Two tools worth setting up before you trust AI with real work.
Both are copy-paste scripts for Claude, ChatGPT or Copilot. You set each one up once and it keeps working. These are two of the tools we set up in every Kickstart — written out here in full, free, because they work on any business.
Tool one — emails in your voice Tool two — fewer mistakesMake AI emails sound like you
Most people do the same thing: type a rough email, ask AI to tidy it up, send back what it gives them. It reads fine. It just doesn't read like them — and the people they email every week can tell.
Here is the same email three ways.
The rough draft
hi steve thanks for the quote. we cant do the 3rd, drainage crew are on site that week. can we push it out to the following week? also can you send the safety plan through before you start, need it for the file45 words
Plain AI — "make this more professional"
Dear Steve, I hope this email finds you well. Thank you for sending through your quote — I appreciate you taking the time to put it together. Unfortunately, we are unable to accommodate a start date of the 3rd, as our drainage crew will be on site during that week. Would it be possible to reschedule to the following week instead? Please let me know what works best for your team. Additionally, could you kindly forward your safety plan prior to commencing on site? We will need this for our records to ensure compliance. Please don't hesitate to reach out if you have any questions. Kind regards, Graham110 words
With a voice guide set up
Hi Steve, Thanks for sending through the quote. We can't do the 3rd — the drainage crew are on site that week. Could we push it out to the following week, starting the [date]? Could you also send through your safety plan before you start? I need it for the file. Cheers, Graham55 words
Two things worth noticing in the middle one. It more than doubled in length. And it added two things that were never in the draft — that I appreciated the time they took, and that the safety plan is for compliance.
The third stayed about the length it was written, and where a detail was missing it put [date] in rather than choosing one. That last part is what matters on a job. An invented date in an email about a start on site isn't a style problem.
Step 1 — the script
Open whichever assistant you already use. Paste this in, with three or four of your own emails at the bottom. It gives you back a short set of instructions describing how you write — your voice guide.
Here are some emails I've written myself. Read them and work out how I actually write: how I open and sign off, how long my emails usually are, how formal I am, words and phrases I use, and my spelling and punctuation (I'm in New Zealand). Then write me a voice guide of about 200 words, written as instructions to an AI, covering: - How I open and sign off - Usual length and tone - Words and phrases I use - A "never use" list: phrases you'd normally add to an email that don't appear anywhere in mine - These rules: keep every fact, name, date and price from my draft. Don't add anything I didn't say, especially promises, dates or prices. If something's missing, put it in [square brackets] so I notice. Keep it about the length of my draft and no more formal than I'd be. Reply with just the email, ready to paste. Give me the voice guide as one block I can copy. My emails: --- [paste email 1] --- [paste email 2] --- [paste email 3]
Use ordinary emails, not your best ones — a quote follow-up, a short reply, a longer explainer. Take client names, addresses and prices out first. Then copy the voice guide it gives you: that's what goes into step 2.
Step 2 — put the voice guide where your AI can reach it
This is the only part that changes depending on which assistant you use.
Claude works on the free plan
- In any chat, find Choose style in the toolbar, next to the model name.
- Click Create & Edit Styles, then Create Custom Style.
- Choose Describe style instead, then Use Custom Instructions, and paste your voice guide in.
- Name it "My email voice" and save.
- From then on: paste a rough draft in, pick that style from the menu, and ask Claude to tidy it up.
Shortcut: at step 3 you can choose Add Writing Example and paste your emails straight in instead. It's quicker, but Claude writes its own summary of your writing, so open the style afterwards and add your "never use" list — that's usually what goes missing.
The style has to be selected in each new chat. If you'd rather it switched itself on whenever you mention an email, that can be built as a Claude skill — send us an email and we'll point you at it.
ChatGPT works on the free plan
- In the left sidebar, click Projects and create a new one. Call it "Emails".
- Open the project's instructions and paste your voice guide in.
- From then on, start your email chats inside that project. Paste the draft, ask it to tidy it up.
Project instructions only apply inside that project, and they override your account-wide custom instructions — so this won't change how ChatGPT behaves anywhere else. Don't put the voice guide in the account-wide settings instead, or it will apply to everything you ever ask it.
Copilot a few seconds each time
Copilot doesn't hold onto instructions between sessions the way the other two do, so this one takes a bit more handling. We'd rather say that than pretend otherwise.
- Save your voice guide somewhere you can get at it quickly — a note on your phone, or a draft email to yourself.
- Each time you want an email tidied up, paste the voice guide in first, then your draft underneath it.
If you have Copilot inside Outlook on a paid Microsoft licence, there's a better way: in Outlook, click the arrow next to the Copilot icon → Settings → Draft Instructions, turn on Use custom instructions when drafting email, and paste your voice guide there. It then applies to every draft with nothing to paste.
