Five AI Habits for GTM Operators

Most companies got very little back from their first wave of AI spending, and the usual post-mortems blame the models. I think that's wrong. The models were fine. We handed people a power tool with no manual and acted surprised when the work came out crooked. The fix isn't a better model. It's teaching the people, on purpose, the way every other serious skill gets taught. (I made the longer version of that argument here.)

So here's the part that comes after the argument: what to actually teach.

What kicked it off was a list. My wife doesn't work in tech, has no patience for jargon, and would rather I stop talking about agents at dinner. She's also been quietly using AI every day to get real work done. At some point I asked what actually stuck from all the advice I'd inflicted on her over the months. She sent back five things. No theory, no jargon, just the moves that changed her output.

That list is sharper than most "AI enablement" decks I've seen, because it survived contact with someone who only cares whether the thing works. So here are the five, pointed at the people I work with all day: growth, sales, marketing, RevOps. The operators who live in the CRM, the campaign, and the quarterly number.

If you teach your team one thing this quarter, teach them these.

One job per session

The habit that pays off fastest, and the one almost nobody adopts by default: start a new chat for a new job.

Most people treat their AI like one long-running group text. Prospect research, then a board email, then a campaign brief, then a quick "what's the weather," all in the same thread for three weeks. By message forty the model is wading through everything you've ever said to find the two sentences that matter. The output goes vague and generic. Not because the model got dumber, but because you buried the signal.

A session should be one objective, scoped, with the relevant background handed over on purpose. When the objective changes, you start fresh. This is the same focused-work discipline that makes people good at their jobs. We just forget to apply it here, because the chat window never tells us to stop.

For an operator that means prospect research is its own session. Drafting the sequence is its own session. Pulling apart why last quarter's campaign underperformed is its own session. Don't let them bleed together.

I'm writing a 3-touch outbound sequence to Heads of Growth at
50 to 200 person B2B SaaS companies.
Context: we sell [one line].
The pain I want to hit is [one line].
Draft touch one only: under 90 words, no exclamation points.

Refine, don't restart

When the first output is wrong, most people do one of two things. They accept it, or they delete the whole thing and re-prompt from scratch. Both are mistakes.

The first draft is rarely the deliverable. It's the starting position. The skill is correcting it: telling the model exactly what's off and letting it adjust while keeping everything it already got right. You built up useful context getting to draft one. Throw it away to start clean and you throw that away too.

The trick is being specific. "Make it better" gets you nothing. "The opener is too generic and it's trying too hard to be clever, so cut the metaphor, lead with the number, and write for a skeptical reader" gets you a real second draft.

The closer is good, keep it.
The middle paragraph reads like a brochure.
Rewrite just that paragraph so it sounds like one operator
telling another what actually happened. Same length.

Argue the other side first

This one surprises people. These models are trained to be agreeable. Ask "is this a good email?" and you'll usually hear yes, because agreeing with you is the path of least resistance. That agreeableness feels good and it's quietly useless.

So make it argue against you. Before you ask for help building something, ask it to attack the thing first. Start from the position you don't hold.

This earns its keep in go-to-market, because we fall in love with our own messaging constantly. Told to play the skeptic, the model surfaces the objection your buyer is actually going to raise, the one you've gone nose-blind to.

You're a skeptical Head of Growth who has seen 200 pitches like this.
Give me the three strongest reasons you'd ignore this campaign.
Don't be polite.
Then help me fix the two objections that are fair.

The same move works on a deal ("steelman why this prospect says no"), on a positioning line ("argue that this is forgettable"), and on a forecast ("make the bear case"). You get a sparring partner instead of a hype man.

Ask it to propose, not solve

There's a real difference between "write me the sequence" and "propose three sequence structures and tell me the tradeoff of each."

The first hands you an answer with the reasoning hidden inside it. You take it or you don't, and you can't see why. The second forces the thinking up to the surface where you can judge it, and it keeps you in the chair as the decision-maker. The "propose a ___" pattern turns the model from an oracle into an analyst.

