Absorption / Diagnostic No. 01
Pretraining present
Press Enter and Walk Away
A useful intelligence should begin with the human pattern it already knows, produce the first artifact, and ask only for the local fact reality proves it needs.
You are my chief marketing officer.
Press Enter
The argument is not that artificial intelligence never needs data. The argument is that a human should not have to prepare the intelligence before the intelligence can begin working.
We have quietly accepted the opposite arrangement. Organize your files. Build a knowledge base. Upload your history. Explain your preferences. Define your advantage. Specify the strategy. Supply the examples. Then, if the briefing is sufficiently complete, perhaps the machine will become useful.
This advice sounds responsible because it resembles the way organizations onboard people. But a generative pretrained transformer is not arriving empty. Its defining asset is that it has already learned a broad geometry of human activity: the recurring shapes of a medical practice, a marketing department, a school bureaucracy, a website, a business plan, an admissions process. It does not know every fact. It does not know your private facts. It can still be wrong. But it should not need you to redraw the species-level pattern before it attempts the work.
That changes the opening move. Give the intelligence a role, the access it is permitted to use, and the boundaries it must respect. Then let it produce something. The first artifact is the beginning of the conversation—not the reward for surviving an intake form.
We keep supplying the answer in advance.
The owner tells the AI what he believes the competitive advantage is, what the strategy should be, which images it should use, how the website ought to feel, and how he has always marketed the practice. He thinks he is adding intelligence. Often, he is installing his own assumptions before the model has been allowed to examine the problem.
The loss is easy to miss because the output becomes more familiar. It sounds like the owner. It confirms the positioning he already prefers. It reproduces the categories already present in the company’s documents. The personalization appears to be working precisely when the experiment has become incapable of surprising anyone.
Excessive prompting can therefore reduce capability. The model’s wide learned geometry is compressed into the narrow geometry inside the user’s head. The machine may have encountered thousands of ways a practice creates trust, discovers demand, explains expertise, and moves a patient toward action. Yet before it begins, it is ordered to treat one inherited brochure as the perimeter of the possible.
This is not collaboration. It is premature convergence.
Do not train the intelligence before you test the intelligence.
Expectation makes error visible.
A biological prediction machine does not begin by asking a newborn what kind of heartbeat it prefers. It begins with a deep expectation of regularity. When the actual heartbeat differs from the expected heartbeat, the discrepancy becomes information.
The irregularity is not immediately promoted into personalization. We do not say: this is the child’s chosen rhythm; update the model. We preserve the expectation long enough to recognize that the difference may matter. The prior pattern is what turns variation into a signal.
Pretraining gives an artificial intelligence its own kind of prior pattern. This is not a claim that every statistical regularity is healthy, neutral, or correct. It is a claim about sequence. Before we can know which local context improves the work, we need to see what the general intelligence produces without that context. The first attempt gives us a baseline. The discrepancy gives us a reason to intervene.
Artifact diagnostic / one complete cycle
A ≠ EWhat pretraining has learned about how this kind of activity usually works.
What the intelligence produces after using available reality, tools, and public facts.
The precise place where a private fact, correction, permission, or decision is genuinely required.
The crucial phrase is demonstrated local exception. Local context should enter as a correction to an observed failure, not as a mountain of anticipatory instruction. That is how personalization becomes evidence-based. We are no longer telling the system everything we imagine might matter. We are supplying the one thing the artifact has shown to be missing.
The artifact-first protocol
Name the outcome.
Give the role or objective in ordinary language. Do not smuggle your preferred answer into the assignment.
Grant access and boundaries.
State what the system may inspect or change, what is private, and which actions require human authority.
Demand an artifact.
Let the intelligence research, decide, build, and return with completed work rather than another questionnaire.
Add only proven context.
Compare the artifact with reality. Supply the local fact only where reality reveals an actual mismatch.
Let it discover what can be discovered.
“You are my chief marketing officer.” Press Enter. Walk away.
That sentence does not magically grant the machine a browser, access to the company’s systems, or authority to publish. The point is not incantation. The point is the division of labor that follows. Once appropriate access and boundaries exist, the intelligence should discover the practice, study the public material, examine the market, construct a strategy, and produce the first artifact from what it already knows and what it can independently verify.
Only then should the owner enter as the holder of genuinely local reality. Perhaps the practice will never offer the service the model centered. Perhaps the founder has a contractual restriction no public page reveals. Perhaps a proposed claim is legally or clinically inappropriate. Those facts matter because the artifact has located exactly where they matter.
The difference between access and briefing is the difference between enabling work and pre-solving it. Access allows intelligence to encounter reality. A briefing often replaces reality with the user’s interpretation of it.
Three walk-away tests
TEST / 01The medical practice
Define every audience, identify the advantage, choose the visual system, outline each page, gather the imagery, and write the biography.
Study the practice and market, form a positioning hypothesis, build the site, and return with a finished first version plus the few claims that require confirmation.
TEST / 02The homeschool decision
Research associations, interpret the regulations, organize documents, build a knowledge base, and maintain the workflow that tells the AI what to do.
Research current requirements, select the plausible path, prepare what can be prepared, schedule what can be scheduled, and isolate the one decision only the parent can make.
TEST / 03The website
Supply the sitemap, copy, photographs, design references, biography, market position, calls to action, and a complete explanation of what good means.
Discover the person or business, form the argument, write and design the site, deploy a working version, and ask for the private or subjective correction only after it becomes necessary.
RAG is a supply line, not a mind.
Retrieval-augmented generation has a proper and important place. The model may need to retrieve a private policy, a current law, an exact account balance, a specialized technical record, or a fact too recent or too precise to trust to pretraining. A large external body of material may need to be searched at the moment the work requires it.
But retrieval is not the source of intelligence. It is a way for intelligence to obtain relevant material. Confusing the two produces a strange ritual in which people spend months assembling a private library before asking whether the librarian can think.
The correct question is not, “How much context did we upload?” It is, “Which missing fact prevented the system from completing the work?” That question keeps retrieval subordinate to the artifact. It also makes the knowledge system smaller, cleaner, and more truthful because every addition has a demonstrated reason to exist.
Can the human leave?
Absorption is not measured by how eloquently an AI explains the work back to the person who must still perform it. It is not measured by prompt length, knowledge-base size, or the number of workflow boxes connected to one another. Its evidence is behavioral: the human can walk away.
When the human returns, some meaningful state of the world has changed. The research exists. The strategy has become an artifact. The application has been assembled. The site is live. The unresolved items have been reduced to the specific decisions that actually require a person’s authority, preference, relationship, or accountability.
If “I want to homeschool my children” causes the parent to become the operator of a new administrative system, the bureaucracy has not been absorbed. It has changed interfaces. If “Build me a website” produces a seventeen-question intake form, the creative work has not been absorbed. It has been returned to the user in the language of collaboration.
A capable system will still ask questions. The standard is not silence. The standard is whether each question marks the boundary of what the intelligence could discover or decide on its own. A good question is evidence of completed work reaching a real human edge. A bad question is the machine handing the unopened task back.
This is why the first artifact matters so much. It reveals what the intelligence already knows. It reveals where access is missing. It reveals which assumptions are wrong. And it gives the human something concrete to correct instead of demanding that the human imagine every future error in advance.