How to fine-tune AI for your business voice (Australian small business guide)

What “fine-tuning” actually means (and why you don’t need it)

In the technical world, fine-tuning means training a custom AI model on your data. It produces a model that has your style baked in at the weights level. The mechanics are real, but for most Australian small businesses it’s the wrong tool:

The voice-file approach gets you 90% of the outcome of real fine-tuning for 1% of the cost. It’s the right answer for almost everyone reading this.

The 90-minute method

Block one Sunday afternoon or one rainy Tuesday. Do these five steps in order.

Step 1: Write the voice file (30 minutes)

The voice file is 200-500 words that tells Claude who you are, who you’re writing to, how you sound, and what’s off-limits. The template:

WHO I AM
I run [BUSINESS NAME], a [TYPE OF BUSINESS] in [CITY/SUBURB], Australia.
We do [WHAT YOU DO IN ONE SENTENCE]. Our customers are mostly [WHO THEY
ARE, age range, situation, vibe]. We've been at it [HOW LONG].

WHO I'M WRITING TO
Most of my writing is aimed at [PRIMARY AUDIENCE: existing customers,
prospects, both]. They know [BACKGROUND ASSUMPTIONS]. They don't know
[THINGS WE SHOULDN'T ASSUME]. They care about [WHAT MATTERS TO THEM,
in their words, not yours].

TONE
We sound [3-5 ADJECTIVES: warm, dry, slightly cheeky, plain-spoken, etc.].
We don't sound like [3-5 NEGATIVES: corporate, salesy, gushing,
American, over-friendly, etc.].

HARD RULES
- No em-dashes
- Australian English (organised, colour, centre, analysed, recognise)
- AUD currency unless I say otherwise
- DD/MM/YYYY dates
- No exclamation marks (or: only in [SPECIFIC CASES])
- Banned words: the standard AI-tell list (the L-word for "use",
 the E-word for "raise", the U-word for "release", the D-word for "explore"),
 plus: curated, full, unlock, solid, clean, big shift,
 empowering, journey, ecosystem, synergy

VOICE SAMPLES
Five sentences that sound exactly like me, copied from real past writing:
1. [paste one]
2. [paste one]
3. [paste one]
4. [paste one]
5. [paste one]

OFF-LIMITS
- [Topics, claims, or angles never to touch]
- [Competitors you don't mention]
- [Confidential things]

OUTPUT DEFAULTS
- Default length: [short / medium / long]
- Default format: [paragraphs / dot points / structured headers]
- If I haven't told you the length, ASK before generating long output.
- If I haven't told you the audience, ASK before generating outreach copy.

Some operators do a longer version (1,000-2,000 words) with more samples. Diminishing returns kick in fast. 500 words is the sweet spot for most.

Step 2: Choose your voice samples (30 minutes)

The voice file’s “Voice Samples” section is where most operators fall down. They paste either nothing, or generic marketing copy from their website. Neither works.

What you want: 5-10 pieces of writing you wrote (not your copywriter, not AI, you) that you’d be proud to publish again. Examples:

• A customer email where you nailed the tone
• A social caption that got good engagement
• A paragraph from a blog post you wrote yourself
• A line from an old newsletter
• A reply to a complaint where you struck the right note

Avoid:

• Anything an AI helped with
• Anything from your “About Us” page (usually too polished)
• Anything you copied from a template

5 strong samples beats 20 mediocre ones. Quality of samples is the limiting factor on quality of output.

Step 3: Set up the Claude Project (15 minutes)

• Open claude.ai, make sure you’re on Pro
• Left sidebar: Projects → Create Project
• Name: your business (or the brand voice you’re building)
• Instructions field: paste your voice file
• Knowledge area: upload your full sample writings (PDFs, DOCX, or paste as MD files). Also upload: product/service list, pricing page, FAQ, brand style guide if you have one.

Save the Project.

If you’re a ChatGPT Plus user, the equivalent is My GPTs → Create a GPT, paste your voice file into the Instructions, upload samples in Knowledge.

Step 4: Test it (10 minutes)

Open a fresh chat inside the Project. Run this prompt:

“Write a draft email to a customer who’s been with us for 12 months, thanking them for the year and inviting them to leave a Google review. Use the voice in my voice file.”

Read the output. Compare it to the same prompt run in a fresh non-Project chat. The difference should be obvious by the second sentence.

