FlintFlintby Torch

Tool map

Which AI should you actually use?

Short answer: pick one general AI and get good at it. The technique Flint teaches transfers to all of them — a good prompt is a good prompt anywhere. Below is what each one is best at, what you stay responsible for, and one control most people never find.

No affiliate links, no referral codes, no prices — those change monthly and we would rather be useful than current. Verify plans and features yourself.

Pick your everyday AI

Claude

The model Flint runs on. Every Forge draft, Skill run, and score here is Claude. Strongest of the four on long documents and careful drafting — paste a contract or a year of reviews and it holds the whole thing.

You stay in the loop

It can be confidently wrong about your local facts, your prices, and your dates. You are the last read before anything reaches a customer.

If you are picking one general AI, pick the one your training is already tuned for. Everything you learn here still works in the others.

Visit Claude

ChatGPT

The most widely used general assistant. Drafting, brainstorming, images, and voice in one place, with the largest pile of community how-tos behind it.

You stay in the loop

Fact-check any name, price, or date it produces before you send it. Check what your plan does with your data before pasting anything about a customer.

Your prompts from Flint work here too. Model names and plan tiers change often — verify what you are actually on.

Visit ChatGPT

Google Gemini

Google's assistant, sitting inside Gmail, Docs, and Sheets. Best fit if your business already runs on Google Workspace and you want AI where the work already is.

You stay in the loop

Confirm figures against the original document rather than the summary. Review anything it drafts from your inbox before it sends.

Capability varies a lot by plan, and the Workspace features are the reason to use it. Verify what your plan includes.

Visit Google Gemini

Perplexity

An answer engine rather than a chatbot. Ask a question about a supplier, a regulation, or a competitor and it answers with links.

You stay in the loop

Open the sources. Citations are the whole point of the tool and they are still sometimes outdated or do not say what the answer claims.

Good for the ten-minute research question. Not where you draft.

Visit Perplexity

Doing the work

Canva

Design with AI built in. Social graphics, flyers, before-and-after posts, and simple logos, from templates rather than a blank page.

You stay in the loop

Check you have the right to every photo, font, and logo you put in a paid ad — AI-generated or not, the liability is yours.

The free tier covers most of what a small business posts. Verify current limits.

Visit Canva

Otter.ai

Records and transcribes meetings and site visits, so the notes write themselves and you can hand the transcript to an AI afterward.

You stay in the loop

Get consent before recording anyone — several states require every party to agree. Never upload a recording that contains health or payment details.

Transcription is accurate enough to work from, not accurate enough to quote verbatim without checking.

Visit Otter.ai

Zapier

Connects the tools you already pay for so a form submission becomes a text, a calendar entry, and a row in a sheet without you touching it.

You stay in the loop

Put a human approval step in front of anything that messages a customer or moves money. An automation that misfires does so at machine speed.

Start with one two-step automation you can explain out loud. The all-in-one build is where people give up.

Visit Zapier

Claude Code

Describe a small tool in plain English and it builds it — a quote calculator, a job tracker, a form that emails you. Flint's build-along lessons use it.

You stay in the loop

Do not paste customer records into anything you built until someone who knows security has looked at where that data goes.

The one entry here that is genuinely technical. Worth it if you have ever wanted a tool nobody sells.

Visit Claude Code

The control almost nobody touches

Turn the thinking up.

All three big assistants let you tell the model how hard to think before it answers. Same idea, three names, and it is a button rather than anything you type. Leave it low for a quick email. Turn it up for a contract, a price, or a decision you will act on.

ClaudeEffort, plus a separate Thinking toggle

low · medium · high (default) · xhigh · max

Click the model name at the top of the chat.

ChatGPTThinking level

Light · Standard · Extended · Heavy

In the model picker, on the plans that expose it.

GeminiThinking level

minimal · low · medium · high

Varies by surface and plan; clearest in AI Studio.

Worth knowing, worth not overselling: more thinking helps on genuinely hard problems, and the research on how much is still mixed. Spend it on the hard things, not on everything.

How Flint stays honest

You'll see lines like “use only the pasted facts” in our prompts. That's not nervousness about the tool; it's the discipline the tool is built on, and how we know it holds:

Gaps come back labeled, never guessed

A fact you didn't paste comes back as a [NEEDS: …] slot for you to fill, not an invented detail.

Scores are computed, not vibes

Lesson grades are arithmetic over per-axis verdicts, so the same prompt gets the same score.

Math gets flagged for checking

Multi-step calculations are shown and tagged CALCULATOR-VERIFY; the last check is yours.

A fabrication screen runs on every release

Automated checks hunt for made-up figures before code ships. Our packaged Skills measure at zero fabricated numbers on that screen.

None of this makes AI infallible. You stay the final editor; the system is built so your edit is easy.

AI-generated — it can be wrong or incomplete. Review before you send, publish, or rely on it.