What every leader walks away with
Four things you can verify. If one is missing, the program is not finished.
-
Leaders learn how AI behaves, and how to think when they use it. Why it gets things right one time and wrong the next, what context it needs, and what to check before trusting it. All of it goes into a Project in Claude: who they are, the rules they work by, and what "done" means.
-
A recurring job becomes an agent that runs on a schedule. The job that eats their Monday afternoon today, now named, running on their own files and wired into whatever the company has already enabled.
-
Multi-step work gets an agent, built with five steps that repeat: decide, test, design, equip, protect. Is an agent even the right answer? Then test against real examples, write clear instructions, pick the tools, and decide what a person signs off on.
-
Last, the second brain. Everything they know about their area, their team, the decisions behind the current plan, the standard a result has to meet, sitting where the AI reads it before every job. Keeping it current takes twenty minutes a week.
Four weeks: foundations, a scheduled agent, how to build one from scratch, and a second brain.
01
Aprenden cómo piensa la IA y cómo tienen que pensar ellos cuando la usan.
- Why the same prompt gives different answers, and what to do about it.
- A different way of asking: name the result, give context and examples, then check the answer instead of taking the first one.
- Whatever they would explain every single time goes into a Project, once.
Aprenden cómo piensa la IA y cómo tienen que pensar ellos cuando la usan.
- Why the same prompt gives different answers, and what to do about it.
- A different way of asking: name the result, give context and examples, then check the answer instead of taking the first one.
- Whatever they would explain every single time goes into a Project, once.
I understand the tool, and my instructions stick from one Monday to the next.
A short questionnaire on what they use and what for. Everyone brings two real prompts from that week, one that worked and one that flopped.
What is happening under the hood, and how that changes the way you ask. Advanced prompting, applied to the two prompts they brought: context, examples, constraints, format. Then into a Project with system instructions, reworked until the output holds up.
A week of real use. Every correction they make by hand goes back into the instructions.
The same prompt can give two different answers. So you check the output, fix it, and record in the Project what a good one looks like, rather than trusting the first thing you see.
Context, examples, constraints, format. One of their own prompts, before and after, on the same job.
What the Project knows without being told: its role, the rules, the voice. Written once and reused.
Deliverable A Project with system instructions, two of their own prompts rewritten, and a week of real use with the corrections written down.
02
A recurring job becomes an agent that runs on a schedule, over their own files.
- What an agent does that a chat cannot: it opens files, works through several steps, and comes back with the job finished.
- Monday's job gets recorded as an agent with a name, then called by that name.
- It goes on a schedule, and they decide what to connect: email, calendar, folders, whatever the company allows.
A recurring job becomes an agent that runs on a schedule, over their own files.
- What an agent does that a chat cannot: it opens files, works through several steps, and comes back with the job finished.
- Monday's job gets recorded as an agent with a name, then called by that name.
- It goes on a schedule, and they decide what to connect: email, calendar, folders, whatever the company allows.
Monday's job has a name now, and it lands finished without me asking.
Pick the job against four tests: it repeats weekly, "done" can be stated in a sentence, the inputs are safe to share, and the leader owns it. Bring the folder of files it runs on, plus a note of where the inputs live: email, Drive, calendar.
The difference between a chat and an agent, shown on real files: it opens them, works the steps, and files the result where it belongs. Last week's Project becomes a named agent, run live on the folder they brought, then scheduled. Connectors are decided in the room, against what the company has enabled.
The agent fires once on its own, with nobody watching. Then they read what it produced and write down the connector decision.
The steps of that job, with their own criteria baked in, saved as a Skill in Claude. Called by name, never explained twice.
The job fires at a fixed time, say Monday at 7, and the result arrives without anyone opening the tool.
Which systems the agent reads, and which stay closed. When something is blocked, a one-page request to security: what, why, and under what safeguards.
The agent runs again on its own at the time they set, and the result arrives on its own. Anything that depends on a connector is taught against what the company has already enabled. We do not promise access that security has not approved.
Deliverable A named agent, scheduled and run at least once on real files, and a written decision on connectors.
03
The five steps for building an agent that handles multi-step work and gets it right every time.
- First question: does this job need an agent, or will fixed rules do?
- Define a good result with real examples, then try the simplest version that might work.
- Clear instructions, the right tools, and a line between what it does alone and what waits for a person.
The five steps for building an agent that handles multi-step work and gets it right every time.
- First question: does this job need an agent, or will fixed rules do?
- Define a good result with real examples, then try the simplest version that might work.
- Clear instructions, the right tools, and a line between what it does alone and what waits for a person.
Five steps, same order every time: decide, test, design, equip, protect.
Everyone brings the multi-step job that costs them the most hours each week, and five real examples with the right result already worked out.
The five steps, worked through on the job they brought: decide whether this needs an agent or fixed rules, test against their five examples, design the steps, equip it with instructions and tools, and set what a person approves. Built live, then run in front of them.
The whole job goes through the five steps, dated. Handed in: the decision, the examples, the design, the instructions, the limits.
An agent earns its place when the work needs judgment, when the information arrives messy, or when the rules are too tangled to write out. Otherwise fixed automation is cheaper and behaves the same way every time.
Five real examples with the right answer attached. Every version gets tested against them. Without them, nobody can tell whether a change helped.
Every job has one named person answerable for it, and anything risky or irreversible waits for that person.
The five steps come from the published guides by Anthropic (Building effective agents) and OpenAI (A practical guide to building agents), reduced to the decisions a leader actually makes without writing code: whether an agent is worth it, what counts as a good result, how much rope to give it, what to hand it, and what stays with a person.
