Build a voice-trained LinkedIn content engine that does not sound like AI

Gather 10-20 source samples from the person whose voice you want to preserve: LinkedIn posts, emails, podcast clips, sales call excerpts, internal rants, and rough notes. Use Descript to transcribe spoken samples and save everything in Airtable with source type, topic, tone, and what made it strong. Do not only collect polished posts; the messy spoken material is often where the real voice lives.

Use Claude to analyze the source library and produce a voice profile with sentence patterns, recurring beliefs, strong opinions, favorite phrases, banned phrases, pacing, humor level, and examples of acceptable and unacceptable rewrites. Save this profile as a living document in Airtable.

Create a LinkedIn voice profile from these writing and transcript samples.

Source samples:
{{source_samples}}

Person role:
{{person_role}}

Audience:
{{audience}}

Output:
1. Voice summary
2. Sentence structure patterns
3. Common opening styles
4. Topics and beliefs they can credibly own
5. Phrases they naturally use
6. Phrases they would never use
7. AI-sounding phrases to ban
8. Examples of good rewrites
9. Examples of bad rewrites
10. A checklist for reviewing future posts

Be specific. This should help a writer produce posts that sound like this person, not a generic executive.

Create an Airtable table for raw ideas with columns for source, rough thought, audience, pain point, proof, story, format, priority, draft status, reviewer, and performance notes. Add weekly intake sources such as voice notes, customer calls, founder observations, sales objections, and product lessons.

Use Claude to convert raw ideas into post briefs, not final posts yet. Each brief should include the hook direction, core point, supporting story, proof needed, risk notes, and the exact audience pain being addressed. Reject ideas that do not contain a real point of view.

Turn these raw ideas into LinkedIn post briefs using the voice profile.

Voice profile:
{{voice_profile}}

Raw ideas:
{{raw_ideas}}

For each idea, output:
1. Recommended post angle
2. Audience pain addressed
3. Hook direction
4. Core argument
5. Proof or story needed
6. What to avoid saying
7. Risk or approval notes
8. Priority score

Reject ideas that are too generic or not credible for this person.

Use Claude to draft one sharp version and one safer version for each approved brief. Include the voice profile, banned phrases, and examples of posts that performed well. Keep drafts under the target word count and require each one to make one clear argument.

Draft LinkedIn posts from these briefs.

Voice profile:
{{voice_profile}}

Banned phrases:
{{banned_phrases}}

Approved briefs:
{{post_briefs}}

For each brief, create:
1. Sharp POV version
2. Safer version
3. Suggested first-line hook
4. Optional closing question
5. Notes on where the draft may still sound too polished

Rules:
- One clear argument per post
- No generic AI phrases
- Use concrete examples
- Do not invent personal stories or metrics
- Keep the voice natural, not corporate.

Before scheduling, run every draft through a QA pass. Check for banned phrases, vague claims, over-polished tone, missing proof, repetitive sentence structure, and whether the author would actually say it aloud. Use AuthoredUp to preview the first two lines and formatting on LinkedIn.

Review these LinkedIn drafts against this voice and quality checklist.

Voice profile:
{{voice_profile}}

Banned phrases:
{{banned_phrases}}

Drafts:
{{drafts}}

For each draft, identify:
1. Lines that sound AI-generated
2. Vague or unsupported claims
3. Voice mismatches
4. Repetitive structures
5. Stronger hook option
6. Final edited version

Do not make the post more polished. Make it more human and specific.

Schedule approved posts and track performance in Shield. Every month, review which hooks, topics, formats, and tones attracted the right comments, profile views, saves, and conversations. Update the voice profile and idea scoring rules based on what actually worked.