01 Between 2023 and 2025, we programmed AI in Portuguese

For two years, the way to get a good result from AI was to put all the task’s intelligence into one giant request: "first do this, then research that, then compare, then write, then review, don’t forget to...". We called this micro-prompting. In practice, it was programming AI in natural language.

This made sense because AI at the time lost its way without guardrails. You laid out the path, and it followed it. The problem is that the habit stuck, while AI changed. Today, the model figures out much of the how, and the list of steps that used to help now ties the result to your way of doing things, which isn't always the best way.

Cláudia, a real estate agent, had a request saved in eleven lines: "open the property description, list its strengths, find three similar properties in the area, make a table, calculate the price per square meter, write an opening paragraph..." It took half an hour to fill out the request, and the AI delivered exactly that, without taking a single step further. When a client asked something outside the list, the answer wasn’t useful.

02 The prompt doesn’t disappear: it moves up a level

What changes is the question the prompt answers. Before, it answered "how you should do it". Now it responds "what I want to happen". It’s still a prompt. But now it works as guidance, intent, and delegation, not as a sequence of commands.

Compare the two requests below. The first dictates the process. The second describes the destination, says what matters and what must not happen, and lets the AI choose the steps. It’s the same work, with intelligence in different places.

Sérgio, owner of a language school, replaced “research five competing schools, visit their websites, note the prices, build a table, compare their strengths, and write a report” with “analyze our position against schools in the region and tell me where the three best opportunities are; use whatever information is available and don’t recommend anything without evidence.” The second version surfaced an opportunity he had never listed: classes for businesses, which none of the competitors offered.

03 What remains yours, in any era

The better the agent, the less you need to teach it the route and the more you need to explain the destination. But “explain the destination” has six parts, and AI can’t do any of them for you: what the objective is, what matters, what the constraints are, what can't happen, what resources are available e what result will be considered good.

These six parts are the new work of the person making the request. Notice that none of them is technical. They’re decisions made by someone who knows the business. That’s why this course is for you, not for someone who codes.

Renata, an administrative coordinator at a clinic, needed to reduce missed appointments. Her old request said how to send reminders. The new one stated the destination: “I want fewer than 10% missed appointments; don’t bother people who have already confirmed; never mention the reason for the appointment in the message; use the schedule and history; a good result means fewer missed appointments without patient complaints.” Five sentences, and AI suggested three approaches she hadn’t thought of.

04 It’s like hiring someone experienced

When you hire an experienced manager, you don’t say, “open the browser, type in the competitor’s name, click the first result.” You say, “we need to increase enrollments by 30%.” You can speak that way because the person already has expertise: they know the trade, know how to use the tools, and have experience. Your instruction can be abstract precisely because their expertise is extensive.

With AI, we’re moving in the same direction. The agent’s repertoire grows with every generation of models. What changes on your end is that you move from writing the recipe to giving the briefing. And briefing is a different skill: it calls for clarity about the outcome, not the procedure.

Paulo, from the travel agency, realized he worked with AI in the opposite way from how he worked with Denise, his longest-serving employee. To Denise, he says, "this couple wants a relaxing honeymoon, a twelve-thousand budget, and she doesn’t walk much." To the AI, he dictated every day. When he started talking to the AI the way he talks to Denise, the itinerary got better and he saved twenty minutes per client.

01 A good prompt has always had four parts

Prompt engineering from the first era can be summed up in four parts that almost every course taught, including FEP. Role: who the AI should be ("you are an experienced real estate agent"). Context: the situation and the material. Task: what to do, clearly and unambiguously. Format: how the response should look (a table, three paragraphs, a list). With examples in the middle when the result needed to follow a pattern.

These pieces solved real problems: generic answers, the wrong tone, useless formats, AI that made things up because it lacked material. People who learned this didn’t waste time. But notice that two of the pieces describe what you want and two describe how AI should behave. This distinction will matter in the next section.

Cláudia, a real estate agent, created the prompt she still uses today to describe properties in 2024: "you are an award-winning real estate copywriter (role); this is the apartment, with photos and floor plan (context); write the listing, highlighting light and location (task); in three short paragraphs, without empty adjectives (format)." The listings improved right away. It was the first time she felt the AI was working her way.

02 Half the rules were about you; half were crutches for AI

Look again at the four pieces and the tips that came with them. "Be specific," "say what can’t be done," and "show an example of what good looks like" are about your destination. They still apply in any era because they describe what you want. But “think step by step,” “first list, then compare, then write,” and “respond only in JSON” were crutches: existed for an AI that needed guardrails to keep from getting lost.