If it’s not, the voice file is too thin. Common fixes:

• Add 2-3 more voice samples
• Tighten the “we don’t sound like” list (be more specific)
• Add more no-no words to the rules
• Add an explicit length default (“Unless I say otherwise, keep emails under 100 words”)

Step 5: Use it for everything for two weeks (then refine)

For the next 14 days, run every writing task through the Project. Emails, captions, blog drafts, internal Slack updates, anything. Notice patterns:

• Claude keeps using a word you don’t use? Add it to the no-no list.
• Claude defaults too long? Add a length cap.
• Claude misses a brand-specific nuance? Add an explicit rule.
• Claude nails it 90% of the time? Now spend the saved time on the parts AI can’t do.

After two weeks the voice file is sharp. From there, monthly 10-minute reviews keep it accurate.

The maintenance cycle

Voice files decay. Your business changes, your audience evolves, new no-no words emerge. The monthly maintenance:

On the first of every month, 10 minutes:

1. Open your voice file.
2. Read it as if for the first time. Is it still accurate?
3. Add anything new (new product, new audience segment, new no-no word).
4. Delete anything that’s no longer true.
5. Re-upload to Claude Project if you’ve made non-trivial edits.

Every quarter, 30 minutes:

1. Pull 5-10 of your best AI-assisted outputs from the quarter.
2. Read them. Are they in your voice? Where do they still feel off?
3. Find the gaps in the voice file. Patch them.
4. Refresh your voice samples (replace older ones with stronger recent ones).

This 10-minutes-a-month discipline is what separates operators who say “AI made my content generic” from operators who say “AI saves me 4 hours a week.”

Multi-brand operators: the rule

If you run more than one brand (agency with multiple client voices, holding company with multiple businesses, founder with a side hustle), the rule is: one Project per voice.

Don’t try to merge them. We’ve tried. The output goes bland.

The pattern that works:

• Project 1: “[Brand name 1] voice”
• Project 2: “[Brand name 2] voice”
• Project 3: “[Brand name 3] voice”

Each Project has its own voice file, its own samples, its own Knowledge files. Switching between them in Claude is one click. The mental load is light.

For agency creative directors, we go a level further: one Project per client, plus one master Project for the agency’s internal voice. Five clients = six Projects. Total Claude cost: $30 AUD/month flat. The time saved on each client per week pays the bill in the first session.

Voice tuning for specific output types

Different writing jobs need different sections of the voice file activated. The pattern that works inside a single Project:

For emails (1:1 replies)

Add to your prompt: “Reply in my voice. Match the tone of [past sample email]. Length: equal to the customer’s email or shorter.”

For social captions

Add: “Three caption variants. Each [N] words. Mention [specific detail]. End with [your usual close, e.g. ‘open from 7am Sydney Rd’). No exclamation marks.”

For blog drafts

Add: “First draft of an article in my voice. Match the tone and structure of [past blog sample]. Target audience: [specific segment]. Word count: [N]. Include: [specific things].”

For ad copy

Add: “Three variants of [headline / primary text / description]. Match the voice in the file. Each variant tries a different angle: [angle 1, angle 2, angle 3]. Max [N] characters.”

These prompt extensions stack on top of the voice file. The file gives the constants (who you are, how you sound); the prompt gives the variables (what specifically, how long, who for).

What can’t be voice-tuned

Be honest about limits.

• Subject matter expertise. A voice file makes Claude sound like you. It doesn’t make Claude know what you know. If you write content that requires deep professional expertise (medical, legal, financial, deep technical), the voice file handles the how; you still provide the what.
• Genuine opinion. AI defaults to balanced. A voice file can tell Claude “we have strong opinions about [topic]” and steer it. It can’t replace you actually having an opinion worth sharing.
• First-hand reporting. If your content requires being somewhere, talking to someone, observing the thing yourself, AI can’t substitute. It can write up your notes after.
• Strategic decisions. Voice = tone. Strategy = direction. Don’t confuse the two. AI’s good at the first, mediocre at the second.

The voice file is a tone amplifier. It makes the wheel spin faster. The wheel still has to be pointed in the right direction by a human.

What’s next

• Claude for the not-quite-beginner for the broader workflow context.
• How to stop AI content sounding like AI for the editing pass after generation.
• Custom GPTs vs Claude Projects if you’re choosing between platforms.
• Book a free 30-minute audit if you want help designing the voice file for a multi-brand business.