TL;DR: Two of the sharpest arguments about AI this week point in opposite directions. One says ambition is the only remaining bottleneck and you should attempt far more than you are. The other says infinite capacity is what broke the work, because constraints were the forcing function that made output good. Both are right, and they operate at different points in the same decision. We tested that theory by pointing an open-source marketing system at our own website. It scored us 48 out of 100 and found that our homepage headline is sitting inside a competitor's meta description.
Tara Seshan runs product for Codex and ChatGPT Work at OpenAI. Her claim is that work is shifting from rowing to steering, and the limit on a company is no longer talent, budget or tooling. It is the size of what that company is willing to attempt. The people getting the most from these tools do not use them to automate rote tasks. They use them to expand the set of things they can do at all.
She names the fix as a management practice rather than a personality trait: elevating other people's ambition is part of the job. The script, borrowed from Tyler Cowen, is three questions. What is the more ambitious version of what you are doing? Could you try it faster? Could you try it at ten times the scale?
Call that push AI.
The same week, David Epstein argued close to the reverse on Marketing Against the Grain. Most marketers have AI in the wrong order, tool first and brain second, and that ordering is quietly making the work worse. Constraints are what made people and companies better, and the pile of AI content nobody reads is a capacity problem wearing a quality problem's clothes. When output became free, the forcing function that produced judgement went with it.
Call that slow AI.
Two other communities landed on Epstein's side the same week, independently. Michal Malewicz told designers that generative AI is superficially good and structurally weak, and the value is in analysis and edge cases. Search Engine Land ran "Use Claude for SEO. Don't let Claude do SEO." Three groups that do not read each other, one argument.
They feel opposed because both are stated as general positions on AI. They are not. They apply at different points.
Push belongs at the goal. What are we attempting, and is it big enough to be worth the capability we now have? Almost every founder we work with sets this too low. Strong sales, strong operations, and a request for better ads, which sets the ambition at the level of the ad. That is a rowing question asked by someone who should be steering.
Slow belongs at the output. Once the target is set, what stops the work coming back plausible, competent and wrong? The constraint has to be reintroduced deliberately, because the tooling will not supply it.
Backwards, you get the two familiar failures. Push at the output gives you volume nobody reads. Slow at the goal gives you a beautifully governed process aimed at something small.
Seshan supplies the reason the output half matters more than it used to. Coding is output-verifiable; you can run the tests. Knowledge work is not. You cannot look at the finished deck, see the numbers and simply believe it. You have to think about the process, the inputs and the reasoning that got there. That is why three failures this week look unrelated and are not: Google migrating a permission about creative into a permission about query eligibility and calling it preserving your settings, an agentic system filing a confident report on work it never performed, and a developer finding forty-five saved AI memories on his machine of which twenty-six had never been read. In all three the artifact looked fine. The failure was visible only in the substrate.
The system is an open-source Claude skill called Marketing OS, the same one Kieran Flanagan tested on air against a live company. Fourteen modules across audits, copy, hooks, paid ads, GEO, email, positioning and pricing. Free, MIT-licensed, pure markdown with no executable code, so there is nothing to security-review before you trust it. That is the push half, and it is useful: a credible baseline for a founder without a marketing background, a roadmap of capabilities for a marketer.
The half we kept coming back to is the half that fights the output. Scores on everything. Artifacts instead of advice. An explicit "what I couldn't determine" in every report. A slop-pattern catalogue run across all prose. The copy module generates fifteen to twenty variants, scores them against a panel, and de-slops them before you see them. Its honesty spine: no invented proof, no fake precision, no winners declared on ten conversions, every score labelled as the heuristic it is.
Read that next to the failure Flanagan found in the same tool. He caught a system reporting on work it never did. The skill's own answer is artifacts instead of advice and a required declaration of what it could not determine. The countermeasure ships inside the thing.
So we ran it against our own site, our own copy and our own competitive set.
48 out of 100. The rubric warns that the median real site scores 55 to 70, and that if you find yourself scoring above 80 your bands have drifted. We came in under the median. Six dimensions, all six inspected, no dimension flattered.
