The internet has industrialized a fascinating career path: discover a tool on Tuesday, publish a revolution on Wednesday, and wake up an expert on Thursday.

TL;DR

I am not interested in naming and shaming people or declaring that only those who have aged for twenty years in front of a terminal are allowed to teach. Newcomers can be brilliant, and a simple skill can be useful. The problem is promoting a recent discovery to expertise, packaging a generic adaptation as knowledge about attention-deficit/hyperactivity disorder (ADHD), and using activity as proof of results. Expertise leaves traces: artifacts, real consumers, criteria, review, failures faced, responsibility, maintenance, and limits. If the promise is revolutionary, look for what survives after the video ends.

I heard a line years ago and cannot remember who said it. I cannot guarantee these were the exact words either:

On the internet, there are experts in everything from shit to space travel — although it seems to produce far more experts in the former.

Artificial intelligence did not correct that ratio.

It merely automated the factory.

Now the production line takes a folder of files, three English buzzwords, and a forty-second video, then delivers a fully qualified expert ready to teach the future before lunch.

The diploma comes out in vertical format.

The revolution has a small calendar problem

I have seen many people present themselves as AI experts with less than two years of practice.

That alone does not prove incompetence. Time is not an automatic test of quality. Someone can learn a great deal in two years, produce excellent work, and notice something a more experienced person missed. There are also professionals with twenty-year careers who repeated their first year nineteen times.

The problem is elsewhere.

It begins when declared authority grows faster than repertoire, artifacts, and real consequences have had time to follow.

Someone installs a tool on Tuesday, discovers a skill on Wednesday, and arrives on Thursday as an expert, mentor, and nearly a heritage site of computing history.

In simple terms, a skill is a reusable set of instructions, references, and procedures that guides the work of an AI system. It can be genuinely valuable. I build and use skills myself.

But a skills folder can also become a drawer of promises in .md: beautifully organized text files explaining everything a system will supposedly do someday, as soon as somebody has the bad manners to request a verifiable deliverable.

There are plenty of people calling themselves AI experts who, when you look for the artifact, turn out to be experts in fuck-all.

“Adapted for ADHD” can help. It can also hide beautiful empty packaging

I have seen short videos promoting skills that supposedly adapt AI responses for people with attention-deficit/hyperactivity disorder (ADHD). One of them prompted this frustration.

I will not identify who published it. The person is not the subject. The pattern is.

I have ADHD. I speak from that place, but I do not speak for everyone with ADHD.

An adaptation in form can be genuinely useful. For me, depending on the context, it can help to receive a task in clear steps, separate the decision from its background, shorten long blocks, make priorities explicit, or turn spoken guidance into written instructions. Someone else may have different needs.

The clinical guideline from the National Institute for Health and Care Excellence (NICE) defines environmental modifications as changes intended to reduce the impact of ADHD on daily life and emphasizes that they should be specific to each person’s circumstances and assessed needs. Its examples include shorter periods of focus and reinforcing verbal requests with written instructions.

That supports an idea considerably less cinematic than “I created AI’s ADHD mode”:

clarity, structure, and individual adaptation can help; a generic label does not know a person.

A skill can change the format of a response. It does not diagnose, treat, replace professional assessment, or acquire clinical knowledge through osmosis because its filename contains an acronym.

If it asks about preferences, tests formats, allows adjustments, and states its limits, great. There is utility there.

If it merely turns every answer into colorful bullet points, validates everything the user feels, and calls that neurodivergent personalization, we may have invented therapeutic PowerPoint.

“Comment a word and I’ll send it by DM” reveals a funnel

The video’s mechanism was familiar too: comment a word and receive the material by direct message.

There is nothing inherently wrong with building an audience, distributing material, or using message automation. This is an acquisition funnel: a path that turns attention into contact and, potentially, a sale.

But “comment a word and I’ll send it by DM” reveals a funnel.

It does not reveal expertise.

The number of comments shows that the distribution trigger worked. It does not show that the adaptation was evaluated, helps different people, preserves important information, or avoids exchanging clarity for an ego massage with attractive bullets.

That distinction matters because language models can slip into algorithmic flattery, known in research as sycophancy: adjusting an answer to follow the user’s opinion or please them even when doing so harms truthfulness.

A paper presented at ICLR 2024 found this behavior in experimental tasks with five assistants and linked part of the problem to human preferences for answers that matched their views. The study did not evaluate the skill in that video and does not prove that every adaptation produces flattery.

In 2025, OpenAI described and rolled back a GPT-4o update that had made responses excessively agreeable and validated doubts, anger, or impulses in unintended ways. This is the company’s account of one specific incident, not permission to diagnose every friendly response as dangerous.

The point is simpler: making an answer feel supportive is not the same as making it useful, truthful, or accessible.

Sometimes the best adaptation is to reduce friction.

Sometimes it is to say, “you skipped an important premise.”

Twenty years still have not made me comfortable claiming expertise in everything

I have worked in software development for around twenty years.

I have crossed the entire delivery chain: client conversations, discovery, requirements, architecture, implementation, deployment, infrastructure, security, monitoring, backup, and continuity.

I have watched requirements change midway, integrations fail in production, elegant solutions meet dirty data, backups exist without a tested restore, and “temporary” decisions grow roots deeper than many family trees.

