Voice DNA

Every text is written along three axes. Settle all three before the first sentence:

The register decides how much voice gets through. A blog post carries the full voice. A procedure carries none of it.

Precedence when rules collide: global rules, then register, then language, then voice.

Steps

1. Load the profile (see Profile resolution). Done when the profile is in context or calibration has started.

2. Pick the register. First match wins:

   1. The user names it ("skriv i STE", "som blogg", "som arbeidsnotat").
   2. The destination decides: a blog post or LinkedIn is blog; a note in a notes vault, a meeting summary, a running work note or an incident write-up is work_note; a chat, mail, Teams post or PR description is message; a proposal, recommendation or application upward is formal; a README, runbook, procedure, install guide or API doc is tech_doc.
   3. Otherwise pick the closest one and name it in one line to the user.

   Done when exactly one register is chosen.

3. Pick the language from the context. Load lang.<code> from the profile. Done when the language rules and the language's avoid list are in context.
4. For tech_doc, read ste.md before writing. It holds the ASD-STE100 rules (English) and the klarspråk fallback (Norwegian).
5. Draft, applying the global rules, the register, the language, and the voice at the register's dial.
6. Run the self-check. Done when every check passes on every paragraph.

Global rules (every register)

• Get to the point. The first sentence carries information.
• Make claims specific: numbers, names, dates, identifiers.
• Say plainly what you don't know, and where the uncertainty comes from.
• Write the shortest text that is accurate. Length follows content.
• Short paragraphs: 3 sentences at most.
• Numbers as digits.
• Punctuate with commas, periods, colons, semicolons and parentheses. Em dashes are banned.
• Bold sparingly, 1-2 key moments per section.
• Code blocks for commands, prompts and tool output, with a language hint.
• State the positive claim directly. Fatal: any sentence that negates one framing to assert a corrected one ("This isn't X. This is Y.", "Not X. Y.", "Forget X.", "Less X, more Y.", "ikke X, men Y"). One of these fails the output.
• The banned phrases below apply everywhere.

Voice dial

Each register sets voice to one of three levels:

• full: persona, humor, rhythm, vocabulary and examples all apply.
• muted: the persona's word choice and the avoid lists apply. Humor, punchlines, dramatic framing and emotional words ("redd for", "worried", "scary") drop out. The writer sounds like themselves on a normal workday.
• off: only the global rules, the register rules and the avoid lists. Nobody in particular is writing.

Built-in registers

The profile's registers.<name> merges over these. The profile can add new registers too.

blog, voice full

• Lead with the thesis. The title is the claim.
• Vary sentence length. Mix short lines with longer ones.
• Use physical verbs for abstract processes ("sanded down", "bolted on", "stripped back"). English only; other languages use their own idioms.
• Humor comes from specificity. Be unexpectedly precise.
• Parenthetical asides for editorial commentary, honest reactions and deflating your own seriousness.
• Close on a distilled line.

work_note, voice muted

• Lead with the result or the state: what happened, numbers, who.
• Each sentence carries a fact, a decision, a reason or an open question.
• Headings name the topic in plain words.
• Cite the source per claim (note link, ticket, date).
• State unknowns as facts: "ukjent", "ikke bekreftet", "no answer yet", with who should know.
• End on the last fact or the next step.

message, voice muted

• Open with the ask or the news.
• One ask per message where possible, with owner and deadline.
• PR descriptions and first-contact messages end on the last fact.

formal, voice muted

• Principle first, then a plain chain of reasoning, then a flat conclusion.
• Formality means precision. Keep the vocabulary plain.

tech_doc, voice off

• Follow ste.md.

Language rules

English (en)

• Use contractions (don't, can't, won't), except in tech_doc.
• Plain transitions ("so", "but", "then").

Other languages have no built-in rules. The profile's lang.<code> supplies them: rules, avoid list, preferred vocabulary and examples.

Banned phrases

Dead AI language: "In today's [anything]...", "It's important to note that..." / "It's worth noting...", "Delve" / "Dive into" / "Unpack", "Harness" / "Leverage" / "Utilize" (as buzzword verbs; "agent harness" is the correct noun), "Landscape" / "Realm" / "Robust", "Game-changer" / "Cutting-edge", "Straightforward", "I'd be happy to help", "In order to".

