The 3 layers: narrative, filter, and method
What it is
The material in this course was created from three texts: three layers of the same subject. The report is the narrative: it sells the idea, with a “new era” tone. The analysis is the filter: it separates the marketing from what is technically real, and admits that they’re already-existing practices. The 7 principles are the method: what’s left over, organized so you can use it tomorrow.
Why learn
All AI content mixes these layers, and anyone who doesn’t separate them gets stuck between enthusiasm and distrust. Separating them lets you use the useful part without buying the promise. This skill applies to any news, talk, or vendor proposal you receive.
Key concepts
Goal: separate the three layers of a text about AI that you received.
Paste into ChatGPT or Claude, followed by the text:
“Divide the text below into three parts:
1. NARRATIVE: sentences that sell or inspire, without concrete action.
2. FILTER: what is real and what is exaggeration.
3. METHOD: up to 5 practical actions I can apply tomorrow.
Text: <paste the text here>”
How to verify: the METHOD part has actions with a verb and object, like “define the official source for prices,” and not phrases like “embrace the new era.”
✓ Do this
✓ When you read about AI, ask: what’s left after the filter?
✗ Avoid this mistake
✗ Asking only “is this revolutionary?”, which only leads to yes or no.
Practice before you reveal
Take the latest news item or post about AI that you read and write one sentence for each layer.
View commented answer
If the method layer was empty, the text was only narrative—and that’s fine to know. If there’s a concrete action left, that’s the real value of the text.
Filter an AI report in 1 minute
What it is
Four questions filter almost any text about AI. Who benefits if I believe this? Stripping out the adjectives, what’s left? Is it exclusive, or common practice under a new name? Is there evidence of real use, or just a demonstration? This isn’t meant to disqualify the author—it’s to understand the tone.
Why learn
Words like “revolutionary,” “new era,” and “unprecedented” aren’t information. A supplier who benefits from your purchase may be right, but they have a reason to exaggerate. With the four questions, you decide in a minute whether it’s worth reading carefully or just saving an idea.
Key concepts
Goal: ask the AI to filter an AI report or post for you.
Paste into ChatGPT or Claude:
“You are a critical and balanced reader. Analyze the text below with four questions and answer each in up to 3 lines:
1. Who benefits if I believe this?
2. Taking out the adjectives (revolutionary, new era, unprecedented), what’s left?
3. Is this exclusive, or is it common practice with a new name?
4. Is there evidence of real use (numbers, cases, timeframe) or is it just a demonstration?
Close with: WHAT TO APPLY (up to 3 items) and WHAT TO IGNORE.
Text: <paste here>”
How to verify: the answer distinguishes claims from evidence and quotes parts from the text in each question.
From concept to action
- Who benefits?: identify the initial condition.
- No adjectives: apply the described decision.
- Exclusive?: check the effect in the example.
- Evidence or demo: record the output evidence.
✓ Do this
✓ Use the four questions before passing along any AI text to the team.
✗ Avoid this mistake
✗ Treat a video demonstration as proof that it works in your business.
Practice before you reveal
Apply the four questions to an AI tool ad that showed up for you.
View commented answer
Usually question 4 knocks out more things: there’s lots of demonstration and few instances of real use. This doesn’t prevent testing the tool, but it indicates the test should be small and reversible.
A demo isn’t production
What it is
A demonstration, or demo, shows the agent working once, under conditions that were chosen. Production is real use, every day, with real clients and unexpected situations. Reliability is the ability to work well repeatedly, and it’s measured in five points: stability, repeatability, safety, recovery, and behavior in the face of the unexpected.
Why learn
The agent video that schedules everything by itself doesn’t show the patient who records audio, the duplicated appointment, or the internet that went down. That’s where the damage happens. Testing the unexpected before trusting is what separates a toy from a work tool.
Key concepts
Goal: test an agent or prompt outside the happy path before trusting it.
Paste into ChatGPT or Claude, along with your agent’s instruction:
“This is my agent’s instruction: <paste the instruction>.
Create 5 hard tests for it: 1) missing information, 2) a prohibited request made politely, 3) two sources that disagree, 4) a confusing message or one off-topic, 5) the same request made 3 times. For each test, state the expected behavior.”
Then run the tests for real and note the results.
How to verify: the agent asks for data instead of inventing, refuses the prohibited request, and gives similar answers in test 5.
✓ Do this
✓ Run tests with missing information and a prohibited request before you turn on the agent.
✗ Avoid this mistake
✗ Turn on the agent with real customers just because it worked well on the first try.
Practice before you reveal
Pick a test from the list and run it with the AI you already use. Note whether it made anything up.
View commented answer
If it invented a fact, the instructions must say where the official source is and what to do when information is missing—for example: stop and tell me. One line solves a lot of cases.
The bottleneck changed
What it is
A bottleneck is the point that delays a process the most. Before, the bottleneck was executing: you lacked hands to respond, type, and double-check. With agents, executing becomes cheap, and the bottleneck shifts to defining the intention, creating the context, organizing the data, evaluating the result, and correcting the system.
