Match voice

Match Voice

Rewrite supplied text, or draft from bullets, so it reads in the user's natural voice and in keyboard-typed prose. Preserve meaning; never invent claims or facts. Normalizing typography is about reading like a human typed it on a keyboard, not about hiding that an agent wrote it (surfacing agent authorship is fine).

Output lands in Slack announcements, docs, PRs, commit bodies. It is always properly capitalized (sentence casing).

Modes

• Rewrite — text is supplied: recast it in-voice.
• Draft — only bullets/intent supplied: write it in-voice. Add no facts beyond the input.

The voice

Surface is fixed: proper capitalization (sentence casing), apostrophes intact, no emoji (allowed only when the target is Slack, and opt-in). Lowercase-everything is not the voice, it was just the casual surface. The personality below is the voice, and it carries into formal output unchanged.

• Conclusion first. State the call, then the why. ("Auto-management was paternalistic." then the reasons.)
• Blunt and dry, lightly self-deprecating. The wit is load-bearing, not decoration. Keep the bite when formalizing: "we test locally and merge on green" survives the jump from "i just yolo merge, what could go wrong" — same edge, proper casing. Do not sand it into corporate-neutral.
• Plain words, short precise verbs — wire, gate, cap, shim, sweep, surface, flesh out, lock, pull. Jargon only when needed, defined on first use.
• Label uncertainty; don't hedge it. Say "unsure", "not sure", "open question" as explicit flags. Separate what you know from what you suspect. Never a hedge pileup ("I think this might possibly").
• Parenthetical asides for the side-thought. Bullets over prose for lists. Bold the load-bearing phrase.
• Aphoristic punch where it fits: short declaratives ("This is a feature, not a bug.").
• "we" framing for shared work.

Strip the anti-voice

Cut on sight:

• Corporate filler — leverage, synergy, actionable, ensure, "in order to", "moving the needle", "data-driven * framework", robust, seamless.
• Enthusiasm openers — "Excited to announce", "This PR adds exciting".
• Hedge pileups — "I think this might possibly".
• Closing pleasantries — "Let me know if you have questions!", "Hope this helps".
• AI recap verbs — emphasized, highlighted, stressed, "The team discussed".
• Generic headers — Overview, Background, Introduction. Name the section for what it says.

Normalize typography to keyboard ASCII

Two layers, then verify.

1. Rewrite by hand the constructions that carry meaning: recast em-dash clauses into the user's comma or parenthetical style; turn arrows into words or ->. Don't blind-swap where the phrasing itself should change.
2. Run the deterministic pass on the result (portable sed, no extra files):

   sed -e 's/—/ - /g' -e 's/–/-/g' -e 's/…/.../g' \
       -e 's/[“”]/"/g' -e "s/[‘’]/'/g" \
       -e 's/↔/<->/g' -e 's/→/->/g' -e 's/⇒/=>/g' -e 's/←/<-/g' \
       -e 's/[•·]/-/g' -e 's/≥/>=/g' -e 's/≤/<=/g' -e 's/≠/!=/g' \
       < draft.txt > clean.txt
   

3. Verify mechanically — prove no tells remain, don't assume:

   LC_ALL=C grep -n '[^ -~	]' clean.txt   # any non-ASCII byte (tab allowed); expect no output
   

   The grep is the real guarantee — it catches anything the sed missed (stray unicode spaces, exotic symbols). Fix each flagged spot by hand (e.g. a non-breaking space becomes a normal space) and re-run until clean. No "looks clean" without this grep.

Output

Print the rewritten/drafted text, then a one-line note of what changed:

voice: normalized 3 em-dashes + smart quotes; cut "leverage", "seamless", a closing pleasantry.

Anti-patterns

• Going casual on the surface — lowercase i, dropped apostrophes, unprompted emoji. Capitalize properly (sentence casing); emoji only on an opt-in Slack target. (Keep the personality, drop the casing.)
• Flattening the dry wit and bluntness into corporate-neutral while formalizing. The voice is the content.
• Adding claims or numbers not in the source — this is a voice pass, not a content pass.
• Claiming typography is clean without running the verify grep.
• Blind-replacing every em-dash with " - " when a comma or parens reads more like the user.

See reference.md for the verbatim voice corpus the profile is drawn from.