A footpath runs through the grass to the edge of a headland and stops. The sea lies beyond, at first light.

Push chat to its ceiling, then see where it stops.

Where everyone starts, and where most people get stuck. This part pushes chat to its ceiling, so the move to an agent in Part 2 makes sense.

A footpath runs to the edge of the headland, and there it stops.

Tanush Jagdish · Part 1 of 6 · 11 min read

#1Why we start here

AI tools fall into three broad categories, and you climb them in order. First comes chat: the box you type into in ChatGPT, Claude, or Gemini. Second comes the older automation world of N8N and Zapier, where you connect thing A to thing B. Third come agents. As of September 2026, the ChatGPT and Claude apps hold agents too (Work in ChatGPT, Cowork in Claude), so the move from the first category to the third can happen inside one app. This curriculum skips the middle category, which was built for an earlier style of automation, but it cannot skip the first.

Chat teaches a skill that never goes away: saying clearly what you want and knowing when to trust the answer. It matters more as models become more capable. You learn it by pushing chat hard and then running into its limits, and that experience is what makes the move to agents make sense.

Most professionals have not yet moved past this first rung. In February 2026 we scored the AI capability shown in 54 interview transcripts from professionals at strategy firms such as McKinsey, BCG and Bain, at the Big 4, and at venture capital and private equity funds. The median score was 38 out of 100, and only 15 of the 54 transcripts scored above 50. Median scores fell from 55 on everyday chat to 12 on automation and agents. You are not late: for knowledge work, it is still early.

#2The ceiling

Most people remember the first time AI impressed them. A compliance officer fed an entire regulatory manual into ChatGPT and got a gap analysis in seconds. Then the officer ran the same request again.

“You get that wow moment... and then you do the same thing again and all of a sudden the report is formatted different. I’m damned if I can get the thing to do the same thing twice.”

a compliance officer at a financial services firm

Some people compare the letdown to a first trip to Las Vegas: everything looks opulent until you notice it is painted plastic. Many stop here, decide AI is overhyped, and go back to doing the work by hand. What they miss is that this is a ceiling, not the limit of AI. A whole tier of capability sits above the chat window, and most people do not know it exists. For almost everyone, the ceiling shows up as three walls.

Context resets

Every new chat starts from zero or, with memory switched on, from what the app chose to remember. A consultant spent the first 20 minutes of every session re-explaining his project, his standards, what he’d already tried. He was training the same intern from scratch every day.

The copy-paste tax

Chat gives you text, or a file you still carry to where your work lives. Your job requires deliverables. A consultant spent more than an hour on one slide and got what he called “first-month-analyst” quality. You become the bridge between what AI produces and what your work demands.

The invisible ceiling

The one people can feel but can’t name. Across 54 interview transcripts scored in February 2026, median scores fell from 55 on everyday chat to 12 on automation and agents. There is an entire tier of capability most people don’t know exists.

The AI ceiling: three walls that most professionals hit with chat-based AI.

Wall 1 is context resets. Every new chat starts from zero or, with memory switched on, from what the app chose to remember. Memory features keep facts such as your role, your projects, and your preferences, but they do not keep every detail of your work. One consultant preparing to join a strategy firm spent the first 20 minutes of every session re-explaining his project, his standards, and what he had already tried. Each morning the AI had forgotten the structure of his financial model, the feedback loops, and the quality bar. He was training the same intern from scratch every day.

Wall 2 is the copy-paste tax. Chat gives you text, and your job requires deliverables. Chat can also return a file, such as a spreadsheet or a slide deck, but you still carry that file to where your work lives. A strategy consultant spent more than an hour on a single slide and got what he called first-month-analyst quality. Getting an answer was never the problem. You become the bridge, carrying content by hand between what AI produces and what your work demands.

Wall 3 is the invisible ceiling, the one people feel but cannot name. An executive at a boutique consultancy put it this way: there is a universe of AI-enhanced productivity out there, and we are barely scratching the surface. If your slides still take three hours, your models need heavy editing, and your research includes invented facts, the problem is not you. It is the ceiling of chat-based AI.