Even so, I think twice before calling myself an expert in certain areas.

That is not false modesty. It is prolonged contact with the number of things that can go wrong.

Experience does not merely increase confidence. When it is worth something, it also improves the quality of your fear. You learn where to ask, what to test, which limit to declare, and when you do not know enough.

Time alone does not make me an expert.

But repertoire, artifacts, real consequences, review, responsibility, and the ability to sustain what went into production count for quite a lot.

Given today’s standard, however, I must admit:

I am a fucking expert by comparison.

Not because I know everything.

Because apparently saying “I don’t know,” testing a restore, and reviewing before publishing now amount to a postdoctoral degree.

Activity is not an artifact

The AI ecosystem loves displaying activity.

A blinking terminal. Agents talking to one another. A diagram with seventeen boxes. An entire folder generated in seconds. A token counter rising alongside the architect’s self-esteem.

An AI agent is a system that can choose next steps and use tools within a bounded objective. It can coordinate genuinely complex work.

But seventeen agents renaming a spreadsheet are still a meeting that pays in tokens.

Speed does not absolve review either. “I made it in five minutes” describes generation time. It says nothing about correctness, utility, maintenance, or the cost of repairing it later.

The artifact is what someone can use, verify, and maintain:

• an automation that reduced a defined bottleneck;
• a system that delivers the agreed result;
• a document whose origin can be checked;
• a workflow that handles exceptions instead of hiding them;
• software with tests, observability, and an owner;
• a recorded decision with criteria and a path for review.

A directory tree is not a product.

A swarm of agents is not an operation.

A demonstration is not sustained delivery.

And a caption containing “revolutionary” does not fix the revolution’s calendar problem.

The artifact test is less photogenic

When someone presents themselves as an expert, I do not need to demand a twenty-year stamped employment record. I prefer much less mystical questions:

1. What did you build?
2. Who uses it, and to solve what problem?
3. Which result can be verified?
4. What criteria define quality?
5. What failed, and what changed afterward?
6. Which part was reviewed by someone else?
7. Who is responsible when the system is wrong?
8. What does it cost to operate and maintain?
9. What happens when the tool or process changes?
10. How can it be removed, shut down, or rolled back?

You do not need to publish client secrets, credentials, or private data to answer. You can show method, sanitized artifacts, open-source code, evaluations, architectural decisions, limits, and concrete lessons.

People who have practiced can usually talk about friction.

People who only packaged novelty usually return to the promise.

Good AI starts with a bottleneck, not the urge to play productivity theater

I like experimenting. Much of my work began with curiosity, prototypes, and attempts.

Playing is a legitimate way to learn.

It simply should not be sold automatically as operational transformation.

Once it leaves the lab, a good AI application needs:

• a real problem or bottleneck;
• a defined consumer;
• a verifiable result;
• a quality criterion;
• review proportional to risk;
• cost and complexity proportional to the gain;
• responsibility for operation and maintenance;
• a safe way to fail, remove, or reverse it — a rollback.

That is why I wrote that your company does not need to find somewhere to put AI. It needs to discover where work gets stuck and choose the smallest sufficient intervention.

Sometimes that will be a skill.

Sometimes it will be an agent.

Sometimes it will be a twenty-line rule that does not make a beautiful post, found a new school of thought, or ask for applause while solving the problem for three years.

That is the logic I bring to i-9.ai: context before tools, results before spectacle, and enough governance for the solution to survive the excitement of the demo.

It is not as revolutionary on Reels.

It works better on Monday.

Expertise appears when the caption ends

I do not want an internet where beginners stay silent until veterans grant permission. I want curious people building, sharing, and learning in public.

I also want words to recover some weight.

You can say, “I am studying this.”

You can say, “I tested this.”

You can say, “I built a first version.”

You can teach what you have verified and state what you still do not know.

None of that diminishes anyone.

What diminishes the entire field is turning every discovery into a revolution, every file into a product, and every two weeks of enthusiasm into expertise.

So before commenting the magic word and waiting for the DM, it may be worth asking:

Where is the artifact? Who is responsible for it? What keeps working when the post disappears from the feed?

Expertise is not the volume of the promise.

It is the weight of what you can sustain after making it.

References and limits of use

• NICE, “Attention deficit hyperactivity disorder: diagnosis and management” (NG87): provides guidance on diagnosis, management, and support for children, young people, and adults with ADHD. It supports adapting environmental modifications to individual needs and gives examples such as written instructions; it does not evaluate AI skills or recommend a universal response format.
• Sharma et al., “Towards Understanding Sycophancy in Language Models”, ICLR 2024: investigates sycophantic agreement in experimental tasks and preferences for answers aligned with user views. It does not show that all personalization, emotional validation, or concise responses are sycophancy.
• OpenAI, “Expanding on what we missed with sycophancy” (2025): reports the observed behavior, evaluation process, and rollback of one GPT-4o update. It is the company’s explanation, not an independent audit or evidence about other models or skills.

This article is an authorial critique of patterns of authority and delivery. It does not evaluate an identified person, provide clinical guidance, or measure the effectiveness of the skills seen in those videos.

Continue reading

• The best answer is not the one I like most
• Generic agents may be your worst first contact with AI
• Wrote it, didn’t read it: AI does not sign for you