Your AI output is mediocre

When growth is often measured at the last click, you’re paying to compete for demand that was created somewhere else.

Reach people in the purchase planning phase before your competitors know these customers even exist.

With high-intent Pinterest signals on Performance TV you can reach audiences earlier where they watch the most.

Get an advertising advantage

Start Your Campaign

I want to start with a sentence that sounds small and is not.

The new generation of AI models does what you typed, not what you meant.

For two years, that was not true. The earlier models guessed. You wrote a lazy, half-formed prompt, and the model filled in the gaps with its best interpretation of what a reasonable person probably wanted. It was forgiving. It covered for you. A vague prompt produced a decent answer because the model was doing a quiet second job in the background, reconstructing your actual intent from your sloppy instructions.

That era is ending. The current models, Claude 4.7 being the clearest example, have been deliberately built to be literal. They follow the instructions you actually wrote. They size the answer to what you specified. They do the scope you named, not the scope you were vaguely gesturing at. Anthropic's own guidance on the latest model says it plainly. Old Claude guessed what you meant. New Claude does what you typed.

This is a genuine improvement. Literal models are more controllable, more predictable, and more honest. But it has a consequence that almost no growth team has absorbed yet. The quality of your AI output is now a direct function of the quality of your prompt. The model has stopped covering for vague instructions. The gap between a team that gets excellent AI output and a team that gets generic slop is no longer a gap in the model. It is a gap in prompting skills.

This edition is about that gap. Why it is now an operational problem worth taking seriously, what good prompting actually looks like, and why the teams that treat prompting as a real skill are quietly pulling ahead of the teams that are still typing the way they did in 2024.

Let me describe what bad prompting actually costs, because most teams have never put a number on it.

A marketer needs a piece of content. They open Claude or ChatGPT and type something like "write a LinkedIn post about our new feature." The model, being literal now, produces exactly what was asked for. A generic LinkedIn post about a feature. No scope, no length guidance, no tone direction, no audience specified, no format named. So the output is generic, because generic was, technically, what was requested.

The marketer looks at the output, feels vaguely unsatisfied, and does one of two things. They either rewrite it themselves, in which case the AI saved them very little time. Or they go back and forth with the model five or six times, adding one missing instruction per round, slowly steering it toward what they actually wanted. That back and forth takes 20 minutes. The task should have taken five.

Multiply this. A growth team of six people, each interacting with AI 15 to 25 times a day, each interaction running two to four rounds longer than it needed to because the first prompt was vague. That is somewhere between 60 and 120 minutes per person per day lost to imprecise prompting. Across the team, that is most of a full-time role's worth of time, every week, evaporating into rounds of clarification that a single well-constructed prompt would have eliminated.

And the time cost is the visible part. The invisible part is the quality cost. Most people do not actually do five rounds of refinement. They do one or two, get something good enough, and ship it. So the real cost of bad prompting is not just slower work. It is a steady stream of mediocre output going out the door, content that is technically fine and genuinely forgettable, because the prompt that produced it was lazy and the model, being literal, faithfully produced laziness.

This is the part that should bother growth leaders. Your team's AI output quality is now capped by your team's prompting skill. If the prompting is sloppy, the output is sloppy, and no amount of better models will fix it, because the better models are precisely the ones that stopped compensating for sloppy prompts.

The good news is that prompting well is not a mysterious art. It is a learnable, teachable, surprisingly mechanical skill. The principles are stable, and there are not many.

Name the scope explicitly. The single biggest failure in prompting is leaving the scope to the model's imagination. A good prompt names the output, the order, and the boundaries. Not "review this contract" but "review this contract, flag the risks, rank them by severity from one to five, suggest one rewrite per risky clause, and return it as a table." The second version leaves nothing to guesswork. The model knows exactly what done looks like.

Define the length. Literal models size their answer according to what they think the task is. Hand a model a long document and say "summarise this," and it will often produce a long summary, because a long input signals a substantial task. If you want a short output, you have to say so explicitly. Name the word count. Name the format. The model is not going to infer brevity from your hopes.

Use positive instructions, not negative ones. This one is counterintuitive and important. Telling a model what not to do works poorly because the literal model fixates on the very thing you named. "Do not use jargon" puts jargon in the foreground. The instruction that works is the positive version. "Write in plain English a sixteen-year-old could read aloud." Describe the thing you want, not the thing you are trying to avoid.

Lead with action verbs. Vague, permission-seeking prompts go nowhere. "Can you help me with this email?" is not an instruction; it is a question, and a literal model treats it as one. The prompt that produces work is a sequence of action verbs, each one shipping a concrete step. "Find the contact. Draft the reply. Keep it under ninety words. Match a confident, casual tone." Every verb produces an output.

