Flint Group · AI Field Guide

you already use AI every day. now use it well.

Everything from training day, in one place you can come back to. Not theory. The moves, the guardrails, and the prompts you'll open Monday. We don't do inspirational. We do useful.

start here
where you landed

you were already in the top 1%.

We surveyed all 23 of you before training. The headline: you don't need convincing to use AI. You're already daily users. What you needed was to understand what you're actually using, so you can trust it where it earns trust and catch it where it doesn't. That's what this guide is for.

83%
of you use AI several times a day
4.1/5
average confidence in your own AI use
4
different AI tools per person, on average
3.4/5
how clear you felt on what's safe (the gap)
the one thing that changes everything

it's predicting. it isn't searching.

In the survey, 65% of you described AI as "searching the internet." It's a reasonable guess. It's also the single most expensive misunderstanding in the room, so we fix it first.

Tools like Claude and ChatGPT are prediction engines. They learned patterns from an enormous pile of text, and when you ask a question, they predict the most likely next words, one after another. There's no library it walks into and pulls the right book. That's why it's fast, fluent, and sometimes confidently, completely wrong.

Why you care: if you think it's looking things up, you trust it like Google, and you forward its answer to a client without checking. It isn't Google. Treat every answer like a sharp intern's first draft - fast, useful, and in need of a human read before it leaves the building.

From your survey: 96% of you knew the word "hallucination." This is why it happens. It's not a bug someone forgot to fix - it's a side effect of prediction. Assume it can happen, and check anything that matters.
calibrate your trust

great at this. quietly bad at that.

Half of you marked at least one thing AI is quietly terrible at as a strength. No shame - it hides it well. Knowing the line is the whole game.

lean on it for
where it's genuinely strong
  • First drafts you'll rework anyway
  • Summarizing long, boring documents and call notes
  • Options when you're stuck: 20 headlines in 20 seconds
  • Matching a voice, if you show it real examples
  • Explaining something complicated in plain English
double-check it for
where it quietly lies
  • Math you can't verify yourself
  • Citing real sources - it will invent them, cleanly
  • Remembering past chats perfectly (it doesn't)
  • Today's facts without searching - it has a cutoff
  • Telling you it's wrong - it won't flinch
the line to remember

Confident isn't correct. The dangerous answer isn't the obviously crazy one. It's the plausible one you don't think to check.

protect the business

the 30-second safety check.

"What's safe to put in" was the number one thing you asked to learn. Two of you wrote, word for word, "honestly, I'm not sure what's safe here." So here's the whole thing. Before you paste anything in, run these three questions.

1

Would you hand it to a stranger on the street?

If no, strip the sensitive bits or skip AI for this one.

2

Does it have a client's name, data, or IP?

If yes, anonymize first. "Client A," "$X," "[name]." Same output, none of the exposure.

3

Is it approved for the tool you're using?

If unsure, use a paid or company-approved tool only.

Not all tools treat your data the same. This is the part most people skip. Here's what actually happens to what you type:

use with caution

Free tools

Free ChatGPT and friends may use what you type to train future models. Treat it like posting on a public forum.

better - check the policy

Paid plans

ChatGPT Plus, Claude Pro and similar don't train on your inputs (per policy). Still stored, so still check your org's rules.

safest for sensitive work

Company-approved

Enterprise Copilot and approved org tools keep data inside your walls. The default for anything confidential.

the gray zone is real - have a default

Not sure it's okay? Ask your manager or check the AI policy first. Thirty seconds now beats a very long conversation later. When in doubt: don't paste. Rewrite it first.

augment, don't outsource

own the thinking. delegate the typing.

Several of you worried, rightly, about AI flattening the work. The answer isn't to avoid it. It's to draw a line on purpose. The rule of thumb: if a wrong answer costs the brand or the relationship, you own it. If it's volume or format, hand it off.

you keep
the judgment that has your name on it
  • The strategy and positioning
  • The core creative idea
  • The hard client conversations
  • The final "is this on-brand and true?" call
hand to AI
the work that just needs doing
  • First drafts and 20 headline variations
  • Summarizing research decks and call notes
  • Resizing one asset into ten
  • SEO, alt text, and formatting
prompting that works

say it like this.

A weak prompt gets a weak answer. The fix is four parts: tell it who to be, give it the context, make a clear ask, and show an example of good. And if you're staring at a blank box, don't. Use meta-prompting - ask AI to write the prompt for you. Tap any card to copy it.

