Look at the map. The last problem standing:
AI won't do YOUR work consistently.
Brilliant on Tuesday. Different on Thursday.
And a colleague you can't rely on isn't a colleague.
Before we fix it, a memory.
At some point in your career, you got the bad handover.
A laptop. A phone. And a "good luck, kid."
Nobody explained the reports. Nobody showed you what good looked like. You figured it out.
But it took months longer than it should have.
Today we do the opposite.
Today you give Bruno the handover you wish you'd gotten.
The Handover Framework
Five steps. And here's the good news: you already know how to run every one of them.
You've done this before — with people.
- Context first. Let him study your world before anything else.
- Show what good looks like. One finished example.
- Hand over the recipe. The inputs, and every detail of how you build it — and how you use it.
- Run it together. Correct the result until it's yours.
- Let him run it alone. The proof every real handover ends with.
Now let's run them, one by one.
Step 1. Context first.
Open Bruno and paste:
"We're starting the handover of my chosen task, the one in your memory file. Before anything else, get your context: re-read your memory file, then study the finished version in `example` and every file in `input`. If any file won't open for you, tell me which one. Then summarize, in plain language, what you understand so far and what you still don't know."
Notice something while he reads.
Yesterday's conversation is gone. Bruno doesn't remember chats — he remembers his memory file and your folders. That's exactly why we built them.
And if you ever want his eyes on one specific file, don't describe it. Drag it straight into the chat. (One detail for VS Code: hold Shift while you drag.)
Remember this rule, it's worth more than every prompt trick ever sold:
The right context beats the perfect prompt. Every time.
Step 2. Show what good looks like.
Think of a chef teaching a new cook to recreate his best dish.
He shows three things: a photo of the finished plate, the ingredients, and the exact sequence.
Step 2 is the photo.
Drag your example into the chat (hold Shift).
Then tell Bruno, in your own words, what makes it good: the structure, the tone, what the reader needs to see first, what must never be missing.
Two or three sentences. Your words.
Step 3. Hand over the recipe.
Now the ingredients and the sequence.
The ingredients are in input. Drag them in.
The sequence is the recipe: every action, in order, exactly as you'd teach it.
Where each number comes from. What gets checked against what. What you round, what you never round. What you do when something's missing. And what this report is FOR — who reads it, and what they do with it.
The more detail you give, the easier it is for Bruno to recreate your dish. Same as the chef.
You're not writing code. You're explaining your Tuesday to a colleague.
Then tell him: "Ask me questions until the recipe is complete."
Answer until his questions run out. His questions will show you what you forgot.
Step 4. Run it together.
Tell him:
"Now produce the task, following the recipe exactly. Write the finished work only into `output`, nowhere else. Then tell me you're done."
Open the result. Compare it to your example.
Something will be off. Good.
Because now comes the most important habit of the whole week:
Don't just fix the mistake. Ask him what wasn't clear.
"The number was 5, not 10" teaches him nothing.
He'll apologize, he'll fix it — and next run, he'll do it again.
Exactly like a person who didn't understand, and pretends they did.
Instead, ask: "What wasn't clear, that made you write 10?"
He tells you. You explain the why. And the why becomes a rule.
It costs you a few extra minutes today. It pays you back on every run after.
And yes, this is micromanagement.
It's also the only situation I know where micromanagement is a good thing.
Run again. Correct again. Until the output is genuinely yours.
Step 5. Let him run it alone.
One last test — the one every real handover ends with.
"Run the task one more time, start to finish, completely on your own. No questions this time. Then tell me what you did and where the result is."
Open it.
If it's right — and it will be — the handover is complete.
He learned your task. Not a task. Yours.
Now put the framework down for a second.
Because what comes next is where I had my first wow moment with AI.
The moment I understood what this actually was.
And the moment I deliver on the promise I made you at the start.
Let's create a skill.
Make it a button.
Paste this, exactly:
"The handover is complete and this output is the standard. Now create a skill from it, in the official Claude Code skill format, so you load it automatically in every future session. Do exactly this. One: pick a short, simple name for the skill based on the task, lowercase, words separated by hyphens. Two: create the file .claude/skills/<that-name>/SKILL.md inside this main folder, creating any missing folders. Three: start the file with the required frontmatter, name: (the skill's name) and description: (one sentence: what this skill produces and when to use it). Four: below that, write the complete procedure in plain language: my recipe, every correction I made and every "why" I gave you, where input and example live, and the rule that finished work goes only into output. Change nothing else in my folders. Five: verify your own work: read the file back, confirm the frontmatter is valid and that this skill will load by itself in a brand-new session, and fix it if not. When everything checks out, tell me three things: the skill's name, the exact words I type to run it next time (the slash command, plus one plain sentence that also works), and the file's location."
He writes it down. He hands you the words.
Write them somewhere you'll see them.
A sticky note on the monitor is allowed. Encouraged, actually.
Because the task that eats your week...
...is now a button.
output, a skill exists with your recipe and corrections inside, Bruno confirmed it will load on its own in a new session, and the run words are on a sticky note in front of you.Congratulations.
Today you did the thing almost nobody does.
You didn't use AI. You onboarded it, corrected its thinking, and delegated a real piece of your job.
The handover you wish you'd gotten: you just gave it.
The button exists and the solo run looked like mine.
It ran, but it's not quite my standard yet.
Stuck.
DM me on LinkedIn: DAY 3 + your color.
Yellow today usually means one more correction round — and that's the system working, not failing.
I read everything.
Tomorrow, my favorite day.
You'll press your new button.
Then I'll hand you one question managers almost never think to ask...
...and your report will come back better than the one you've been sending for years.
See you tomorrow.
Filippo