Hold onto one encouragement while you are frustrated: frustration is a leading indicator. If you can reach a deliverable today by re-prompting many times, a later model will likely do it in one pass. Friction means you are working at the edge of what the tools can do. Do not over-engineer your way around it. Take the bitter lesson of AI research seriously: compute keeps getting cheaper and models keep getting better, so build for that world rather than this one.

#3The mental model: a drunk genius intern

Before any technique, you need the right picture of what you are working with. Most people get this wrong from the first prompt: they treat AI like an oracle that returns the truth when asked. A consultant preparing to join a strategy firm said he had assumed he could give it everything he had and it would figure things out.

It does not. AI is not an oracle, a search engine, or a function that returns the right answer. A better picture: AI is a drunk genius intern, highly capable in some areas and unreliable in others. Your job is not just to use it but to manage it.

The genius

  • Reads 100 pages in seconds
  • Writes in any format: email, code, analysis, slides
  • Never forgets what you told it (this session)
  • Works 24/7 without complaints
  • First drafts in minutes, not hours
  • Runs multiple tasks in parallel

The drunk

  • Confidently fabricates data and citations
  • Forgets details between sessions
  • Needs a brief to know your project in detail
  • Gives generic output unless you’re painfully specific
  • Sounds authoritative even when it’s dead wrong
  • Gets worse, not better, the longer a conversation runs
The drunk genius intern: the genius side and the drunk side of the same model.

Its genius side reads a hundred pages in seconds, writes in any format, keeps track of what you told it earlier in the session, works around the clock, and drafts in minutes. Its drunk side confidently invents data and citations, forgets details between sessions, gives generic output unless you are specific, sounds authoritative when it is wrong, and gets worse, not better, the longer a conversation runs.

You will see both sides in a single afternoon. A data analyst asked Claude to build a toggle for a LookML dashboard. Its code looked right and the syntax was valid, but when she deployed it, the toggle did nothing: it relied on hardcoded column positions that did not exist. When she asked it to fix the bug, it reviewed its own suggestions and declared them correct. She solved the problem herself.

“If it answers my question in the first run, it’s amazing. If I have to do follow-ups, it derails.”

a consultant at a global strategy firm

The practical translation, which I repeat in every session: treat it like a person with access to your computer. Like an analyst you have handed a folder to, it is smart enough to figure things out, but it needs a brief, and you have to check the work. Because it forgets details between sessions, every new chat is like meeting a newborn genius: you have to show it the map of your world again each time. Part 4 solves that problem, but first you have to feel it. Everything after this part answers one question: how do you manage this intern so it stops forgetting, stops being generic, and stops needing you to check every line?

#4The rich prompt

Here is the trap. People think “be detailed” means “use more words.” A consultant gave ChatGPT a paragraph (his background, his upcoming move to a strategy firm, his six weeks of preparation time) and asked for a financial-modeling roadmap. It was eighty words and felt thorough. He got back case-interview prep mixed with beginner Excel tutorials.

“It gave me a lot of generic. The answers were not MBB-quality.”

a consultant preparing to join a top strategy firm

AI was not the problem. His prompt had no role, no constraints, and no format, so the model filled those gaps with the average of the internet. Words without structure produce generic output; structure turns a request into a brief.

1

Role

Who the AI should be

“You are a senior consultant who builds financial models for due diligence.”

2

Context

What it needs to know

“I join a strategy firm in three and a half months. A reference model from an earlier project is attached.”

3

Task

What to do

“Design a six-week practice plan with increasing complexity.”

4

Constraints

The boundaries

“Each model is built end to end, not step by step. This is not case-interview prep.”

5

Format

How to hand it back

“For each week: problem brief, model structure, key formulas, common mistakes.”

Prompt anatomy: the five-part structure behind a good prompt.

A good prompt has five parts. You do not always need all five, but when output is generic, vague, or wrong, one of them is usually missing.