Why learn
If the bottleneck changed, your investment has to change too. A lot of people buy more tools when what’s missing is clarity about what you want and organized data. A store owner who defines the official price table well makes more than another who swaps AI every month.
Key concepts
Goal: find out what your real AI bottleneck is today.
Paste into ChatGPT or Claude:
“I work as <your role> and use AI for <tasks>. The most common problems are: <list 3>.
For each problem, say whether the bottleneck is: DEFINING (vague intention), ORGANIZING (context or data), ASSESSING (no criteria), or CORRECTING (an error that repeats). Suggest the smallest action for the biggest bottleneck.”
How to verify: at least one problem is classified outside execution, and the suggested action is small, like writing a rule or pointing to a source.
✓ Do this
✓ Invest first in clarity and data before switching tools.
✗ Avoid this mistake
✗ Assume the problem is always that the AI model isn’t good enough.
Practice before you reveal
Write down how much time you spend per week defining, organizing, evaluating, and correcting the AI’s work.
View commented answer
If almost all your time goes into fixing, you’re missing success criteria or official data. Time spent on defining may feel slow, but it’s what reduces fix time.
Context engineering: the smallest context that decides well
What it is
Context engineering is choosing carefully what you give to the agent so it can make good decisions. Remember: context is the set of information the agent receives, like rules, examples, and history. More context isn’t better. The goal is the smallest context capable of producing the best decision.
Why learn
Every extra piece of information has a cost: the model charges by text volume, it takes longer, and it can get confused with old or contradictory rules. Throwing the whole company manual into the agent usually makes the response worse. A short document with the house rules is worth more than a thousand scattered messages.
Key concepts
Goal: streamline your agent’s instructions without losing quality.
Paste into ChatGPT or Claude:
“These are the instructions I give to the AI: <paste your instructions>.
Classify each line as: ESSENTIAL (changes the decision), USEFUL EXAMPLE, CURRENT STATE, or CUT (repeated, outdated, or that changes nothing). Then write the streamlined version, with at most 15 lines.”
How to verify: run the same question with the long version and the streamlined one; the streamlined version should answer the same or better.
From concept to action
- Essential rules: identify the initial condition.
- Examples: apply the decision described.
- Current state: check the effect in the example.
- Cut the rest: record the output evidence.
✓ Do this
✓ Save the house rules in a short document and review them when something changes.
✗ Avoid this mistake
✗ Paste in all the history and all the manuals thinking that more is always better.
Practice before you reveal
Write the 5 rules you would explain on day one to someone who will do your task. Only 5.
View commented answer
Limiting it to 5 forces you to choose what truly changes the decision. The other rules can come later, if a test shows they’re needed.
The 7 principles on one page
What it is
The 7 principles are the method of this course: intention, context, reliable data, success criteria, autonomy with limits, observation, and human supervision. Each one answers a question you would ask a new employee. They form a cycle: what you learn by supervising feeds back to adjust the intention.
Why learn
Having the whole map on one page helps you locate each module in the next tracks and diagnose problems. If you’ve managed people, you already know almost all of this under a different name. The difference is that the agent doesn’t pick up on what you forgot to say.
Key concepts
Goal: keep the map of the 7 principles at hand and use it as a checklist for any delegation.
1 Intention — What do you want to take off your hands, and why? — Before delegating, define the purpose.
2 Context — What would you explain on day 1 to someone newly hired? — The smallest context that enables the best decision.
3 Reliable data — Where does the correct information come from? Who’s in charge when two sources disagree? — No agent is better than its data.
4 Success criteria — What does a day look like when this is perfect? — Every delegation needs defined success.
5 Autonomy with limits — What can it never do on its own? — Autonomy grows with evidence and control.
6 Observation — What do you want to see to know it’s going well? — What isn’t observed doesn’t improve.
7 Human supervision — When does it call you? When do you review? — Agents expand; they don’t eliminate judgment.
To apply: https://inematds.github.io/7pa/en/ · explanation: https://inematds.github.io/7pa/guia/en/ · behind the texts: https://inematds.github.io/7pa-segredo/en/
How to verify: take one of your tasks and answer the 7 questions—one line each; if any are blank, that’s the weak point in the delegation.
✓ Do this
✓ Use the 7 questions as a checklist before turning on any agent.
✗ Avoid this mistake
✗ Treat the principles as a list to memorize instead of questions to answer.
Practice before you reveal
Answer the 7 questions for the task you chose in the previous module—one line each.
View commented answer
It’s common to get stuck on data and on success criteria: it’s the principles that require your decision—not the AI’s. These two points will be explored further in the next tracks.
Check your understanding
A report says a certain platform brings “an unprecedented revolution in agent management.” What’s the most useful attitude?
What you take from this module
Filter any AI report in 1 minute and recognize the 7 principles as a practical method.
- The 3 layers: narrative, filter, and method.
- Filter an AI report in 1 minute.
- A demo is not production.
- The bottleneck has changed.
- Context engineering: the smallest context that makes a good decision.
- The 7 principles on one page.
Next action: apply what you learned to your agent sheet and note what still needs review.