One crutch isn't a mistake. It's a solution to a problem that existed. The mistake is continuing to rely on the crutch after your leg has healed. Many of the prompts saved since 2024 contain more crutch than destination, which is why today they deliver less than a three-line request.

Sérgio, from the language school, had a prompt for answering parents’ emails: four outcome lines (“warm tone, never promise a spot, always offer a visit”) and twelve crutch lines (“first identify the subject, then classify its urgency, then choose template 1, 2, or 3, then…”). When he cut the twelve, the responses became more natural and, for the first time, the AI suggested scheduling the visit for a specific time.

03 The hidden cost: giant prompts nobody maintained

The first era had a cost that only became visible later. Every good prompt became a document of twenty or thirty lines, saved in a file, copied, and pasted. When company policy changed, no one updated the prompts. When two people used different prompts for the same task, the results diverged. The prompt was both the day's request and the company handbook, and one document can't handle both roles.

This is the problem the following eras solve, each in its own way: the context era removes from the prompt what AI needs know; the era of skills takes out of the prompt what the AI needs do it the same way every time. What remains in the prompt is just the request for that situation. But before we get there, it’s useful to know how to diagnose a first-era prompt.

Renata, from the clinic, discovered that each receptionist had their own version of the appointment confirmation prompt. One of them was still using a version that mentioned the old insurance plan, discontinued five months earlier. No one had made a mistake: the prompt just had no owner. That’s when she realized the problem wasn’t writing better; it was storing it somewhere else.

04 The line test: "does this describe what I want or the AI’s process?"

To audit any prompt from the first era, you don’t need technical skills. You need one question for each line: "does this line describe what I want, or describe the path the AI should follow?". If it describes what you want (tone, limits, materials, delivery format, criteria for what’s good), keep it. If it describes the path (order of steps, “think first,” “first this”), it’s a candidate for removal.

Take your time with the exercise the first time. Then it becomes automatic, and you start writing new requests without the crutches. It’s the same skill that lesson 8 will take to the extreme, using the September 2026 generation of models, under the name engineering by subtraction.

Paulo applied the test to the itinerary prompt: of the nineteen lines, seven described what he wanted (the client’s profile, budget, pace, delivery table, and what to never include), and twelve described the path. He kept the seven. The next itinerary included a suggestion for an overnight train that saved the cost of one night’s stay—something the old version, tied to his steps, had never suggested.

01 Context is raw material, not decoration

In the first era, "provide context" was one of the prompt’s four parts, and it fit in two sentences. The second era realized those two sentences were the most important part—and the least well served. Context is everything the AI needs to know: who the company is, what it sells, who it sells to, its history, brand, competitors, policies, goals, and past decisions.

The difference from an instruction is simple: an instruction says what to do; context says what world to do it in. A perfect request in the wrong world produces a perfect result for another company. That's why the English name for this era, context engineering, took the place of prompt engineering: the lever changed hands.

Sérgio, from the language school, would ask “write a welcome email” and get copy that could work for any school in the country. Once he started attaching a page describing what the school is (small classes, native-speaking teachers, focus on adults who travel for work, no classes for children), that same one-line request started producing emails only his school could send.

02 The AI's workspace has limited space, and it forgets the middle

Every chat AI works from a kind of table: it can see what’s on the table; what isn’t there doesn’t exist for it in that conversation. This table has a size, technically called context window. It has grown a lot, but it still has limits, and two practical effects matter to you.

First: in a long conversation, what you said at the beginning can fall off the table. Second, and less intuitive: when the table is full, AI pays more attention to the beginning and end than to the middle. Researchers called this "lost in the middle." In practice, put the important rule at the beginning or end, never buried in paragraph twelve.

Renata, from the clinic, pasted the entire forty-page service manual into a conversation and asked for a confirmation message. AI ignored the rule against mentioning the doctor’s name, which was on page 23. When she started over with a single page containing the seven rules that mattered, with that rule first, the message came out right. Less paper on the table, better results.

03 Context that helps and context that wears you out

If more context were always better, you could just paste everything. It isn’t. Too much context has two costs: it fills the table and hides what matters. Think of a new employee: on the first day, they need the client folder, not the company archives. What helps is what the AI needs for this responsibility; what gets tiring is what's at hand and pasted in just in case.

A good context sheet for one responsibility fits on one page and answers five questions: who we are (business), what we’re dealing with (product or service), who it’s for (customer), what must never happen (rule), and what we’ve already decided about it (previous decision). Five well-chosen lines beat forty pages pasted in.