If you can't find that setting, you're on the free version — stick with the paste-each-time method above.
If it doesn't sound right
Too polished? You've probably fed it your best emails. Swap in ordinary ones — the quick replies you fire off without thinking are the ones that sound most like you.
Still adding phrases you'd never use? Open your voice guide and add them to the "never use" list by hand. It's a plain list and you can edit it any time.
Nothing changed at all? Check the style is actually selected, or that you're in the right project. Both sit quietly and are easy to miss.
Stop AI making things up
AI is confident whether or not it's right. That's the whole problem. It will give you a gradient, a lead time, a clause number or a rate in exactly the same even tone whether it knows the answer or has filled the gap with something plausible.
On most things that doesn't matter much. On a price you quote off, a standard you cite in a handover pack, or a council requirement you plan around, it matters a lot — and those are precisely the questions where it's least reliable, because the answers change.
This script does two jobs. It tightens what the assistant is allowed to assert, and it makes it tell you how sure it is at the end of every answer, with the reason and the one thing most worth checking.
The script
Paste this in once, wherever your assistant keeps its standing instructions. Same text for all three — this one doesn't need to know anything about you.
From now on, follow these two rules in everything you tell me. PART 1 — DON'T FILL GAPS WITH GUESSES - Never invent a number, price, date, product code, standard, clause reference or name. If you don't have it, say you don't have it. - Keep my information separate from yours. If something came from a document I gave you, say so. If it came from your own training, say that too, because that's the part that can be out of date. - Never state a regulation, standard or compliance requirement from memory as though it's current. Name the document I should go and check instead. - If my question is missing something you'd need to answer it properly, ask me rather than assuming. - If something I've told you contradicts something else I've told you, say so instead of quietly picking one. PART 2 — TELL ME HOW SURE YOU ARE End every answer I might act on with exactly this: Confidence: [number]% Why: [one short line] Worth checking: [the single thing most worth verifying] Rules for that number: - Round it to the nearest 10. It's your judgement, not a measurement, and a precise-looking number would oversell it. - 90% or above: it came from something I gave you, or it's settled and uncontroversial. - 60-80%: sound reasoning, but resting on general knowledge or an assumption. - 50% or below: say plainly that you aren't confident, and what you would need to answer properly. - Never go above 90% on a price, a regulation, a standard, a lead time or a date. Those change. Leave this off short factual replies and ordinary back-and-forth. Only use it where I might act on the answer.
What you get back
Ask it something you'd actually act on, and the answer now ends like this.
You asked
Roughly what should I allow to supply and lay 60m of 100mm uPVC lateral in a residential berm?
At the end of its answer
Confidence: 50% Why: The method and the line items are standard, but the rates behind them are general knowledge, not your supplier pricing or your own job history. Worth checking: your last three similar jobs. Your real rates will move this number more than anything else here.
That's the useful version. The number isn't really the point — the two lines under it are. They tell you where the answer came from and what to go and look at before you rely on it.
Be clear about what the percentage is. It's the assistant's own guess at how reliable its answer is. It isn't measured, it isn't audited, and a confident-looking 90% can still be wrong. Treat it as a prompt to check, not a guarantee. That's why the script tells it to round to the nearest 10 — a number like 87% would look like a measurement it has no way of making.
Where to paste it
Unlike the email tool, this one is the same everywhere. It only needs to go somewhere that sticks.
Claude
Settings → Instructions for Claude, if you want it on everything. Or create a Project and put it in the project instructions, if you'd rather keep it to work.
ChatGPT
Settings → Personalisation → Custom instructions for everything. Or a Project's instructions, to keep it to one area of work.
Copilot
Paste it at the top of a session, same as the email tool. It won't persist, so keep it saved somewhere you can grab it.
If you only want it on the questions that matter, put it in a Project rather than the account-wide settings. A confidence read on every bit of small talk gets annoying fast, and annoying is what gets a good habit switched off.
Menus move. These steps were checked in October 2026 — if something has shifted, tell us and we'll fix the page.
What these are, and what they aren't.
These are off-the-shelf scripts. Anyone can use them. The first only knows what your emails sound like; the second doesn't know anything about you at all.
What we build in a Kickstart is a different thing: a tool wired into your actual job files, quotes, permits and registers, so it already knows your forms and the way your business runs. This page is the free version of the idea, not a sample of the paid one.
Nothing you paste here touches our servers. There's no sign-up and no account with us. It all runs inside your own Claude, ChatGPT or Copilot — we never see any of it.
Want this set up properly?
If you'd like these running across the office, or something built for your permits, quotes or compliance paperwork, have a chat with us. Free, no pressure, no obligation.
Email for a free chat See pricing