Voice DNA · 5 min read
Training Your Instagram Bot: What to Feed It and What to Fix
"Training" is the wrong word and it sets the wrong expectation, so it is worth spending a paragraph on before anything else.
You are not training a model. Nothing you do here changes how the model works. What you are doing is assembling a reference — a set of replies you wrote, which new replies get written against. That distinction matters practically, because it tells you what fixing a bad reply looks like. You are not adjusting weights. You are adding evidence, or you are moving the message somewhere it should have gone in the first place.
What goes in
Replies the account owner actually wrote. That is the whole eligibility rule, and everything else follows from it.
Which means these do not count, however tempting:
- Your website copy. Nobody DMs the way they write an About page.
- Your captions. Captions are performance. DMs are conversation. Most creators run two completely different registers and know it.
- A description of your voice. "Friendly but not unprofessional, a bit playful" is a sentence about you, not a sample of you. It is also, usually, inaccurate.
- Somebody else's replies. Your VA's, an agency's, a competitor's screenshot you liked. Feed those in and the profile is genuinely accurate — to them.
How much is enough
The useful question is not how many. It is whether the range is covered.
Volume without range gets you a profile that is confident about the four questions you answer daily and lost on everything else. Range with modest volume gets you something usable. Here is the coverage checklist worth running against your own outbox:
- Your most common question, answered on a normal day. The baseline register. This one you have hundreds of.
- The same question answered when you were busy. Shorter, blunter, fewer pleasantries. This is half your real voice and most people never think to include it.
- A refusal. Out of stock, not taking clients, we don't ship there. Saying no is the register generic models cannot reach, and if your outbox contains none, nothing can synthesise it.
- An apology. A late reply, a delayed order, a mistake. Not a complaint handled — those should not be automated at all — just the ordinary "sorry, missed this".
- A reply to someone you know. Repeat customers get a different register from strangers, in almost everybody's inbox. If only stranger-replies go in, every reply comes out at stranger distance.
If a category is empty, the honest fix is not to write specimens for it. It is to answer that kind of message yourself for a while, and let the record fill in from real ones. Specimens written for the profile come out in the register people use when they know they are being observed, which is the formal one.
Do the templates first
This is the step people skip, and skipping it is why some accounts conclude that AI replies are expensive and unreliable at the same time.
Templated replies are free and unlimited. A credit is spent only when the AI writes a new reply — one credit, one reply. If four questions cover most of your inbox and you write four good templates, most of your automation runs at no marginal cost and is perfectly in voice, because you typed it.
Then test on yourself before anything goes live. The Test on myself button runs the complete pipeline and delivers the result to your own account, so you read the reply as its recipient rather than as its author. Text you wrote reads correctly to you in the builder in a way it will not read in a DM.
When it gets you wrong
Three failure shapes, three different fixes. Diagnosing which one you have is most of the work.
Wrong fact, right voice. The reply sounds exactly like you and says something untrue — a price you changed, a delivery window you no longer promise. This is not a voice problem and no amount of extra reference material fixes it. Move that answer into a template, where the fact is something you typed rather than something inferred.
Right fact, wrong register. It sounds stiff, or over-eager, or four sentences long for a two-word question. This is a coverage gap. Look at what kind of message it was and check whether your reference contains any of that kind. Usually it does not — refusals and rushed replies are the two that are almost always missing.
Wrong message entirely. The reply answers something nobody asked. This is a trigger problem, not a voice problem. Check the keywords, and check the negative keywords, which is where not interested should live. Matching ignores case and punctuation, so Price? and price are already the same thing and the mismatch is somewhere else.
Most "the AI got it wrong" reports are routing failures wearing a voice costume.
What not to do
Do not fix it by writing instructions. Adding "be more casual" somewhere does not undo a reference built from formal replies. It layers an adjective on top of evidence, which is the tone-dropdown problem you were avoiding in the first place.
Do not expand what gets automated to justify the setup. The messages worth handing over are the repetitive ones. Which questions belong to a bot at all is the prior question and it does not have a generous answer.
Do not judge it from the drafts. Judge it from whether you edit them. Send three unchanged in a row and it has your register. Rewrite the second sentence every time and something is missing from the input.
For what the record can and cannot hold in the first place, the deep dive covers it. For the edge cases — low confidence, the review queue, when generation should refuse outright — that is the advanced post.
