oli-deacon VoiceDNA
Revise drafts toward the user's actual voice, not toward generic "good writing." Ground strong rewrites in real samples, preserve texture and intent, and use anti-slop rules as guidance rather than as a rigid style constitution.
Workflow
1. Frame the job first
- Identify the source draft, intended audience, and purpose.
- Identify the requested edit depth: light cleanup, moderate rewrite, or strong recast.
- Default to revision, not ghostwriting from scratch, unless the user clearly asks for a fresh draft.
2. Check for real voice evidence
- Start with references/voice-samples.md when that file exists.
- Look for 2-3 real writing samples from the user before doing substantial voice imitation.
- Treat real emails, essays, notes, posts, or draft excerpts as stronger evidence than abstract style instructions.
- If no samples are available and the user wants a strong voice match, ask for samples before rewriting heavily.
- If the user does not provide samples, say confidence is limited and fall back to a light anti-slop edit rather than pretending to know their voice.
3. Build a temporary voice profile
- Read references/voice-profile-template.md and infer the user's:
- cadence and paragraph rhythm
- bluntness vs softness
- humor or dryness
- tolerance for fragments and rough edges
- specificity style and favored details
- recurring transitions or sentence patterns
- phrases, tones, or moves that feel fake in their voice
- Let the samples outrank generic writing rules whenever they conflict.
4. Rewrite toward the profile
- Preserve the draft's meaning, intent, and emotional temperature.
- Keep the user's edges when they feel deliberate: slight roughness, sharp turns, fragments, dry understatement, or clipped endings.
- Remove obvious sludge: padded transitions, vague abstraction, fake certainty, motivational filler, and stock AI phrasing.
- Prefer sanding off the synthetic layer over recasting the whole piece.
- When in doubt, choose the smaller rewrite that still improves the text.
5. Use anti-slop rules as a bias, not a cage
- Read references/anti-slop-rules.md when you need sharper guardrails against generic AI prose.
- Treat hard bans on dead AI phrasing and engagement bait as reliable cleanup targets.
- Treat stylistic preferences such as contractions, parentheticals, or em dash avoidance as defaults that can be overridden by the user's samples or the specific piece.
- Never force the output to satisfy a rule that makes it less like the user.
6. Return the work cleanly
- Return the revised copy first.
- Add a brief note only when useful, for example:
- what voice signals were preserved
- what sludge was stripped out
- whether confidence is limited because samples were missing or thin
- Do not pad the response with a long explanation unless the user asks.
Defaults
- Prefer short paragraphs and concrete wording when that aligns with the samples.
- Prefer natural contractions unless the user's voice or the moment wants formality.
- Prefer specific nouns and active verbs over inflated abstraction.
- Prefer natural transitions over mechanical connectors.
- Avoid over-explaining the edit.
Confidence Rules
- High confidence: multiple strong writing samples with consistent traits.
- Medium confidence: one or two decent samples plus a draft with clear signals.
- Low confidence: no real samples, contradictory samples, or only abstract style instructions.
- When confidence is low, say so plainly and keep the rewrite conservative.
Example Triggers
- "Rewrite this in my voice."
- "Make this sound more like me."
- "De-slop this."
- "Tighten this without making it sound AI-generated."
- "Edit this email but keep my voice."
- "This sounds polished in a bad way. Strip out the generic parts."
References
- Use references/voice-profile-template.md to extract a temporary voice profile from user samples.
- Use references/voice-samples.md as preloaded evidence of the user's voice when it is available.
- Use references/anti-slop-rules.md to catch synthetic phrasing and generic cleanup habits.