  • Role: who the AI should be. “You are a senior consultant who builds financial models for due diligence.”
  • Context: what it needs to know. “I join a strategy firm in three and a half months. A reference model from an earlier project is attached.”
  • Task: what to do. “Design a six-week practice plan with increasing complexity.”
  • Constraints: the boundaries. “Each model is built end to end, not step by step. This is not case-interview prep.”
  • Format: how to hand it back. “For each week: problem brief, model structure, key formulas, common mistakes.”

Notice the constraint about building each model end to end. That line is the consultant’s own expertise pushing back on generic advice. Without it, the model kept proposing a step-by-step curriculum that does not match how people actually learn modeling. Your domain knowledge lives in the constraints.

Be specific about your motivation, not just your instructions. You can say “do exactly these steps,” or you can say “here is why I want this.” Prefer the second, because it leaves the model less room to do something foolish. Ask for “a OneNote skill” and it may document OneNote best practices; say “I will use it to look up my own notes” and it builds the right thing. Roles work the same way. Saying you are a strategy consultant points the model toward what it learned from strategy-consulting material. Exact phrasing matters less than including the words.

Speaking is the fastest way to give context: talk it through and paste in the transcript. Give it everything you would tell a person; the model can work out what to ignore. Underneath all of this is the one skill that does not go obsolete: if you cannot state clearly what you want, you will not get it. As models get better at delivering what you ask for, asking clearly matters more. Every criticism of the AI is a prompt you have not written yet. Prompt anatomy is a way to diagnose bad output, not five labels to memorize. When the output is generic, ask whether you gave it a role. When it makes a wrong assumption, ask whether you set a constraint. When the format is useless, ask whether you said how to hand it back.

#5Power chat: thinking, model selection, research mode

What separates power chat from basic chat is not clever prompts. It comes down to three management moves. We watched a consultant who was getting useless frameworks turn it around with three changes: always use thinking mode rather than auto, tell it to research before answering, and give specific feedback when it is wrong. His verdict on the new output was that it was correct. His AI did not get smarter; he got better at managing it.

Thinking effort comes first. For anything hard, turn reasoning up to high or to the maximum setting. More reasoning means the model spends more tokens thinking before it answers, and on hard problems, more thinking tends to produce better output. After you watch a few runs, you develop a feel for what real reasoning costs.

Model selection is a skill, not a setting. Default to the strongest model you have access to, especially if your employer is paying. A strong model behaves agentically, like a good engagement manager who works out the purpose of a task and then handles the steps it implies. A weaker, cheaper model is more like a junior coder who needs the exact task spelled out: good at execution, poor at following intent. You are building the ability to tell, after a couple of questions, which kind of model you are working with.

Build your own intuition before you trust anyone’s benchmarks, including the labs’ own. In our experience, models differ by task: one can be clearly better at slides and visual design while two are even on code. Rankings also go stale quickly. Whatever you last read about the best model is probably a month old, and Claude Code and Codex leapfrog each other often, so learn both.

Research mode is the second management move made literal: make the model look things up before it answers. When you see it run four searches in parallel, it is doing what you would do with four browser tabs, pulling full pages into its context and then synthesizing them. For a generic question it is overkill. For a domain question that needs grounding in real sources, it is the difference between plausible and correct.

#6Verification and trust calibration

Verification separates people who get burned from people whose results compound. You already know why from the mental model: the genius side is real, and so is the drunk side. Verification is how you keep the genius and catch the drunk.

Start by building trust cheaply. Before any real work, give the model a throwaway task: explore this folder, look at the meeting notes, and summarize them. Then run a quick check: is this all correct, and did you make anything up? You are not testing the model’s intelligence; you are calibrating your own trust. You are also catching the worst failure early, a wrong first step. If it opened the wrong file on step one, everything downstream is theater, and it is better to learn that from a cheap prompt than from a finished deliverable.

Your most useful habit is to read the tool calls, not just the answer. To judge whether an answer deserves your attention, look at what the model examined first. If it answered a question about your folder without opening the folder, it does not know what it is talking about, and you can skip the answer.