Deliverable An agent for multi-step work, built through the five steps: the agent-or-fixed-rules call, five test examples, the step design, the instructions and tools, and the approval limits.
04
The second brain: what they know about the business, their people, and how they make calls, kept where the AI reads it before every job.
- Into one folder goes what currently lives in their head, in their inbox and in scattered notes: the team, the decisions and the reasoning behind them, the standard a result has to meet.
- Before a difficult conversation or a decision, they ask it, and the answer comes back with their context in it: what did I promise this person, what did we settle on last quarter.
- Keeping it alive: the agents from weeks 2 and 3 log what they did, and the leader adds one lesson a week.
The second brain: what they know about the business, their people, and how they make calls, kept where the AI reads it before every job.
- Into one folder goes what currently lives in their head, in their inbox and in scattered notes: the team, the decisions and the reasoning behind them, the standard a result has to meet.
- Before a difficult conversation or a decision, they ask it, and the answer comes back with their context in it: what did I promise this person, what did we settle on last quarter.
- Keeping it alive: the agents from weeks 2 and 3 log what they did, and the leader adds one lesson a week.
Everything I know about my area lives in one place, the AI reads it before it works, and it gets better every week.
The three weeks come together. Everyone lists what they currently re-explain to the AI on every single run: the team, the decisions, the promises, the standard. Notes from the last three one-to-ones come too.
Six or seven documents about their area, written in the room, that the AI reads before every job. Three real questions put to it, with the same questions put to a blank chat for comparison. Then the twenty-minute weekly habit that keeps it current, and the line on what goes in and what never does.
A second dated run of the scheduled agent. Before a real meeting, they put a real question to the second brain, and hand in the answer with the first weekly digest.
The team, the priorities, the decisions, the quality bar, worked examples, and the vocabulary of the area. The AI fills in the rest through use.
What did I promise this person. What did we decide about that. How should I open this conversation. Answers with their own context in them, not generic advice.
The agents log their own work into it, and a digest lands every week. The leader adds one lesson. Twenty minutes, once a week.
This is each leader's own second brain, holding their data under their rules. Sharing one across a team, with permissions and governance of its own, is a later project and does not fit in a single session.
Deliverable A second brain holding six or seven real documents, three questions already answered with their context, the first weekly digest received, and two dated runs to prove it.
How the program runs
Live · 1 h 15 min
Enough theory to act on, then a demo on a real job. The rest is hands on their own work in Claude Cowork. The last 10 to 15 minutes are questions.
Async work · 1 h 15 min
Before the first session: the pre-assessment, a job picked against the filter, and the folder of files that job uses today. Between sessions, dated runs on that same job. Real output, not exercises.
Before and after, on real jobs from their own area
No generic case studies. Every leader arrives with one of these, or with their own.
Finance
Three days pulling spreadsheets out of five departments→the consolidated file already sitting there on Monday
Hunting by hand for what moved and why→the variances explained, without asking
Two hours of drafting the night before→15 minutes of edits on a draft that already sounds like them
Credit and risk
"Let me look it up and get back to you"→the answer, with the clause quoted, on the spot
Everyone rules on them by their own judgment→the same criteria applied, whoever picks it up
Asking someone else to build the model, then waiting two days→their own agent running the scenarios on their assumptions, the same day
Their team
Twelve reviews written from scratch every six months→twelve evidence-backed drafts, ready to review
Chasing everyone's progress through chat threads→the report ready before the meeting, every week
14 files in 14 different formats→one comparable dashboard, on a single page
Regulation and governance
Hunting for the document it was buried in→the answer with its source cited and what is left to confirm
Scattered notes nobody opens again→agreements with an owner and a date, written in the meeting itself
A job that is not on this list still fits. The mechanics are the same for anything that comes round every week.
Who this is for
For
Leaders who open Claude every day but still ask it questions instead of handing it work, and who copy and paste between files to assemble something an agent could deliver finished. They want one recurring job, credit, control desk, regulation, staffing, budget, to come out at the standard they demand, run again without being asked, and be something a colleague can pick up. They also want to understand the technology well enough to judge which AI projects to back in their own area.
Not for
Beginners, who are better served by the AI Literacy program. Anyone who needs software wired into company systems, which is the AI Builder Bootcamp. And anyone who wants a talk without opening the tool.
How we run it
Enough to act on, then act. One short idea, one real demo, and the rest of the time on their own work.
Assess
A questionnaire for every leader before we start. Not a test. We ask what they use, how, what for, and which recurring job they would bring. The answers set the pre-work and calibrate the room before session one.
Coach
Every session opens with one short idea and a demo on a real job. The rest is practice on their own work, with us in the room.
Run
Between sessions the Project and the agents run on real work, dated. Each week leaves behind one run they can show.
From asking AI questions to handing it work, with Claude Cowork
Today: one-off questions, every Monday
- Every Monday starts by explaining to the AI who they are and what they want, again.
- Copying and pasting between files to assemble what an agent could already deliver finished.
- Nothing they work out can be handed to anyone. It stays in their chat history.
After: agents with written instructions, context and rules
- Instructions and standards written once, into a Project and a named agent.
- Agents built through five repeatable steps: decide, test, design, equip, protect.
- A second brain that holds what they know about their area and grows with every job the AI runs.
It does not stop at the last session
Certificate
A shareable certificate once the four weeks are done and the proof of recurrence is in.
Opportunity report
A report on what the ten leaders built and what is still running: where it worked, where it did not, and what is worth extending. A set of next steps, not a satisfaction survey.
Platform access
Exercises, materials, their Projects and their agents stay on the AdapttoAI platform after the program ends. The setup does not die in a chat history.