1. role2. context3. the ask4. an example
the meta-prompt · when you're stuck
I need to [describe your task in plain English].
Write me a detailed, reusable prompt for this. Ask me any questions you need first, then give me the final prompt.
client email in your voice
Act as a calm, confident account lead at a marketing agency.
Write a follow-up email to [name], who was interested but went quiet after [stage].
Tone: warm, low-pressure, not salesy. Under 120 words, end with one clear ask.
Match this voice - here are two emails I actually liked: [paste two].
brief to three directions
Act as a senior strategist. Here is the client brief and our past work: [paste].
Give me three on-strategy campaign directions, each with a one-line hook and why it fits.
Then flag anything missing from the brief I should ask the client before we start.
make a long document make sense
Summarize this for a busy exec who has 60 seconds: [paste].
Give me the one thing that matters, three takeaways, and any risk or red flag I should not miss.
Plain language. No filler.
teach it your voice · do this once per project
Here are 5 things I wrote that sound like me: [paste].
Describe my voice as a few rules - sentence length, tone, words I use and avoid.
From now on, write everything in this project that way unless I say otherwise.
your toolkit

which tool, when.

You don't need all of them. Most days you need two. Here's the job each one is best at, and the one tip that makes it click.

Claude

your deep-work partner

Long documents, nuanced writing in your real voice, careful reasoning. The one for actual work, not party tricks.

Tip: paste in a few things you've written and ask it to learn your voice before you start.

Cowork

claude, working on your files

Give it a task and your actual files. It reads them, builds the report or the deck, and doesn't send anything until you say so.

Tip: one project per part of your work - a client, a service line - so it always has the context.

ChatGPT

the all-rounder

Fast everyday help, images, quick answers. The easiest first stop. Reach for Claude when it's a long doc or your voice matters.

Tip: great for images and a fast second opinion. Not your vault for sensitive client data.

Superwhisper

voice to text, anywhere

Tap a key, talk, and the text lands at your cursor in any app. You speak about three times faster than you type.

Tip: talk your messy thoughts at AI, then let AI clean them up. Fastest combo there is.

Otter.ai

meetings, written down for you

Auto-joins your calls, transcribes, and hands you the recap and action items. No more frantic note-taking on a client call.

Tip: get everyone's consent before you record. In about a dozen states it's the law - put it in the invite and say it out loud.

Lindy

agents & automation

Build AI agents in plain English that handle repetitive work across your apps - drafting, follow-ups, triage. No code.

Tip: start with one low-stakes agent and keep a human on anything that sends, spends, or posts.
agents, safely

let it mess up your day. not your life.

Five of you asked for agentic AI in your workflows, so here's the honest version. An agent doesn't just answer - it acts. It reads, decides, and takes steps across your apps on your behalf. That's the power, and the catch.

The catch: an agent reads your instructions and random text it finds on the web the same way. So a hidden instruction on a page can hijack what it does next - that's prompt injection. You don't fix it by trusting the agent more. You fix it by shrinking what it's allowed to touch.

safe to start
  • Summarize meetings and transcripts
  • Draft replies and copy (you still hit send)
  • Research and competitor scans
  • Sort and tag your inbox
not yet · keep a human on it
  • Emailing clients on its own
  • Spending money or making purchases
  • Posting online as you
  • Touching client files or live systems
the rule

Give it its own limited login, make it ask before anything risky, and start where the worst case is a wasted hour, not a lost client. You stay the boss.

plain english

the words, without the jargon.

You knew the vocabulary in the survey. Here are the definitions you can actually explain to a coworker.

LLM

Large Language Model. The engine behind Claude and ChatGPT. It learned patterns from a huge pile of text and uses them to predict words. There's no database it looks things up in.

Prompt

What you type in. The clearer and more specific, the better the output. Vague question, vague answer. Garbage in, garbage out - it's just polite about it.

Hallucination

When AI states something false with total confidence. Not a switch you can turn off - a side effect of how prediction works. Assume it can happen and check what matters.

Training data

The text it learned from, up to a cutoff date. So it doesn't know last week unless it searches, and it can be wrong or out of date, because the internet is.

Context window

How much it can hold in its head at once - your files, the chat, your instructions. Big, not infinite. Long messy chats make it lose the plot. Start a fresh one.

Agent

AI that doesn't just answer, it does. It takes steps for you across your apps. The most useful, and the one to keep on the shortest leash.

make it stick

your monday.

This is where most training dies - everyone feels inspired and changes nothing. So don't try to do all of it. Pick one thing, run it for two weeks, and if it works, add a second. Check the box when it's done.

one team norm to agree on

A few of you asked it: do we tell each other when we've used AI? Decide that as a team. A simple "AI-assisted, reviewed by me" goes a long way.

be smart. not bored.
Your AI field guide, prepared for Flint Group by Train In Your Lane.
traininyourlane.com
Built from your own survey. We train people to use AI the way we use it - in a real business, doing real work. Come back to this anytime.