Free · the masterclass kit

AI work you're
proud of.

Seven things that make the output better. Everything from the talk, written up so you don't have to remember it, plus what makes it hold up when the reader is picky. Copy-paste. Any tool. No code.

From Impossible to Replace: The 2026 AI Masterclass
Saturday, September 19, 2026 · 2:30–4:30 pm · HanaHaus Newport Beach, 3366 Via Lido

Start here

Pick two. Not seven.

Everything on this page works in whatever tool your company gives you: Copilot, ChatGPT, Claude, Gemini. None of it needs code. Read the list once. Pick the two that fit the work on your desk on Monday. Use them for a week before you add a third.

They run in the order you'd bring on a new hire: brief it, show it what good looks like, work in rounds, make it show its work, make it argue back, speak its language, write it all down once.

01 Move two · the hat

Brief it like a new hire, or let it write the brief.

It can't see your inbox. Most bad output is a missing brief, not a bad model. The line that matters most is the last one: what "finished" means.

Start today

  1. Four lines before the ask: who's reading (and how long they have) · what happened before · what's fixed (numbers, deadline, length) · which hat it wears: hard worker, problem solver or storyteller.
  2. One more line, "This is finished when…": the format, the length, the one thing that must be right.
  3. No time to write it? Make it interview you. Answer one line each, then say go.

The brief

Be my [hard worker / problem solver / storyteller] on this. Reader: [who, how long they have]. Background: [what happened before]. Fixed: [numbers, deadline, length]. Previous version: [paste]. Task: [the ask]. This is finished when: [format], [length], [the one thing that must be right]. Check each before you show me.

Or let it write the brief

Before you do anything, ask me questions until you understand what I need. One at a time, only the ones that change the answer. Then tell me in three lines what you'll do, and do it.

Thirty seconds of brief saves three rounds of fixing.

02 Move one · your voice

Show it what good looks like. Start with your voice.

It copies what it's shown far better than what it's told. "Professional but warm" means nothing. Five emails you're proud of mean everything.

Start today

  1. Paste five emails or documents you're proud of. Ask for one paragraph on how you write. Cut what isn't you. Save it.
  2. That's your voice paragraph. It goes at the top of every prompt, or into the skill file (07) so it's there without asking.
  3. For anything with a shape (a one-pager, a dealer letter, a deck), attach the best one you've ever sent and say "like this".

Your voice

Here are five pieces of my writing I'm proud of. Describe how I write, as one paragraph I can give to any AI tool: tone, sentence length, how I open and close, how I handle numbers, what I never do. [paste the five]

Or, for a format

Match this example: same structure, length and level of detail. Tell me in two lines what you kept from it, then write. [paste the example]

Mine came out of Copilot in one go. It was right about the "no pleasantries" part.

03 Move three · the loop

Work in rounds. Never take the first answer.

The first answer is the average of the internet. The second is a reply to you. A rewrite throws away what was good; one change per round keeps it.

Start today

  1. Anything longer than an email: outline first. Fix the outline, then ask for the prose. Five bullets are easier to fix than five paragraphs.
  2. Say exactly what's wrong. "Too long" isn't exact. "Cut paragraph two, keep the numbers" is.
  3. One change per round. Stop at the round you'd put your name on, usually the third.

Outline first

Don't write it yet. Give me the structure: the sections, the one point each makes, and the number that proves it. I'll edit that, then you write.

The round

That's a draft. What's wrong: [one thing]. Fix only that. Keep everything else exactly as it is, and tell me what you changed.

Ask. Read it. Say what's wrong. Ask again. The full version, with loops you can steal, is in the Loop Library.

04

Make it show its work.

Given a gap, it fills it with a confident guess. Given a number, it may have invented it. Neither goes into an email until you've seen where it came from.

Start today

  1. Give it the raw file, not your summary of it. Your summary already dropped something.
  2. Two lines in any prompt with facts in it: list your assumptions; if something is missing, ask, don't invent.
  3. Every number gets its source next to it. Spot-check three. If one is wrong, check them all.

Copy & paste

List every assumption you made. Mark anything you're unsure of with [?]. If information is missing, ask me. Do not fill it in. For every number, show the file, the row or cell, and the calculation next to it. If you can't trace one, say so.

05

Make it argue against you before your reader does.

It's a better critic than writer. The hardest questions are cheaper to hear from the machine.

Start today

  1. Before anything important goes out, two extra prompts: the toughest reader, then an editor scoring it against your finish line.
  2. Fix what both flagged. Ignore the rest.

The reader

You are the [CFO / dealer principal / client] about to say no. Your five hardest questions about this. And which one am I least ready for?

The editor

Now review it as a demanding editor. Score it 1 to 10 against the finish line I gave you. What would you cut? What's missing?

06 Move three · the toolkit

Speak Simplified Technical English.

The standard aircraft manuals are written in, so a mechanic anywhere can't misread a safety instruction: short sentences, one instruction each, one meaning per word, no idiom, no hedging. The machine misreads the same things people do.

Where it earns its keep: instructions that must not be misread (your skill file, anything for a global team) and prompts that keep coming back "almost". For a quick email, optional.

Start today

  1. Rewrite your next prompt to the rules: under twenty words a sentence, one instruction per sentence, no "maybe", no "if possible".
  2. When the output goes to a global team, ask for it in the same standard.

Copy & paste

Write in Simplified Technical English: sentences under twenty words, one instruction per sentence, one meaning per word, active voice, no idiom, no hedging. Read my requests the same way.

ASD-STE100: about 900 words and 53 rules, in use in aviation since 1986.

07 Move three · the toolkit

Write it down once. The skill file.

AI is a new hire every single morning. Everything above either lives in a file or gets retyped forever.

Start today

  1. One page: your voice paragraph (02) · your finish-line template (01) · how you write numbers · who gets copied and what stays out · the STE rule (06).
  2. Save it where your tool keeps instructions (Custom Instructions in ChatGPT, a Project in Claude, a Notebook in Copilot), or as a pinned note you paste first.
  3. The new-hire test: could someone use it on day one? If not, it isn't finished.

Copy & paste

Here is my skill file. Follow it for everything in this conversation unless I say otherwise. If a request conflicts with it, tell me before you start. [paste the file]

Two rules that aren't techniques

Before any of it leaves your hands.

Know what you can paste.

If you wouldn't email it outside the company, it doesn't go into a public tool. Strip the names and the figures, or use the version your company pays for. Quality work includes not becoming the cautionary tale.

Add the line only you could write.

The specific name. The thing that happened on Tuesday. The detail no model has. It's what turns "AI wrote this" into "she wrote this, fast."

From the talk

Nobody can guarantee your job. Practice anyway.

Fifty openings. Thirty applications. In your voice, not its voice. That's what practicing buys you, and it starts with two of the seven, on Monday.

Bookmark this page so it's there when Monday comes: irinagrak.com/masterclass.