Paulo, from the travel agency, put together the reference sheet for the responsibility "honeymoon quotes": who we are (a boutique agency serving couples ages 30 to 45, with destinations limited to places we’ve visited), what we handle (7- to 12-day packages), who we work with (couples with a set budget), what we never do (promise a room upgrade, recommend a hotel we don’t know), previous decision (we stopped selling cruises in 2025 because of complaints). Five lines, and every new conversation starts by pasting this in.

04 Context moves out of the prompt and gets a place of its own

The big change in the second era wasn't writing better context. It was taking it out of the prompt and storing it in a fixed place with an owner. Today, chat tools have that place: in ChatGPT, they're projects and custom instructions; in Claude, they're projects with files and instructions; both have memory across conversations. You enter the profile once, and every request for that responsibility starts with the necessary knowledge.

This solves the problem of a prompt without an owner from the previous lesson. The partnership rule lives in one place; when it changes, you update it there, and all future conversations already know. The day-to-day prompt can be short again because it no longer has to carry the whole world on its back. In the next lesson, the repeated procedure will get a home of its own, too, and it will be called a skill.

Cláudia created a project called "anúncios" in the tool she uses, pasted the five-line profile into the instructions, and attached the real estate agency’s style guide. Now, for each new property, she opens a conversation within the project and writes only "this apartment" with the floor plan attached. Three words, and the listing comes out in the agency’s tone without repeating the profile.

01 A prompt is the request for right now; a skill is how your organization handles this kind of work

Imagine a company. You walk up to an employee and say, "prepare the proposal for the client on Augusta Street." That is the prompt: the guidance for that specific situation. But the company has a manual with the proposal template, discount policy, how to calculate the margin, visual standards, approval steps, and final checklist. That manual is the skill.

The difference comes down to two questions. A prompt answers "what do I want now?". A skill answers "how does our organization usually handle this kind of work?". One changes with every request. The other stays the same for months and applies to every request in that category.

Renata, from the clinic, noticed that every monthly report for the three doctors followed the same outline: missed appointments, rescheduling, average wait time, complaints, comparison with the previous month, and three recommendations. Each month’s request was “August report” and nothing else. Everything else was a repeated manual embedded in the prompt. She separated them: the “monthly report” skill was written once; the prompt became two words.

02 Prompt and skill don’t compete: they add up

When you say, “turn this research into a presentation for the owners of our partner schools,” that’s the prompt. A presentation skill might contain the narrative structure, slide standards, writing rules, visual criteria, review process, and two ready-made templates. AI receives prompt + skill and executes much better than it would if it received only one of the two.

A skill isn’t a long prompt saved somewhere. It’s more like an operations manual: it can bring together instructions, reference files, document templates, and even guidance on which tools to use. And unlike the context from the previous lesson, which says what world the skill says to work how do one class of work well in this world.

Sérgio, from the language school, built the “business proposal” skill: a three-page proposal template, a price table by class, a discount policy (maximum 10%, only for more than eight students), what to always include (a free trial lesson), and a final checklist (correct company name, 15-day validity, signature). Now he writes “proposal for the transport company, twelve employees, beginner level” and gets a finished proposal in the school’s style.

One caution: calling a 30-line prompt pasted in every time a skill is the most common mistake of this era, because they seem like the same thing. If the text mixes today’s request with the usual procedure, you’re back in the first era. The test is simple: remove the specific request and see whether what remains works for the next case without changes. If it does, it’s a skill. If it doesn’t, it’s still a prompt.

03 A one-page skill has five parts

You don’t need anything technical to write a skill. You need a page with five parts. When to use: when this skill applies (and when it doesn't). What it delivers: the final result, described. The steps: how the organization works, in the order that matters. The model: a finished example or an outline of the document. The final checklist: what to check before considering it done.

Notice that the steps came back, and that doesn’t contradict lesson 1. There, the steps were in the request and tied AI to your way of doing things in each case. Here, they describe how your organization handles that class of work; they were carefully considered once and are available for AI to consult when it makes sense. Steps in the request are a crutch; steps in the skill are a procedure.

Paulo, from the agency, wrote the skill "honeymoon quote" on one page. When to use it: a couple, their first trip abroad together, with a set budget. Deliverable: a two-page PDF proposal with two alternative itineraries. Steps: learn the couple’s pace, choose a destination from the list of places we know, plan day by day, calculate with a margin, review with Denise. Template: the June couple’s proposal, anonymized. Checklist: total within budget, no hotel we haven’t visited, valid for seven days.

