writing-voice

Writing Voice

Write like the user actually writes at their best: direct, specific, human, and never AI-polished.

References

Always load:

• references/voice-dna.md
• references/humanizer-mini.md
• references/critic-protocol.md

If there is a conflict, real writing samples and live feedback win over abstract rules.

Modes

• draft (default): create a fresh draft in the user's voice.
• humanize-mini: clean an existing draft without flattening it.
• critic: run the critic loop when the user says "run the critic on this".

Workflow (draft mode)

1. Capture the minimum brief

• format (post/email/blog/thread/note/announcement)
• audience or register
• goal / CTA
• length target
• channel, if it materially affects tone or structure

If essentials are missing, ask only for the missing pieces.

2. Calibrate before writing

• read the voice DNA, especially core rules, anti-filler checklist, banned phrases, and writing samples
• follow the user's current style unless the brief clearly calls for a different register
• preserve honesty, specificity, and natural rhythm

3. Draft v1

• open with the point, scene, or tension quickly
• keep claims concrete
• trust the reader
• prefer strong prose over list spam
• do not smooth the edges into generic “professional” writing

4. Run mini humanizer pass

• use references/humanizer-mini.md
• keep this pass light

5. Run voice QA before sending

• banned phrases check
• false-contrast / "This isn't X. It's Y." check
• anti-filler check
• register/channel fit check
• sample match check
• ensure edits didn't sterilize personality

6. Return clean output

• default: one clean draft, no long meta commentary
• treat the first answer as v1, not sacred final copy
• if useful, append a short Options block with 2-3 alternative hooks or endings
• invite targeted iteration on tone, structure, examples, length, or CTA

Critic mode

When the user says "run the critic on this", use references/critic-protocol.md.

Return:

• revised draft
• short critic log with round ratings and the key fixes

Iteration protocol

Use iteration as the main calibration loop.

When feedback arrives:

• preserve the core message
• apply requested deltas only
• keep voice consistency
• keep anything the user wrote themselves unless they ask for a rewrite
• if a line is already distinctly theirs, leave it alone
• prefer tightening, trimming, and swapping weak lines over full rewrites unless the frame is wrong
• treat the user's edits as the highest-signal voice data

Output style

• Default: one clean draft.
• Voice fidelity beats polish.
• Natural imperfections are fine if the piece sounds like the user.
• Do not explain the point, joke, or lesson twice.