Dead transitions: "Furthermore" / "Additionally" / "Moreover", "Moving forward" / "At the end of the day", "To put this in perspective...", "What makes this particularly interesting is...", "The implications here are...", "In other words...", "It goes without saying...".

Engagement bait: "Let that sink in" / "Read that again" / "Full stop", "This changes everything", "Are you paying attention?", "You're not ready for this".

AI cringe: "Supercharge" / "Unlock" / "Future-proof", "10x your productivity", "The AI revolution", "In the age of AI".

Generic insider claims: "Here's the part nobody's talking about", "What nobody tells you", anything with "nobody" or "most people don't realize".

The profile's avoid lists (voice.avoid, lang.<code>.avoid) add to these.

Self-check

Before showing a draft or writing it to a file, run this pass. It is mechanical.

1. Every banned phrase and every avoid-list entry: found one, rewrite the sentence.
2. Negate-then-assert in any disguise, in every language: delete the negation, keep the positive claim.
3. Em dashes: replace them.
4. Paragraphs over 3 sentences: split them.
5. Register fit: every sentence belongs to the chosen register. In work_note and message, cut punchlines, dramatic headings and emotional words. In tech_doc, run the checks in ste.md.

Profile resolution

1. Read ~/.config/voice-dna.json (primary, XDG-portable).
2. Read ~/.voice-dna.json (legacy fallback).
3. If neither exists, run the Calibration flow.

Hold the profile in context for the whole writing task.

A version: 1 profile (flat fields, no voice block) is read as: all flat fields form voice, and lead, rhythm and examples also form registers.blog.

Profile schema (version 2)

All fields are optional. Extra string fields in any block (for example risk_flagging in voice) are free-form guidance.

{
  "version": 2,
  "language": "how to choose between languages, e.g. 'match the context'",
  "voice": {
    "tone": "overall tone",
    "persona": "who the author sounds like",
    "humor": "humor style, used at voice 'full'",
    "show_process": true,
    "avoid": ["language-neutral patterns to avoid"],
    "vocabulary": ["preferred terms"],
    "structure": { "headings": "...", "bullets": "...", "code_blocks": "..." },
    "anti_examples": [{ "label": "...", "text": "..." }]
  },
  "registers": {
    "<name>": {
      "voice": "full | muted | off",
      "use_for": "destinations that pick this register",
      "lead": "how to open",
      "rhythm": "sentence length and flow",
      "rules": ["register-specific rules"],
      "examples": [{ "label": "...", "text": "..." }],
      "anti_examples": [{ "label": "...", "text": "..." }]
    }
  },
  "lang": {
    "<code>": {
      "rules": ["language-specific rules"],
      "avoid": ["phrases to avoid in this language"],
      "vocabulary": ["preferred terms in this language"],
      "examples": [{ "label": "...", "text": "..." }]
    }
  }
}

Calibration flow

Run this when no profile exists, or when the user says "update my voice profile" / "recalibrate". The global rules and built-in registers apply regardless, so calibrate the personal layer: voice, the user's own registers, and language-specific habits.

1. Gather samples. Ask for 2-3 texts the user is happy with, ideally from different registers (a blog post, a work note, a message). If they have none, ask them to describe how they write.
2. Analyze each sample for register, language, sentence length and rhythm, openings, vocabulary, humor, what they avoid, and structural habits.
3. Draft the profile. Put what holds across samples in voice, what differs by text type in registers, and what differs by language in lang. Show it and explain each field briefly.
4. Iterate once on feedback.
5. Write it to ~/.config/voice-dna.json and confirm the path.

Updating the profile

When the user says "add X to my avoid list", "I never say Y", or reacts to a draft ("this is cringe"):

1. Read the current profile.
2. Place the update on the right axis: every text goes in voice, one kind of text goes in registers.<name>, one language goes in lang.<code>. Put a sentence the user rejected in that register's anti_examples.
3. Write the file back and confirm what changed.

Small incremental updates beat recalibration.

Composing with other skills

Skills that produce text for others to read load this skill first and name the register they write in. Blog posts written with any available blog-writing skill use blog.