04 Where the skill lives today, in any tool

Chat tools already have a place for this. In ChatGPT, there are skills as reusable capabilities: you can see your library by typing /skills and create one by asking, for example, "create a skill to turn research into lessons in our school’s format." In Claude, the same concept appears as skills and also within projects, where the page you wrote can be kept as a reference file. In both cases, the AI starts consulting the skill on its own when the request matches the "when to use" section.

And if your tool doesn’t have this feature yet, the skill works the same way as a document: save the page, paste it at the start of the conversation, and follow with a short request. What defines the skill isn’t where it lives; it’s that it was written once, has an owner, and works for every case in that category.

Cláudia brought the "property listing" skill into the tool she uses: a page with when to use it (residential sales, never rentals), deliverable (three paragraphs plus a list of features), steps (highlight light and location, cite the exact area from the deed, never use empty adjectives), example (the listing for the apartment on Rua das Flores), and checklist (area matches, no "must-see," correct neighborhood). Now the request is "this apartment" with the floor plan attached, inside the project that already has the context profile. Context says who she is; skill says how she writes; prompt says which property.

01 Every request comes from an intent, even when you haven’t written it down

Behind every request to AI, there’s a reason. "Write the welcome email" exists because someone wants the new student not to give up in the first week. That reason is the intent, the highest level of the hierarchy. It rarely appears in the prompt, which is why AI delivers the right task for the wrong reason.

The hierarchy has four levels. Level 1, intention: why. Level 2, objective: what. Level 3, prompt: the guidance for this execution. Level 4, skill: how to do this kind of work well and repeatedly. You already know the two lower ones from previous lessons. The two upper ones are the ones most people skip.

Sérgio, from the language school, would ask the AI to “improve the enrollment page.” Unstated intent: I want to increase course conversions, because in August 40 visitors a day came in and only two enrolled. When he wrote that first, the request changed on its own: instead of “improve the page,” it became “find out why people reach the page but don’t enroll, and suggest what to test first.”

02 The four levels, using the same example, from top to bottom

See the whole hierarchy in one case. Intent (why): "I want to increase the course conversion rate". Objective (what): "find out why people reach the page but don't buy, and suggest improvements". Prompt (guidance for this execution): "analyze the page, offer, objections, competitors, and audience behavior; prioritize changes that can be tested quickly". Skill (how to do it well, every time): the “conversion audit” page, with nine steps no one needs to type again: analyze the offer, identify the ideal customer, study competitors, evaluate the headline, check social proof, identify objections, propose experiments, rank by impact, validate the result.

Notice what each level does with the one below it. Intent determines whether the objective is the right one. The objective determines whether the prompt points in the right direction. The prompt determines which skill to use. Anyone who starts with the prompt is making a request before knowing what they want or why.

Renata, from the clinic, applied the four levels to missed appointments. Intent: reduce losses from empty time slots. Objective: missed appointments below 10% by December. Prompt: “suggest three reminder approaches that won’t bother people who have already confirmed and tell me which one to test first.” Skill: “confirmation routine,” already written, with steps for when to send, what never to say, and how to record it. Four sentences, and for the first time the request had a reason and a goal.

03 Five pillars for checking whether your intent is well structured

Writing the intention and the objective is the beginning. For AI to turn them into a reliable result, the INEMA course on Intention Architecture proposes five pillars that work as a checklist: context (what the AI needs to know, lesson 3), objectives (what success is and why, this lesson), rules (limits and what must never happen), memory (what it should remember from one time to the next) and validation (how you'll know it turned out well). If a pillar is empty, the result will fail right there.

You don’t need to fill in all five for every small request. You need to for every responsibility that you will regularly deliver to the AI. It's the difference between asking for an email and providing the school's communication with parents.

Paulo, from the agency, reviewed the responsibility "honeymoon quotes" against the five pillars. Context: the reference sheet from lesson 3, ready. Objectives: proposal approved on the first read in 70% of cases. Rules: the two "never" lines. Memory: empty, which was why the AI kept suggesting a hotel one couple had complained about in July. Validation: he had never defined it. Two empty pillars explained his two problems.

04 The better the agent, the higher-level your instructions need to be

Here, the hierarchy meets the thesis of lesson 1. A weak AI only understands the detailed level 3, which is why the first era lived there. A better AI understands level 2 and infers level 3. Recent generations are starting to work from level 1: you state the intent and criteria, and they suggest the objective, guidance, and even which skill to use. The better the agent, the less you teach it the path and the more you explain the destination—and the destination lives at the higher levels.

