Training Montage: Your AI Assistant

AI is basically a fantastic intern; it just needs direction.

We’re still in the relatively early days of AI adoption, and I often hear users complain that AI-generated content sounds like it was AI generated. You’re right to push back, it’s not X, it’s Y, the murder of the em dash, etc. etc.

You can take my em dashes from my cold, dead hands.

It’s easy to say that’s a skill issue with the prompt, but in practice it’s a lot harder to solve. At the end of the day, the primary use case for AI-generated content is all about generating momentum and scaling output. Unless you’re creating a skills that your preferred harness always references, most AI writing tools make you do the same manual task over and over: pasting your positioning, your tone rules, and your do-not-say list into a prompt, every single time, and hoping the output lands somewhere near your brand’s voice. Then you burn more tokens trying to dial in the output, and even then it can be hit and miss.

In veloquent, that work happens once on the Assistant settings page. You can configure your own brand rules, and even add custom skills to further hone how veloquent’s assistant talks about your brand. After that, every reply and every draft starts from your brief.

Here's how it works.

The business brief

The brief is what the assistant knows about your business on every turn. Who you are, in one line. What you sell, in your words rather than your website's. Who buys it, and the job they're trying to get done. How you position it, including the claims you never make.

Without a brief, the assistant infers your business from tool names and context clues. That's where wildly off-base drafts come from. In some of the early testing of veloquent, when I asked the assistant to build an onboarding nurture, the content veered sharply away from the target user –marketing operations and marketing teams – and towards IT and development. Obviously not what I wanted. With a consistent brief, the assistant writes from the same source of truth your team does, every time.

Shorter headline: spend time on voice and tone once, not every time.

A writing framework you can actually enforce

Tone in most tools is a dropdown you pick once and forget. Ours is a framework.

You pick tone traits, then sharpen them in your own words. Ours reads "never salesy, always to-the-point, but light, casual, and human-sounding whenever possible." You define voice: who's speaking, to whom, at what distance. Ours is operator to operator, plainspoken, a little contrarian, concrete over abstract. Basically I want what my veloquent assistant writes to sound as much like me as possible (or at least, I think that’s what I sound like…).

Then come the rules. Always rules, like naming the manual task before naming the feature, or limiting every email to one primary call to action. Never rules, like no exclamation points and no sounding like AI. There’s also a banned words list. If a draft contains a banned word, it fails QA. Not a warning. A fail.

Fun fact: this post was initially drafted with that exact framework loaded (and subsquently followed by a light human edit, because no one should “YOLO ship” anything). Zero exclamation points, and you'll notice we never promised anything would be "seamless."

Reference material grounds it in reality

Rules tell the assistant what to avoid. Reference material shows it what good looks like. You add pages from your site, upload markdown or text files, or paste samples directly. The Analyze function reads them and proposes a brief you approve, so you're editing a draft instead of staring at empty fields.

Skills: playbooks that load when the task calls for them

The brief applies to everything. Skills are different. A skill is a procedure with a "use when" line, and the assistant pulls it in only when a request matches. How you write a webinar sequence will probably be a slightly different than how you write a win-back flow or a nurture – how you talk about pricing might be different than how you talk about product. Situational detail belongs in a skill rather than cluttering the brief.

Skills import as plain markdown files, so your existing playbooks are one upload away from being enforced.

To show what that looks like in practice, we built one: an email marketing skill that encapsulates the fundamentals. Subject line discipline, preview text that extends rather than repeats, one CTA per send, mobile-first length. The assistant loads it whenever the task is drafting an email, and ignores it otherwise.

The same request, with and without the skill

We asked the assistant for the same email twice. Once with only the brief loaded, once with the email marketing skill added. This was the prompt used across both emails:

Write an email to our trial users who haven't launched a campaign yet. We want them to try the new agentic campaign builder. Mention that it drafts the whole campaign for them, that QA review is built in, and that there's a live walkthrough webinar next Thursday they can join. Trials expire after 30 days.

Here’s the first sample, without the email best practice skill loaded:

Good, but not great… The subject line and headline are both a little accusatory, and a user might not immediately grok the intent of the email in the half second they see it while flicking through their inbox. It’s completely serviceable, but you probably wouldn’t want to let this one rip without making a few edits.

Here’s what the same prompt generates with some email writing skills loaded:

The email is clearer, more action-oriented, and still fits our desired brand style and tone. The subject line uses the CURVE method and overall the writing is just tighter – in this case, it invites a little curiosity. It’s more on message and on brand. Both versions are a touch wordier than I’d like, but this is a much better draft for iterating.

The skill is what turns an on-voice draft into an on-strategy one: tighter subject line, a single CTA, structure that respects how people actually read email.

No black boxes

The settings page shows you the saved brief exactly as it reaches the model on every turn, plus the skills it can choose to load. You can read precisely what the assistant sees. If a draft misses, you know where to look, and it's editable in plain language rather than buried in a prompt somewhere.

That's the whole idea. Writing rules you'd normally enforce through review cycles and Slack corrections become configuration. Set the brief, add your skills, and the assistant drafts like someone who's been on your team for a year.