Set the tone deliberately. The current models default to a direct, low-warmth register. If you want a specific voice, you cannot hope the model finds it. The reliable method is to paste two or three sentences written in the voice you want, and tell the model to match the rhythm. Show, do not describe.

Be literal, because the model is. This is the principle underneath all the others. Spell out the output, the order, the length, the tone, and the format. If you do not say it, you do not get it. The model is no longer filling in your blanks. The blanks are now your responsibility.

None of this is difficult. But notice that almost nobody on a typical growth team is doing it consistently. Most people are still typing one-line prompts and hoping. The skill exists. It is just not being practised.

Here is the part that separates teams that use AI casually from teams that use AI as genuine operational leverage.

If you find yourself writing roughly the same prompt twice, you should not be writing it a third time. You should be turning it into a reusable asset.

Every modern AI platform now supports some version of this. Saved prompts, custom instructions, projects, or what the newest tools call skills, are saved commands with all the detailed instructions already built in. The idea is simple. You do the hard prompting work once, carefully, getting every instruction right. Then you save it. From that point on, you and your whole team invoke the good prompt with a single command instead of rebuilding it from scratch every time.

This is the difference between a team where one person is good at prompting and a team where everyone produces excellent AI output. The skilled prompter builds the prompts once, turns them into shared skills, and the entire team inherits the quality. The new hire who joined last week, who has never thought about prompting in their life, invokes the same saved skill and gets the same high-quality output as the most experienced person on the team.

Think about what this means operationally. Your content team has a saved prompt for turning a rough draft into an on-brand LinkedIn post, refined over dozens of iterations. Your sales team has a saved prompt for drafting a follow-up email in your exact tone, under your exact length. Your research function has a saved prompt for analysing a competitor with a consistent structure every time. The prompting skill is no longer locked inside the heads of one or two people. It is encoded into shared tools that everyone uses.

This is the move most growth teams have not made. They treat AI as an individual activity, each person prompting on their own, at their own skill level, with their own inconsistent results. The teams pulling ahead treat prompting as infrastructure. The good prompts are built once, saved, shared, and improved over time, exactly the way you would treat any other operational asset.

You might be reading this and thinking it is a productivity tip, not a growth strategy. I want to argue it is genuinely the second.

We have spent several editions of this newsletter on the same underlying theme. The AI Tool Tax edition argued that the teams winning with AI are not the ones with the most tools; they are the ones who use a small stack with real skill. The edition on AI content and consumer trust argued that the brands losing are the ones flooding the world with generic AI output that audiences have learned to detect and discount. The thread connecting all of it is this. The competitive advantage in AI was never access to the model. Everyone has the same models now. The advantage is the skill of the operator.

Prompting is the most concentrated form of that skill. A team that prompts well produces sharper content, faster research, better customer communication, and more reliable output, from the same models their competitors are using. A team that prompts badly produces generic, forgettable output from those same models, and then wonders why AI did not deliver the transformation it was promised.

The model is not the differentiator. The prompt is. And the prompt is a skill, which means it can be deliberately built, measured, and improved, the same way you would build any other capability that affects your output.

The action here is concrete and it is not expensive.

First, audit how your team actually prompts. Ask a few people to show you the last ten prompts they typed. You will almost certainly find one-line, vague, scope-free prompts. That is the gap, and now you can see it.

Second, run one short internal session on the principles above. Scope, length, positive instructions, action verbs, tone, and literalness. An hour is enough to move the whole team from 2024-style prompting to something dramatically better. This single hour is one of the highest-return training investments available to a growth team right now.

Third, start building a shared prompt library. Identify the five or ten prompts your team uses repeatedly. Content formatting, email drafting, research structure, whatever they are. Build each one properly, once, and save it as a shared skill or saved prompt that the whole team can invoke. The hard work is done once. The quality is inherited by everyone, forever.

Fourth, make prompting a named skill on your team, not an invisible assumption. The same way you would expect a marketer to know how to brief a designer or read an analytics dashboard, you should now expect them to know how to instruct an AI model precisely. It is a core operational competency. Treat it like one.

The honest reframe is this. AI did not make the skill obsolete. It moved the skill. The skill used to be in the doing, the writing, the researching, the drafting. Now, a large part of the skill is in the instructing. The model will do the work, faithfully and literally, but only if you tell it precisely what the work is.

The teams that understand this are quietly building a real edge. They are producing better output, faster, from the same models everyone else has, because they treated prompting as a skill worth developing instead of a thing you just type.

The prompt is the job now. Get good at the prompt.

See you at the next edition, Arindam