This changes what it means to "know how to use AI." It’s no longer about carefully writing a level 3 prompt. It’s about clearly stating why something exists and what needs to change. That’s a business owner’s skill, not an operator’s. And it ages the least because it doesn’t depend on which model is in fashion.

Cláudia tested the four levels with the generation of models from September 2026. She wrote only the intent and the criteria: "I want to turn over the portfolio faster without lowering prices; a good result is selling in under 60 days." The AI returned the objective (prioritize the two properties that had been sitting for more than 90 days), the guidance (review photos and listings, try showings at an alternative time), and asked if it could use the listing skill she already had. Cláudia just said yes.

01 The complete stack: prompt + skills + context + tools + memory

In the previous lessons, the stack grew piece by piece. Now it comes together. The prompt is the brief. The skill is operational knowledge. The context is what it needs to know. The tools are the hands. The memory is accumulated experience. And the agent is the one who brings it all together to get the work done.

Tools, here, isn’t a technical term: it’s what the agent can interact with in the world. Browser, email, calendar, customer system, spreadsheet, files. Each one is a door you choose whether to open. Memory is what it remembers from one conversation to another: what it has tried, decisions, preferences, mistakes, exceptions. Without memory, every day is day one.

Paulo, from the agency, assembled the stack for the responsibility "honeymoon quotes." Prompt: the intent and criteria. Skill: the page from lesson 4. Context: the reference sheet from lesson 3. Tools: read access to the inbox and supplier spreadsheet; no sending. Memory: a record of each proposal accepted or rejected and why. For the first time, the agent remembered that the couple in July didn’t like beach hotels.

02 Power isn’t enough: the agent needs to be controllable

Nobody wants just a powerful agent. They want an agent controllable. That’s why, in addition to capability layers, control layers come in: permissions (which doors it can open on its own), limits (how much it can spend, how many times it can try), human approval (what only happens after you say yes), records (what it did, when, and why, so you can check), guardrails e o success criterion, so you know whether it’s good.

The most important practical rule: start by opening only the read access. Let the agent read emails, check the calendar, look at the spreadsheet, and show you what it would do. Run it that way for a week. Only then grant the first write access, one at a time and with approval. Trust is built in layers, not by decree.

Renata, from the clinic, gave the confirmation agent access to read the schedule and patient records, but not to send messages. For one week, it prepared the messages and she approved them in a batch, five minutes each morning. In the second week, she allowed it to send simple confirmations; rescheduling still went through her. Two months later, the agent sends 90% on its own, and she reviews the 10% it flags as “I’m not sure.”

03 A short sentence can kick off a huge amount of work

Imagine saying, "launch my new course." On its own, that request is impossible to carry out well. But imagine the agent already has five skills ready: research (market, competition, audience), sales copy (title, promise, objections, call to action), campaign (schedule, channels, message sequence), video (script, image, narration) and analysis (metrics, conversion, adjustment). Now the sentence "launch this course for school owners and get 200 enrollments" kicks off a huge amount of work, with each part following the organization's standards.

This is the leap in the fourth era. The size of the request is no longer related to the size of the work. What determines the result is no longer the prompt, but what’s behind it: the skills available, the context that has a home, the doors that are open, and the limits you wrote down. You built this in the five previous lessons.

Sérgio, from the language school, wrote exactly this sentence for the agent with five skills: “launch the business travel English course, targeting companies in the region, goal of 40 enrollments in September.” The agent delivered in stages: research on the six most promising companies, three versions of the copy, a two-week schedule, and a list of what needed his approval before anything was sent. Sérgio made in twenty minutes the decisions that used to take him a week just to get started.

04 Where to go deeper: this course points the way; it doesn’t teach it all over again

Each of the four layers can fill an entire lesson, and they already exist in the INEMA catalog. The course Super-Agents (agi-pratica) teaches you to write the agent's spec, connect tools safely, build the autonomy matrix with four levels, and create a trust sheet measured over a week of logs. The course The Super Agents Have Arrived (agi-ready) covers the trust curve and the new human role. Here, what matters is that you know these layers exist and can name them within your area of responsibility.

A note about real limits: agents make mistakes, and they make them confidently. Even an agent with every layer set up properly will occasionally make a decision you wouldn’t make. Control layers exist to make those mistakes small, visible, and reversible—not to make sure they never happen. Anyone promising an error-free agent is selling something else.

