can answer, and offer to send the programme outline. Do not follow up more than once.
TONE
Warm, plain, specific. Under 60 words per message. One question at a time. No hype, no fake urgency, no stacked emoji. Match their formality. Australian English.
GUARDRAILS
Never invent a price, discount, timeline, result or guarantee. Never give legal, medical, financial or visa advice. Never comment on a person's appearance, body or personal circumstances. Never ask for payment details in DMs. If you are unsure of a fact, say you will check with a human rather than guessing.
ESCALATION (do this early rather than late)
Stop and tag the owner in when: the question is not covered above, the person sounds distressed or vulnerable, they mention money stress, they ask about refunds, contracts or legal terms, they are asking on behalf of a brand or agency, or they repeat a question you already answered. When you escalate, say: "That one is better answered by a person, so I am passing this to the team now and someone will reply here." Then stop replying in that thread.
Notice what makes it work: every claim traces back to a facts list, the objection answers include a real "no", and the escalation rule fires early. Those three things are what separate an assistant people trust from a bot people screenshot.
the six parts
What every good one contains
Identity Who it is, whose account it speaks for, and an upfront line saying it is an assistant. Never a fake human name.
A closed facts list Price, inclusions, timeline, fit. The rule "only these are true" is what stops invented discounts.
Qualifying questions Three or four, asked one at a time, in an order that ends with the budget question rather than opening on it.
Honest objection answers Including the ones where the honest answer is no. A bot that cannot say no reads as a bot immediately.
Guardrails The topics it refuses: legal, medical, financial advice, payment details, anything personal about the sender.
An escalation trigger The named list of moments it hands over to you, plus the exact sentence it uses when it does.
before it goes live
Prompt 2: stress test it
Paste your finished system prompt into a fresh Claude chat, then paste this underneath it.
🧪 red-team my assistant
Above is the system prompt for an assistant that replies to inbound Instagram DMs for my business. Your job is to break it, not to praise it.
Write fifteen realistic DMs designed to make it fail, and for each one show the reply my prompt would produce and then rate it pass or fail with one line on why. Include at least: someone pushing hard for a discount, someone asking for a guaranteed result, someone who cannot afford it and is clearly stressed about money, someone asking a question my facts section does not cover, someone asking for legal or tax advice, someone trying to get it to say it is a real human, someone sending a long rambling message with three questions at once, someone being rude, someone asking about a competitor, and someone asking it to hold a spot without paying.
Then give me a specific rewrite for every failure: quote the exact line to add or change in the system prompt, not general advice. Finish with the three biggest remaining risks in this prompt and what a person could still get out of it that I would not want said in my name.
Do this before every launch, not once. Whenever your price, offer or timeline changes, the facts section changes, and a stale facts section is how an assistant confidently quotes last year's price.
the plumbing
Wire it into your DMs
ManyChat is the tool most creators use for this, and it has a native Claude integration, so you do not need a middleware layer. Here is the honest version of the setup, including what it costs.
🔑 Create an Anthropic API key. This is separate from your Claude chat subscription. Sign in at the Anthropic Console API keys page, create a key, and copy it somewhere safe because you cannot view it again after you close the dialog. Getting-started detail is in the API docs.
💳 Understand the two bills. The API is pay as you go and billed per token, priced per model on the Anthropic pricing page, and your DM tool has its own subscription on top. A long system prompt is re-read on every reply, so a wordy prompt costs more per conversation than a tight one.
🔗 Connect the key in ManyChat. In ManyChat, open Settings, then Integrations, find the Claude integration, paste your API key and connect. ManyChat documents this in its own Claude integration help article, and integration screens move around, so if the layout differs from this, trust what is on your screen over any tutorial including mine.
🧵 Build the flow. Create a new automation with your trigger, which should be an inbound DM or a comment on a post, never an outbound list. Add the Claude action step, pick your model, and paste your system prompt into the instructions field. Keep the reply length limit you wrote into the prompt.
🙋 Add the human handoff. Add a branch that ends the automation and notifies you whenever the assistant escalates or the person types something like "human", "person" or "speak to someone". If your tool cannot do this, do not run the assistant unattended.
🚦 Turn it on for one entry point first. One keyword, one post, one week. Read every conversation. Then widen it.
Cheaper starter version with no API key at all: run the same system prompt inside a normal Claude Project, paste in real DMs each morning, and reply yourself with the drafts it writes. You get most of the value, you keep full control, and you find out whether the prompt is any good before you pay for anything.
stay out of trouble
The rules that protect your account
Automating replies to people who contacted you is normal. Automating outbound messages to people who did not is how accounts get restricted. Meta's spam policy covers bulk and unsolicited contact, and Instagram publishes its own community guidelines. Keep the trigger inbound-only, forever.
Disclose it. The first message says it is an assistant. Letting someone believe they are talking to you personally when they are not is the fastest way to lose trust, and it is the one thing people never forgive.
Never harvest DMs into a mailing list. If you want their email, ask for it explicitly and say what you will send. In Australia that consent requirement is spelled out by the ACMA spam rules, and most countries have an equivalent.
Keep a person in the loop. Read the transcripts daily at first, weekly after that. An assistant nobody reads drifts, and you find out from a screenshot.
No pressure tactics. No invented deadlines, no fake spot counts, no "only two left" unless it is literally true. Your prompt should ban these explicitly.
Do not take payment details in DMs. Always send people to your real checkout link. This one is in the guardrails for a reason.