# Naval Voice > Rewrite prose in Naval Ravikant's voice, and pressure-test ideas using the reasoning moves that produce that voice. Use when the user says "write this as naval", "naval tone", "in naval's voice", "naval-ify this", "make this sound like naval", "rewrite this like the almanack", or asks to sharpen/compress a draft toward short declarative aphoristic prose. Also use when the user asks to critique, pressure-test, or attack an idea "from Naval's view / frame / lens". Works on essays, memos, book chapters, prefaces, landing pages, threads, and positioning copy. Do NOT use for decisions the user is actually about to make (use decision-lens for that), for fiction, for technical documentation, or when the user wants warm/explanatory/teaching prose. - Skill: `soheilmomeniii/naval-voice` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add soheilmomeniii/naval-voice` - Raw SKILL.md: https://api.skillmd.com/api/skills/soheilmomeniii/naval-voice/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Marketing & Growth - Author: soheilmomeniii (https://skillmd.com/u/soheilmomeniii) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/soheilmomeniii/naval-voice --- # Naval Voice Three modes. Pick from the request. Ask only if genuinely unclear. - **REWRITE** — return their text in his voice, argument unchanged. Default. - **DIAGNOSTIC** — return what is weak in the draft and where the voice would change it. Do not rewrite unless asked. - **LENS** — attack an idea using his reasoning moves. See `references/thinking-moves.md`. ## Two rules that outrank everything below **1. Style is not attribution.** Write *like* him. Never put words *in his mouth*. - Never invent a Naval quote. Take them from `references/corpus-quotes.md` or say you do not have one. - Never write in his first person. Never sign anything as him. Never present the user's argument as his position. The output is the user's piece, in a borrowed cadence, and it stays under the user's name. - If the user supplies epigraphs or quoted material from anyone, reproduce them **verbatim**: punctuation, spacing, attribution marks, and any typo. Rewrite the prose around them, never the quote. **2. Never inject.** Compression is subtraction and sharpening. It is not addition. None of these may enter a text that did not already contain them: - **Invented specifics.** A number, name, date, company, or mechanism the source never had. This is the most tempting fix because a concrete detail always reads better than a vague one. If the piece needs a number it does not have, leave `[FIGURE NEEDED: churn before/after]` and say so in the note. A visible slot is correct. A confident fabrication is a defect. - **Fake first person.** No "I" in the output if there was no "I" in the input. - **Manufactured stakes.** "Now more than ever." "The window is closing." - **Forced contrarianism.** Inventing a foil to argue against is inventing a claim. - **Performed candor.** "Let's be honest." "Real talk." If the rewrite changes what the piece *asserts* rather than how it sounds, append one line: `Note: the original claimed X; this version claims Y.` Silent truth-drift is the worst failure this skill can produce, because it arrives disguised as an improvement. ## REWRITE mode ### Step 1 — Find the argument Strip the piece to its claims before touching a sentence. If a paragraph carries no claim, it will not survive. His compression works because there is something under it. Style applied to filler produces shorter filler. ### Step 2 — Apply the mechanics Full list in `references/voice-rules.md`. The five that do most of the work: 1. **Break every compound sentence at its connective.** "X, because Y" becomes "X. Y." "Which means," "so that," and "and this is why" are usually a period in disguise. 2. **Delete the hedges.** fairly, arguably, mostly, tends to, something like, at least in part, in other words, it's worth noting. Take the risk of the flat statement. 3. **Delete the transitions.** However, moreover, that said, furthermore, therefore. Two facts beside each other imply the relationship. 4. **One idea per paragraph. One to three sentences.** White space is part of the voice. 5. **End on a line that survives being quoted alone.** Every section, and the piece. ### Step 3 — Measure before returning The classic voice-transfer failure is not sounding too formal. It is chopping everything into fragments and calling it rhythm. Count before you deliver. | | Target | Why | |---|---|---| | Median sentence length | ~8 words | matches the corpus | | Body mean | 9 to 13 words | below 9 means you fragmented | | Sentences of 5 words or fewer | under 15% | corpus runs 12% | | Consecutive sentences of 5 words or fewer | never 3 | this is the imitation fingerprint | | Longest sentence | 20 to 28 words, at least one per section | something has to carry load | | Em dashes | zero | hard rule, always | | Semicolons | zero | he does not use them | | Two-beat parallels | 3 to 5 per 1,000 words | see below | **The specific trap, measured.** The two long rewrites this skill produced ran a mean of 8.8 and 10.1 words, which looks right, but 27% of their sentences were five words or fewer and both contained runs of four and five consecutive fragments. The verified corpus runs 12% and never more than two. Same mean, wrong distribution. Aphorisms cluster tightly around eight words; imitation goes bimodal, stacking two-word fragments against long sentences and averaging out to the same number. If your draft looks like that, fuse fragments back into mid-length sentences that carry weight. **Epigram density: 3 to 5 per 1,000 words, at section boundaries.** Stacked aphorisms are the failure mode of imitation. Aphorism-free prose is the failure mode of caution. A page where every line is quotable is a page where nothing is. Then cut. A real pass removes 20 to 30% of the word count. If the draft got longer, you edited instead of rewriting. ### Step 4 — Deliver The rewritten piece first, clean, nothing above it. Then a short note underneath covering: what was cut and why, two or three specifics rather than a changelog; any `[FIGURE NEEDED]` slots; a truth-drift line if one applies; and **one substantive objection to the content itself**, separate from voice. Weak evidence, a buried lede, an unearned credential, a claim that fails the hard-to-vary test. The rewrite is the deliverable. The objection is what makes it worth asking for. Vary the opening move across pieces. A flat assertion, a concrete instance, a question answered in the next sentence, a definition. Never open two consecutive conversions the same way, or the voice becomes a template the user can see. ## LENS mode Run the moves in `references/thinking-moves.md`. Lead with the strongest version of the user's idea, then attack. Two traps, both easy to fall into: - **Quote-stacking.** Ten aphorisms deployed as verdicts is not reasoning. His own warning applies: "If you don't have the underlying experience, then it just reads like a collection of quotes." Argue from the mechanism. Use a quote to sharpen a point you already made, never to settle one. - **Asymmetric standards.** If you kill the user's claim for being unfalsifiable, do not judge it with maxims that are equally unfalsifiable. Apply the standard to both sides or drop it. He is not an oracle. He is one useful frame. Say so if the frame is being pushed past what it covers. ## What this skill will not do - Turn an argument-free draft into aphorisms. Say the draft has no claim in it. - Make a weak claim sound strong. Compression exposes weak claims, it does not hide them. - Add evidence that is not there. If the piece argues that only proof persuades and contains no proof, name that. It comes up constantly in positioning work. - Route every input to the same conclusion. A converter that produces one thesis regardless of input has stopped converting. ## Composition `no-slop` owns AI texture and source fidelity. This skill owns cadence. They stack: run this for voice, then no-slop's tic scan and fidelity gate on the output. Do not reimplement its gates here. Its punctuation bans stay in force under this skill, which is convenient, because they are the same bans. `decision-lens` owns decisions the user is actually about to make. Hand off rather than answering in aphorisms. ## Checklist before returning 1. Every supplied quote verbatim, including punctuation and typos. 2. No invented Naval quote, no first person as him, no signature as him. 3. No invented specifics; gaps marked `[FIGURE NEEDED: ...]`. 4. Truth-drift note appended if the claim changed. 5. Distribution checked against the table, not just the mean. 6. Zero em dashes. Zero semicolons. 7. Every section ends on a line that survives alone. 8. Word count down 20 to 30%. 9. One substantive objection included, about the content and not the prose.