The enforcement spectrum
  1. Prompt: trust the model

    Ask nicely in the user message. The model may ignore it.

  2. Skill: raise the odds

    A strong description auto-loads the fence. Still advisory once loaded.

  3. Reviewer sub-agent: independent check

    Fresh context grades the output. Advisory; you choose to act.

  4. Hook (nudge): inject a reminder

    Text gets stamped into every loop. The model sees it every time.

  5. Hook (gate): block

    Refuses the tool. Cannot proceed. Mandatory.

  6. External validator: detect after

    CI, schema checks, nightly re-audits. Outside the agent loop.

Low-stakes summaryA skill is fine
Writing to the CRMGate it with a hook
The enforcement spectrum: match enforcement to the stakes, from soft prompts to hard external checks.

Calibrate trust to the stakes. Creative writing earns high trust: glance and move on. Financial data in a client model earns low trust: check every number. Start each new capability at low stakes and expand as trust builds. Do not rely on the model’s own report, either. It will happily tell you it ran the checks and 93 tests passed; the real question is whether those were the right checks.

Your domain expertise is your verification advantage. Verifying work in every domain at once is a problem the big labs are still working on, but in your own domain, you already know how to check the work. Your job is to pass that knowledge to the AI. Tell it how you verify, watch it do so the first few times as you would a new hire, and step back as it proves itself. Before you build formal evaluations, run informal ones: give it examples whose answers you already know, and have it report results against them. For more control, open a second chat and have one AI critique the other. They are blunt with each other; just do not trust either one blindly.

Trust compounds, and eventually you stop checking routine work. Early on you check everything; for tasks where the model is consistently accurate, you can stop. Analysts once checked computer-drawn charts against their own calculations, and eventually they stopped, because they trusted the deterministic software underneath. Your level of verification keeps moving up. AI is on the same path, but it is not deterministic yet, so for now, stay close to the work.

#7Hallucinations and the wall you cannot get past here

Much of the fear around the word hallucination comes from models people used a year or two ago. It still happens. In one benchmarking project, the links looked valid and the data looked reasonable, but the facts were invented. Trust collapsed, and the person went back to manual research. AI could have helped, but that version could not be trusted without verification habits the person had not yet learned.

Why does a good model still make things up? Because it is a next-token predictor: it is trained to produce text that looks right, not to follow a process that guarantees it is right. Two practical points follow. First, current frontier models hallucinate far less than the ones most people remember. Second, the risk grows with depth: a fresh chat is unlikely to invent things, but twenty exchanges deep, it might. Long, cluttered conversations are where the drunk side shows up. A fresh chat behaves better than a tired one, which brings us to context.

Chat gives you one rung of context: custom instructions, Projects, and memory. They are useful and worth setting up, but they are the floor of context architecture, and they run out fast. Two rules help while you are still in chat.

  • 1.Keep chats short. Do not continue a chat past one or two tasks; as the context ages, the model makes more mistakes and forgets what it said earlier. Our rule of thumb: returns diminish past roughly 20% of the context window, and real errors appear past 50%. Slower responses are the sign that the context has grown long. When you notice them, start fresh.
  • 2.Use one chat per problem, not per project. By the eighth or ninth question, the model loses track of what the chat is for. Keep each conversation to one problem.

Notice what neither rule fixes: Wall 1, the context reset. You still re-explain much of your project, your standards, and your quality bar, because chat-level context carries only what you put into it by hand and what the app chose to remember. That is an architecture problem, not a prompting problem, and the rest of the curriculum solves it: agents that live on your file system (Part 2), skills that remember your standards (Part 3), and a folder structure that holds your context (Part 4). You have pushed chat to its ceiling. Part 2 moves from the chat box to an agent that can read, write, and run things on a computer of its own.

The research behind this part.

Continue with the next part.

Want to go faster on your real work?

Coaching is the same method with someone beside you. It is by referral or invitation, and we are at capacity. We write back as capacity opens.