🤲 Phase 3 — Hands: the first skill that proves its value

Remember the new employee from module 0.1? They just started and still don’t know the quirks of the place, and you have two options: give them a clear task and see how they do, or hand them the entire list of ten responsibilities right away—organizing the schedule, answering email, handling finances, writing reports, monitoring competitors, etc. Nobody hires that way in real life. It’s the same with a digital assistant: phase 3 of the capstone is for you to give them one skill, see it work from start to finish, and only then think about the next one.

The reason isn’t just caution — it’s technical. Each new skill uses context space (the agent’s “short-term memory” during a session), and each connected MCP (module 2.1) is another component that can fail. Adding too much at once multiplies the failure points, and when something goes wrong, you won’t know which component caused the problem. Adding one skill at a time and validating each before the next is the only way to build a reliable assistant instead of a fragile Frankenstein.

New here? One skill, as you saw in Track 3, is a folder of instructions that Claude Code loads on demand—a “process manual” it reads when a task calls for that specific knowledge. Phase 3 doesn’t call for just any skill: it calls for the first the right skill, and there’s a three-question test to find it in your own routine.

New here? MCP (Model Context Protocol) is the "standard outlet" that connects the agent to external services — instead of each tool needing a different cable, they all use the same protocol. You explored this in depth in Track 2: an MCP server exposes functions (e.g., "create event," "list pending tasks") that Claude Code can call directly, without you having to copy and paste anything manually.

For most first skills, the most useful MCPs are those for calendar (e.g., Google Calendar — read and create events, check for scheduling conflicts) and tasks (e.g., Google Tasks — list to-dos, mark them as complete, create new ones). They turn a skill from "I think and type" into one where "the agent checks what my day actually looks like before acting."

Caption: the scheduler (the phase taught in Track 6) triggers the skill on its own; the skill checks the calendar MCP and the task MCP at the same time; the combined result is ready for you, without you having to ask for anything at the time.

With the skill chosen (topic 2) and the MCPs connected (topic 3), the next step is to ask Claude Code to create the skill file—following exactly the 6-step framework you learned in Track 3. The block below is a complete prompt, ready to paste: it describes the goal, asks for a structured process, and tells the agent which MCPs to use.

Crie uma skill chamada "" na pasta de skills do
meu assistente pessoal. O objetivo dela é: .

A skill deve:
1. Usar o MCP de calendário para listar os eventos de hoje.
2. Usar o MCP de tarefas para listar pendências e as que já venceram.
3. Cruzar as duas listas e montar um resumo curto, em ordem de
   prioridade (o que tem horário marcado vem primeiro).
4. Nunca criar ou apagar nada nos serviços sozinha — só ler e
   resumir, a menos que eu peça explicitamente uma ação.
5. Seguir o formato de skill do curso (SKILL.md com frontmatter,
   descrição clara do gatilho, passo a passo).

Depois de criar, rode a skill uma vez de teste e me mostre o
resultado antes de eu confirmar que está pronta.

A skill that only runs when you remember to ask isn't a digital employee — it's a shortcut. The difference between the two is what Track 6 taught: scheduled tasks (the skill runs automatically at a set time, e.g., every day at 8 a.m.) or loops (the skill runs at intervals, checking something that changes frequently). For a daily routine like the one in topic 4, a simple fixed-time schedule — covered in depth in module 6.2 — is usually enough.

The key point: scheduling the skill closes the loop on phase 3. You’ll no longer open the terminal in the morning and type “run my routine skill” — the schedule does that for you, and the result arrives ready (by notification, file, or the channel you configured in module 6.6). That’s when the assistant stops depending on your memory.

To wrap up phase 3, two pairs of mistakes and fixes sum up everything this module asked of you. The first is about scope—how much to give at once. The second is about scheduling—whether the skill actually runs on its own or still depends on you remembering.