This matters more the higher the stakes. For a throwaway subject line, just ask for ten. For your territory plan, your pricing page, or your exec-update narrative, you want the options and the logic, not one confident answer you'll have to reverse-engineer later.

Propose four homepage headline directions for [product, one line].
For each one, tell me what buyer belief it's betting on
and who it might alienate.
Don't pick a favorite yet.

Give it tools, or you're just talking

The last one is the bridge to where all of this is going.

A model in a chat window with no access to anything is a brilliant intern with no logins. It can advise you about your pipeline, but it can't see your pipeline. It can suggest an outreach plan, but it can't pull the enrichment, check the calendar, or update the CRM. It's an aquarium: impressive, contained, and not the actual ocean.

The leap happens when you connect it to real systems, your data and your tools and your sources of truth. That's the difference between a thing that talks about your work and a thing that does your work. The whole field is moving this way, through connectors and integrations and the protocols that let a model reach into the tools you already use. You don't need to understand the plumbing. You need to know the question to ask: what could this do for me if it could actually see and touch my systems, instead of just hearing me describe them?

Here's read access to my CRM and my enrichment tool.
Pull the 25 accounts that went quiet this quarter after being active,
rank them by how likely they are to re-engage,
and tell me which signal you used for each.

This is also the honest warning. The teams that never learned to manage a context window are the ones who'll hand an agent the keys and not check its work. Tools are where AI gets powerful, and where it gets consequential. The four habits above (scoped sessions, real correction, built-in skepticism, showing its work) aren't only for better emails. They're the exact discipline you need before you let one of these things act on your behalf. Earn it in the chat window first.

How to stay current without drowning

The tactics change with every model release. The principles above mostly don't. So the goal isn't to chase every "10 mind-blowing prompts" post, because half of those are stale the day a new model ships. (OpenAI now tells people to throw out their old prompts and start fresh with each new model. Take the hint.) The goal is a thin, high-quality input diet, tested against your own work.

Here's where I'd point an operator.

Start here (free, an afternoon each):

Generative AI for EveryoneAndrew Ng's no-code foundation, and the clearest one out there. Google AI EssentialsBuilt for non-technical roles like marketing, sales, and ops. Anthropic's interactive prompt-engineering tutorialFree, hands-on, and you learn by doing.

Go to the source, not the hack blogs. When you want to know how to prompt a model well, read the people who built it:

Anthropic's prompt engineering docs OpenAI's prompting guide OpenAI Academy Google's Workspace Gemini Prompt GuideOrganized by job and use case, which is rare and useful.

Read weekly (pick one or two, not all):

Lenny's NewsletterProduct and growth, the default for a reason. Growth UnhingedKyle Poyar on PLG and modern go-to-market. Demand CurveTactical growth-marketing playbooks. Ben's BitesDaily AI news, digestible, won't eat your morning.

How to judge any of it: prefer primary sources over secondhand "ultimate guides," be suspicious of anything promising magic words, and test every piece of advice on a real task you actually have. If it doesn't move your output, it doesn't count, no matter how many people shared it. More than two newsletters and your inbox starts working against you, so pick one for your lane and one for peripheral vision, and let the rest go.

The point

None of this is hard. That's what bugs me about how badly most teams are doing it. Five habits (one job per session, refine don't restart, argue the other side, propose don't solve, give it tools) and you've cleared a bar that most "AI users" never get over.

My wife runs all five without thinking about it now, and her output beats plenty of people who do this for a living. The difference was never raw talent or a fancier model. It was someone taking ten minutes to show her the moves.

That's the whole thesis, made concrete. The model was never the bottleneck. The willingness to teach the human was. So teach them this. Then go build the thing that earns the tools.

If you're standing up AI literacy on your own team and want to compare notes, you can reach us through civic.com.