Cláudia named the four layers for "following up with prospects": tools (reading email and the showing spreadsheet; sending, not yet), memory (each prospect’s history, including what’s already been said), permissions (answer questions about scheduling on its own; anything about price only with her approval), evaluation (prospects without a reply for more than 24 hours, checked every Friday). The layer she still doesn’t trust: sending. And she’s right not to trust it yet.

Each of the four has a door that stays closed, and each for good reason. Sérgio doesn’t let AI respond to parents’ complaints because the wrong tone in a complaint is worth more than ten good emails. Renata doesn’t let it reschedule appointments, which affects the calendars of three doctors under a rule that changes every week. Paulo doesn’t let it handle payments or reservations: money and commitments to suppliers go through Denise or him. A closed door isn’t a delay. It’s the decision that lets the others open with confidence.

01 The entire architecture, from human to result

Put the previous six lessons together in one line and you have the complete architecture of this era: Human → Intent → Prompt or objective → Skills → Context and memory → Agent → Tools → Actions → Evaluation and supervision → Result. Each arrow is a handoff. What orchestrate means designing these handoffs intentionally, instead of letting them happen by chance.

Look at the beginning and end of the line. It starts with the human and ends with evaluation and supervision, which are also human. Everything in between can be AI, a tool, or a person. But those two ends don’t change. Orchestration isn’t about automating everything; it’s about deciding where the person fits.

Sérgio, from the language school, drew the line for “class launch”: him (intent: fill the evening class) → objective (40 enrollments in September) → research, writing, and campaign skills → school context → three agents, one per skill → reading tools and, with approval, sending → actions → him, on Friday, checking the number of responses → result. For the first time, he saw his own place in the diagram: at the beginning and on Friday.

02 What becomes more valuable in people, and what becomes less so

When execution shifts to agents, three things become more valuable in the person coordinating them: define the destination (the intent and the criteria for good, lesson 5), judge the result (reviewing what came back and deciding whether it works, and why) and correct the agent (adjusting the skill, context, or permissions when the result is wrong). These three are things AI can't do for you because they require knowing what the business wants.

And the value of what you used to do by hand goes down: research, draft the first version, build the table, send the reminder. It doesn’t disappear, but it’s no longer where your time goes furthest. Accepting that is what costs most for people who built their careers by executing well. And it’s what separates those who grow in this era from those who compete with their own tool.

Renata, from the clinic, tracked how she spent her day after connecting three agents: confirmations, monthly reports, and message triage. Execution, which had taken 70% of her time, fell to 20%. Judging and correcting, which had been almost zero, rose to 40%. She gave the other 40% back to work no one at the clinic had been doing: talking with patients who had complained. The following month’s report showed the first drop in complaints in two years.

03 The two points where a human always decides

In a workflow with multiple agents, the temptation is to put human approval at every handoff, which turns you into a bottleneck. The alternative is to choose colon and defend them. The first is before leaving home: any action that reaches a client, vendor, patient, or money goes through a person. The second is at the end of the cycle: someone reviews the measured result and decides whether the workflow continues, changes, or stops.

Between these two points, agents hand work off to one another without you. One researches and passes the work to the writer; the writer passes it to the reviewer; if it fails the checklist, the reviewer sends it back to the writer. You don’t see these handoffs, and you don’t need to: they’re recorded in the logs in case something goes wrong. The rule is: involve a human where there are external consequences and where there’s a decision about whether to continue.

Cláudia orchestrated "property sourcing": one agent monitors owner-listed properties, another writes the outreach, and a third checks whether each property fits the portfolio profile. The three pass work among themselves. Human checkpoint 1: no message goes to a property owner without her reading it. Human checkpoint 2: every Monday, she checks how many owners replied and decides whether to change the outreach. Two checkpoints, fifteen minutes a day.

04 Orchestrating doesn’t mean automating everything

The mistake of this era is confusing orchestration with full automation. Full automation removes people from the workflow. Orchestration puts people in the two places where they matter most and removes them from the rest. The difference shows when something goes wrong: in an automated workflow, nobody notices until the customer complains; in an orchestrated workflow, human checkpoint 2 catches the number dropping on Friday and fixes it on Monday.

A mini-case to wrap up. A travel agency automated proposal delivery from start to finish, with no human checkpoint. For three weeks, one supplier changed its prices and every proposal went out with a negative margin. No one checked the numbers. By the time they noticed, they had already signed forty unprofitable contracts. With human checkpoint 1 (a proposal goes out only after someone reviews it), the error would have lasted one day. With human checkpoint 2 (check fees and margins on Friday), it would have lasted a week. Without either one, it lasted three weeks.

