AI writing sounds generic when it has to guess. If the system does not know what changed, what the company believes, which facts are current, how the founder actually writes, or what the reader should gain, it fills the gaps with safe marketing language.
The fix is not a longer prompt full of tone adjectives. Build a small source system, give the draft one job, provide a voice fingerprint, and review with rules that can reject fluent nonsense.
Separate truth, voice, and the current assignment
- Product truth: what the product does, audience, positioning, proof, prohibited claims, current limits.
- Voice fingerprint: sentence length, vocabulary, rhythm, humor, formatting, phrases you use and phrases you avoid.
- Source note: the specific event, evidence, decision, or question behind this post.
- Assignment: reader, platform, job, desired response, and what the post must not become.
Replace tone adjectives with observable patterns
A model can follow patterns it can see. Include three representative writing samples and annotate why they sound right. Also include one polished-but-wrong example and explain why it should never ship.
Too vague
Sound bold, human, authentic, and thought-provoking.
Observable voice
Open with the concrete observation. Prefer sentences under 18 words. Use one dry aside at most. No rhetorical question in the first line. Avoid 'unlock', 'game-changing', 'seamless', and any claim without a source.
Give the post one reader outcome
A draft becomes generic when it tries to educate, announce, build authority, sell, and start a conversation at the same time. Pick the primary job before generation.
- Teach one method.
- Explain one decision.
- Challenge one assumption.
- Document one result and its limits.
- Ask one question the audience can answer from experience.
Ground every claim in a source
The grounded version is less grand and more credible. It preserves the sample, distinguishes a test from field data, and does not manufacture an activation result.
Source
Three of five interviews mentioned setup confusion. The new flow is 15 minutes shorter in our internal test.
Unsafe draft
Our revolutionary onboarding eliminates friction and dramatically boosts activation.
Grounded draft
Three of five users we interviewed got stuck on the same setup step. We made it optional; our test path is now 15 minutes shorter. We still need enough real usage to know what that changes downstream.
Write rejection rules before generating
- Reject if the post could describe any software company.
- Reject if it uses an unsupported superlative or performance claim.
- Reject if the product appears before the reader receives value.
- Reject if the platform version is only a character-trimmed copy.
- Reject if the first sentence is a generic question or motivational slogan.
- Reject if a private customer detail or invented quotation appears.