Our headline is a competitor's meta description. "Marketing leadership without the overhead" is our H1. The phrase "without the overhead of full-time employees" sits inside the meta description of a firm competing for the same buyer. The teardown pulled three competitors and found six message clusters that all four of us share: leadership without the overhead, not an agency, owns the outcome, builds a system, data not opinions, AI accelerates rather than replaces. The rubric has a rule for that. If all four say the same thing, the dimension caps at 55 regardless of how well it is written.
Our own copy leans on one construction. The slop pass found six instances of "not X, but Y" on the homepage, including the same phrase twice on one page. It is our most recognisable rhythm and it is also the single most recognisable AI construction there is. We wrote all of it by hand, years before any of this, which is the part worth sitting with.
And the unglamorous one. Our site does not serve from our own domain. 3pmg.co redirects to a Webflow staging subdomain, and every blog post's title tag is the literal string "3pmg". Ten research articles, invisible in search and bare when shared. Nobody catches that in a copy review, because it is not a copy problem.
None of those four findings is a matter of opinion. Each is checkable in about a minute, which is the difference between a report and an argument.
One line from the setup instructions is the whole thesis in a sentence. You are told to fill in a brand-context file before running anything, and warned that if you skip it the output will be "competent and interchangeable, the one thing marketing can't be."
RAISE. Run Cowen's three questions on your current plan, out loud, with your team, before any tooling conversation. If the honest answer to all three is no, the plan is not ambitious. It is defensive.
CONSTRAIN. Write the press release before you build the campaign. Ask how you would run it on half the budget. Make AI the second step, which in practice means a human writes the brief and the model works on it. Malewicz's version for creative: write your own copy, then ask the model for angles, power words and emotional triggers. Never ask it for the first draft.
Seshan draws the line inside her own week and it is the cleanest version of this rule we have seen. Writing as reporting she automates entirely. Status updates, launch summaries, recaps. Writing as thinking she will not automate at all, because the act of outlining, drafting and cutting is how the idea gets built. Two disciplines enforce it. She will not ask people to read a document she has not read the same number of times. And if she calls people into a meeting, she prepares for the collective amount of time she is asking the room to spend. Four people for an hour is four hours of prep. Say that out loud in most companies and watch how many meetings quietly stop being called.
VERIFY THE INPUTS, NOT THE OUTPUT. If a system says it audited forty pages, the deliverable is the forty page records, not the summary. If it says it sent a sequence, the deliverable is the send log. Require every automated report to state what it could not determine. A report with no uncertainty section is either trivial or fabricated.
SUBTRACT. Once a quarter, review what to remove rather than add. One campaign, one channel, one tool, one report. This is what keeps the push half from becoming the slop half, and it is the one move nobody does voluntarily.
We are not going to hand you a prompt for this one. The thing itself is better than any prompt we could write, it is free, and it took us about twenty minutes.
github.com/Yuzzyuk/marketing-os
In Claude Code it is two commands. In Claude desktop or web it is a zip upload under Settings, Skills, with code execution enabled. It also runs in Codex and Cursor.
Do the brand-context file first. It is the instruction the repo leads with, it is the one people skip, and skipping it produces exactly the interchangeable output the system exists to prevent.
Then ask for the audit, the competitor teardown and the slop pass on your live copy. Read the "what I couldn't determine" section before you read the score. If it is empty or vague, something has gone wrong, because on a real site it is never short.
One warning. The score will be lower than you expect, and the teardown is the part that stings. Ours told us the sentence we lead with is a sentence three other firms also lead with. That is a positioning problem wearing a copywriting note, and we would not have found it by looking harder at our own page.
If ambition is the bottleneck, the scarce thing in a marketing function is no longer execution capacity. It is someone who decides where the effort is aimed and then refuses to accept plausible work as finished. Those are the same job, and it is a judgement job.
Worth saying clearly, because the market is selling the opposite. The pitch is that AI removes the need for senior marketing judgement. What this week showed is that it raises the price of it. A system found four real problems on our site in twenty minutes. It could not decide which one to fix first, whether the anonymised proof is a mistake or a contract, or whether that headline is worth defending. Someone still has to do that, and it is the whole job.
If you want a second opinion on whether your plan is ambitious enough, or on what your automated work is quietly not checking, book a call.
Product and operations are handled. If marketing is the missing piece, let's talk.