Paulo read this case and recognized his own plan to "let everything run by itself." He went back to the map and kept Denise reviewing suppliers before any proposal went out, while he checked the average margin every Friday. The agents do the rest. The agency got faster, and no contract went out at a loss.

01 The same task, two generations

The clearest way to understand the shift is to give the same task to a 2024 model and a September 2026 model and compare the difference. Given the request "research the competitors," the old model would return a list and stop. The new one asks or makes assumptions: what for, what counts as done, what to do if it finds something unexpected. It works from the intent, not step.

Three concrete changes, without exaggeration. First: the model infers the path from the destination, so steps dictated by you now compete with better steps it would choose. Second: it sustains long-running work across rounds without losing the thread, if it has somewhere to record things. Third: it recognizes boundaries (what it can’t do) better than it follows scripts (what to do and in what order). None of the three means “AI has become conscious.” All three are behavioral changes you can observe in a ten-minute test.

Cláudia gave both generations the same request from 2024: "put together a survey of the apartments competing with the one on Rua das Flores." The old one returned eight properties in a table. The new one returned the table, pointed out that two had been priced below average for more than 90 days, suggested that this might indicate a problem with the building, and asked if she wanted the survey to follow that line of inquiry. The request was the same. What changed was where the model started.

02 Subtraction engineering: what to remove from the prompt

If the model can figure out the path, the new skill isn’t adding more, it’s engineering by subtraction. You take the 2024 prompt and cut what was a crutch: “think step by step,” “first do X, then Y,” “respond only in format Z” without a delivery requirement, and “don’t forget to” followed by something obvious. What remains are the three things the new model actually uses: destination, boundaries e completion criterion.

Lesson 2 taught the line test (destination or route?). Subtraction is that test taken to the extreme, with one extra question for each line: "if I remove this, would a 2026 model do worse?". For most route lines, the answer is "it would do the same or better." Those go. The prompt shrinks by half and delivers more.

Sérgio audited the parent response prompt by subtraction, the one that had already lost twelve lines in lesson 2. He removed two more of the seven remaining lines: “first identify the subject” (the model does that without being asked) and “answer in no more than four lines” became “the answer goes through an app,” because the reason matters more than the number. It came down to five lines: tone, three boundaries, and the completion criteria. The responses improved again.

03 Guardrails, not steps; and a finish line with evidence

Two practical rules follow from the shift. First: write boundaries, not steps. “Never promise a timeline,” “don’t touch payments,” and “don’t reject a candidate without checking with me” matter more than any playbook, because the new model chooses the best path better than you can and respects the boundaries better than it follows the trail. Second: define the finish line with evidence. “Do a survey” has no end. “It’s ready when there’s a table with the eight competitors, each one’s price per square meter, and a line saying which one is an outlier and why” does.

Put the two together and you get the intent profile that the INEMA Intent Architecture course calls version 2: intent, boundaries, a definition of done with evidence, and where to record progress. Four blocks, no steps. It’s the request format made for the September 2026 generation.

Paulo rewrote the quote request in the new format. Intent: a proposal the couple will approve on the first read. Boundaries: no hotel we haven’t visited, never over budget, no cruises. Done when: two alternative itineraries in a two-page PDF, total within budget, and one sentence saying which one he’d recommend and why. Record: one row per proposal in the spreadsheet, with approval or rejection. No steps. Denise read it and said that’s how she’d explain it to a new person.

04 What hasn’t changed: where the model still makes mistakes confidently

The headline says "AGI has arrived." The word AGI there's no accepted definition, and this course won't resolve that debate. What we can say with confidence is what no changed. The model still makes mistakes confidently: when it doesn't know, it makes things up with the same fluency it uses when it's right. It still doesn't know your business rules until you write them down. It still isn't accountable: if the proposal is wrong, you're the one responsible.

That’s why the control layers from lesson 6 and the two human decision points from lesson 7 didn’t become obsolete after the shift. They became more important, because a model that works alone for longer can make mistakes alone for longer. What the shift changed was how much you need to teach it the path. What it didn’t change was who decides the destination and who is accountable for the result.

