Training Voice From Writing Samples

Lesson 1: How Many Samples Is Enough

The honest answer: more than you think, less than you fear.

The numbers

1. Bare minimum — 10 samples. Voice will be detectable but slightly wobbly. Use this only if you genuinely don't have more material.
2. Recommended — 25-30 samples. This is the sweet spot. Voice is locked in, patterns are clear, and edge cases (long posts, short posts, replies) are covered.
3. Diminishing returns — past 50 samples. More samples won't sharpen the voice further; they'll just slow training. Save them for refreshes.

Why the floor is 10

Below 10 samples, the system can't separate signal from noise. Your "voice" looks like 5 specific posts on 5 specific topics, and the agent generalizes badly.

Why 25-30 is the sweet spot

Around the 20-sample mark, the system starts seeing the patterns across samples — repeated hooks, repeated rhythms, repeated avoidances. Below that, it sees individual posts. Above that, it sees the writer.

What "samples" can include

• Past social posts (own account or other platforms)
• Email newsletters
• Blog excerpts (use 200-500 word chunks, not whole articles)
• Long DMs or chat messages where you wrote thoughtfully
• Even meeting notes if they capture your voice

The medium matters less than whether the writing actually sounds like you.

Lesson 2: What a Great Sample Looks Like

Not all samples are equal. The right 20 samples beat the wrong 50.

Length

• Sweet spot — 100 to 500 words.
• Too short (under 50 words) — not enough signal. A single tweet doesn't carry your structure.
• Too long (over 1000 words) — patterns dilute. Use a 300-word excerpt instead.

Variety

Aim for a mix:

• 70% in your typical length range — whatever your usual social post length is.
• 20% slightly longer — captures your structure for longer-form posts.
• 10% replies / shorter — captures your voice in conversation, not just broadcast.

Quality

Pick samples that you would post again today. Not your old voice. Not posts you cringe at. The system trains on what you give it — give it what you want it to produce.

Topics

Mix topics. If all 25 samples are about productivity, your agent will gravitate toward productivity even when you want it to write about something else. Voice is mechanics, but if topics are constant in training, the model lazily ties them to voice.

Lesson 3: The Mistakes That Poison Training

Most weak voice profiles aren't underfed. They're trained on the wrong things.

Mistake 1 — Mixing voices in one profile

You upload your social posts (casual founder voice), your customer emails (slightly more formal), and your About page (corporate). The agent averages them. Output is muddy.

Fix: train per voice. Upload only the samples that match the voice you want this profile to produce.

Mistake 2 — Including AI-written posts as samples

You used another AI tool last quarter, and some of those posts are in your "samples" pile. The agent now trains on AI-flavored samples and produces more AI-flavored content.

Fix: only include posts you wrote yourself, by hand. AI-edited is fine. AI-written is not.

Mistake 3 — Using your old voice

You sound different now than you did 2 years ago. If your samples are from 2024, your voice profile is preserving an out-of-date version of you.

Fix: prefer recent samples. If you must include older ones, weight recent samples 70%+.

Mistake 4 — Quoting yourself in samples

A blog post that quotes a tweet you wrote, then a customer testimonial, then your own writing. The system can't tell whose voice it's training on.

Fix: strip quotations. Train on raw your-writing only.

Mistake 5 — Training on hot-take posts when you want a thoughtful voice

You upload your top 25 highest-engagement posts. They're all hot takes. Your agent now writes only hot takes.

Fix: include some posts that didn't go viral. Include some that are thoughtful, slow, contemplative. The agent should learn your range, not just your peaks.

Mistake 6 — Refreshing too often

You retrain voice every two weeks because output drifted. The agent never gets stable. Output is inconsistent.

Fix: voice should be stable for at least 60-90 days. Drift is usually a persona problem, not a voice problem.

Common myths

1. "More samples is always better." Not after 50, and not when half are off-voice.
2. "I should include my best posts only." No — include your typical posts. The agent will produce typical, not best, on most days.
3. "Sample order matters." It doesn't. Bolta's training is order-invariant.

Action Steps

1. If you trained voice from samples, open the sample list. Count how many of those samples you'd actually want to publish today. If less than 70%, plan a refresh.
2. Pick 25 of your best-fit samples — recent, varied in length and topic, all hand-written. Save them somewhere.
3. Schedule the retrain. Block 30 minutes — that's enough.

Sample-trained voice is the highest-fidelity option you have. It just requires picking the right samples. Next lesson: training from connected socials — when it's the right call.