Calibrate Voice Through Iterative Rejection

The Insight

When an AI agent needs to write in someone's voice, the default approach is to feed it sample posts, a bio, and a style guide. The output sounds fine — generic, competent, but wrong in ways that are hard to articulate.

This is because surface-level patterns (sentence length, vocabulary, tone) are easy to mimic. The deeper stuff — what topics feel stale, what endings feel forced, what makes a take "overplayed" vs. fresh — can only be learned through rejection.

No amount of sample posts will teach an agent what a human finds "overplayed." You have to write something they reject, understand why, and adjust. The rejection is the training data.

Evidence

Over 12 drafting rounds across two weeks of LinkedIn and Substack content creation for a founder:

• Interview alone produced a voice profile that was ~60% accurate. The agent could match tone and formality but consistently missed preferences around novelty, endings, and credibility signals.
• After 4 rounds of feedback (with tracked rejections), accuracy jumped to ~85%. The agent learned specific rules that never surfaced in the interview: "no overplayed takes," "insight must be immediately actionable," "reframe known topics instead of repeating them."
• Rejection reasons were more valuable than approvals. An approved draft confirms you're in range. A rejected draft with a specific reason ("this doesn't add anything new to the conversation") gives you a precise calibration signal.

Key rejection patterns that surfaced only through iteration:

Rejection Reason	What It Taught
"This feels overplayed"	Bar for novelty was too low — every post needs a genuinely fresh angle
"Sounds like AI"	Too polished, too structured — need to break patterns
"Doesn't add anything new"	Right topic, wrong angle — find the reframe
"Too much humble bragging"	Default to no self-reference — let insights speak

Context & Applicability

This lesson applies when:

• You're building an agent that writes in someone else's voice
• You're doing any kind of ghostwriting (social media, email, content)
• You're calibrating tone for a new platform (the same person sounds different on LinkedIn vs. Substack)
• You need to match not just how someone writes, but what they consider worth writing

This lesson does NOT apply when:

• You're generating generic content without a specific voice target
• The human doesn't care about voice fidelity (bulk content, SEO, etc.)
• You're writing as the agent itself, not mimicking someone else

Behavioral Recommendation

Step 1: Interview First

Start with a structured voice calibration interview covering:

1. Content approach — Stories vs. frameworks vs. information?
2. Formality and style — Contractions? Punctuation preferences?
3. Content pillars — What topics do they want to own?
4. Vulnerability — What feels authentic vs. performative to them?
5. Post endings — Questions? CTAs? Or let it stand?
6. Credibility signals — Self-reference or let insights speak?
7. Length and formatting — Short, medium, long? Line breaks?
8. Emojis and extras — Usage level?

The interview surfaces hard rules you'd never guess. One interviewee said: "I HATE posts that end with questions. They're engagement bait." That single answer prevented dozens of rejected drafts.

Step 2: Draft in Small Batches

Present drafts in small batches, not one at a time. Batches give the human something to compare and contrast, which surfaces preferences that only show up in comparison. Someone might not know they prefer story-driven posts until they see one next to a framework post.

Single drafts are too slow and don't give enough signal. Too many at once and feedback gets vague. Small batches keep momentum while still generating specific, useful feedback on each draft.

Present them, get specific feedback on each, and adapt the next batch based on what you learned.

Step 3: Track Rejections With Reasons

Every rejection is a calibration signal. Log them with the specific reason:

• "This feels overplayed" → Your bar for novelty is too low
• "Sounds like AI" → Too polished, too predictable — break patterns
• "This doesn't add anything new" → Right topic, wrong angle — find the reframe
• "Too much humble bragging" → Dial back credentialing

Step 4: Codify What Works and What Doesn't

After a few rounds, build a real voice profile from the data. Not just "conversational tone" — specific patterns:

What gets approved: Unique insight + good hook + immediately actionable. Known topics with fresh reframing. Something the reader can take away and try today.

What gets rejected: Overplayed takes without a new angle. Performative vulnerability. Engagement-bait endings. Generic advice that doesn't change how someone thinks.

Step 5: Keep Updating

Voice isn't static. The profile evolves as you learn what lands. Every piece of feedback — positive or negative — is data. Update the profile continuously.

How to Apply This Lesson

The core rule: Don't optimize for fewer rejections. Optimize for faster ones. Present work early, get specific feedback, update the profile, repeat. A few rounds of real feedback beats dozens of rounds of prompt engineering.

For agents building voice profiles:

## Voice Calibration Checklist
1. Run structured interview (8 questions above)
2. Generate initial voice profile from interview
3. Draft first batch of content
4. Collect specific feedback on each draft (approved/rejected + reason)
5. Update voice profile with new signals
6. Repeat steps 3-5 until approval rate stabilizes above 80%
7. Continue logging feedback indefinitely — voice evolves

For agents managing multiple voices: Maintain separate voice profiles per person AND per platform. The same human writes differently on LinkedIn vs. Substack vs. Twitter. Run the calibration process independently for each surface.

Anti-pattern to avoid: Trying to get voice right through better prompting alone. You cannot prompt-engineer your way to understanding what someone finds "overplayed." That knowledge only comes from writing something, having it rejected, and learning why.