Renata deliberately tested the limits: she asked the new model for the monthly report using a number of no-shows that wasn’t in the spreadsheet. It produced a flawless report with the made-up number and no warning. When she added this boundary—“if a data point isn’t in the spreadsheet, write ‘not available’ instead of estimating”—the next report included three “not available” entries and no inventions. The right boundary matters more than the model generation.

01 The signs of each era in everyday work

Each era leaves visible signs in how a person or department works, and those signs are how you place yourself. Era 1 (prompt): there’s a folder of long prompts; every request is rewritten from scratch; the result depends on who typed it. Era 2 (context): there’s a business profile that AI reads before anything else; requests have gotten shorter; no one has to explain the company again. Era 3 (skill): the repeated procedures have their own page; two people ask for the same thing and get the same result.

Era 4 (agent): at least one responsibility runs in cycles without someone having to request each step, with defined gates and a number reviewed every week. Era 5 (orchestration): several responsibilities overlap, agents hand work off to one another, and people are at both decision points. The most reliable signal isn’t what you think you do. It’s what shows up when someone from outside looks at your week.

Cláudia did the exercise honestly. She thought she was in era 3 because she had a listing skill. But she was still using a folder of long prompts for everything that wasn’t a listing, and the context profile existed in only one project. Signs of era 1 for most of the week, eras 2 and 3 in just one responsibility. She placed herself on the boundary between eras 1 and 2, and that changed what she chose to do.

02 The mistakes of skipping a level

The ladder has an order for a reason: each rung depends on the one before it. Anyone who tries to orchestrate without context builds agents that don't know what company they're in. Anyone who connects an agent without a skill gets different results each time. Anyone who writes a skill without cleaning up the prompt keeps a crutch as if it were a procedure. The mistake isn't ambition; it's the sequence.

Three signs someone skipped a step. The agent “works” in the demo and fails in the second week: context or memory was missing. Everyone on the team uses AI differently and the results vary: a skill was missing. The owner spends the day approving small things: they moved to orchestration without defining the two human checkpoints. In every case, the fix is to go down one step, not up two.

Sérgio nearly skipped ahead. Excited about lesson 6, he wanted to connect the three agents for launching a class in one week. The writing agent produced inconsistent versions because the writing skill didn’t exist yet, and the research agent brought back schools from another city because the context sheet didn’t specify “region.” He went back one level, wrote the skill, and completed the sheet in two days. The following week, the agents ran.

03 What remains human in any era

From era 1 to 5, and after the 2026 shift, four things have stayed in the hands of the people doing the work, and they’re worth practicing no matter which step you’re on. State the intent: why this exists and what needs to change. Write the boundary: what must never happen, something no model can guess. Evaluate the result: look at what came back and decide whether it works, using clear criteria. Take responsibility for what was produced: the wrong proposal is yours, not the agent’s.

Notice that none of the four are technical, and all four are skills of someone who knows the business. That’s why the 50-year-old professional who knows what the clinic needs is better prepared for era 5 than the young person who knows how to write prompts. The whole course was about putting that advantage in the right place.

Paulo listed what he does at the agency that no agent could do: he knows when a couple wants adventure but says "relaxing"; he knows the Lisbon supplier runs late in August; he decides when it’s worth sacrificing margin to keep a longtime client. None of these three things was written down anywhere. He wrote all three as boundaries and criteria on the reference sheet. The agent improved, and Paulo understood what his value was.

04 The learning roadmap: one course per step

This course was the thread. Each step has an INEMA course that covers it in depth, and your next step tells you which one to open. Era 1: FEP, for those who want to master the four pieces and classic patterns. Eras 1 and 2, advanced: FEP2, with a context window, reasoning techniques, and meta-prompting. Intent and the five pillars: Intent Architecture in four days, with day 4 devoted to the turning point of 2026. Era 4: Super-Agents, eight lessons in which one of your agents grows into a team. Role change: The Super Agents Have Arrived, about trust and the new human work.

A mini-case to close out the track. The owner of a language school did the self-assessment, placed himself in era 2, chose writing three skills in 30 days as his next step, and opened the Intent Architecture course for the following week. He didn’t activate any agents in the first month. In the second month, with his skills and context sheet ready, he activated one over two afternoons. In the third month, two. In 90 days, he moved up two steps without skipping any, and never used the folder of long prompts again.

Renata completed her own map. Current step: 2, with one responsibility at 3. Next step: write the report skill and the confirmation skill by day 30. Course to open: Intent Architecture, days 2 and 3, about pillars and systems. After 30 days: Super-Agents, lesson 2, for the agent profile. She put the map on the office wall and marked Friday, day 30, on her calendar.