Naturalize

Naturalize

Rewrite text so it reads like the user actually wrote it, not like an LLM drafted it — grounded in a profile calibrated once from the user's own writing (not guessed from a generic "sound more human" instruction), then reused silently on every later request until the user asks to change it.

<SKILL_DIR> below is this skill's own announced base directory (the path shown when this skill loads, ending in .../skills/naturalize) — references live at <SKILL_DIR>/references/.

Step 1 — Determine request mode

Default behavior is profile mode: the saved profile (if any) governs everything. The user can override this for a single request by explicitly asking to skip it — "naturalize this generically," "without using my profile," "ignore my profile for this one," "just de-AI-ify this, don't personalize it." That sets explicit_generic mode for this request only.

• profile mode (default) → go to Step 2.
• explicit_generic mode → do not read or write ~/.naturalize/profile.md at all for this request — not even to check whether it exists. Skip straight to Step 5, using the checklist's stated defaults (references/checklist.md): everything on except typos, no custom instructions, no voice/sample matching. This is a one-off override — it never triggers calibration and never modifies a saved profile.

If it's ambiguous whether the user means this request only or wants to change their standing default, ask; don't silently reinterpret a one-off request as a permanent profile change (that's what Step 4 is for) or vice versa.

Step 2 — Check for an existing profile

Look for a profile at ~/.naturalize/profile.md (the user's home directory — same external-storage pattern the estimator plugin uses for ~/.estimator, keeping personal writing samples out of the skill package itself).

• Found → read it and skip straight to Step 5.
• Not found → run Step 3 (first-run calibration).

Never re-run calibration just because the skill was invoked again. Only an explicit ask ("recalibrate," "redo my profile," "start over") re-triggers full calibration; smaller asks ("turn off typos," "update my formal sample") go through Step 4 instead.

Step 3 — First-run calibration

Run this once, in order. The full toggle catalogue lives in references/checklist.md — load it before Step 3c rather than reconstructing it from memory. The saved-file structure lives in references/profile-template.md — load it before Step 3g.

3a. Default voice. AskUserQuestion: Informal or Formal, with Informal as the recommended option. This becomes default_voice, used in Step 5 whenever a request's register can't be resolved any other way.

3b. Phase 1 — writing samples. Ask for a real informal sample (e.g. a Slack message or text) and a real formal sample (e.g. an email or doc excerpt). For whichever one the user can't supply, offer two options:

• Write a substitute — a few sentences spanning past, present, and future tense, enough to observe structure and habits without a real artifact. That register still gets personalized.
• Skip it. If they skip, say plainly what that means before moving on: that register will run in generic mode — the Phase 2 checklist (de-AI-ifying rules) still applies, but output won't be tailored to their specific voice, since nothing was learned about it. It'll read like a general humanizer pass for that register, not "sounds like you." They can add a sample later (Step 4) to personalize it whenever they want.

Track each register's resulting mode — personalized or generic — for Step 3g.

3c. Phase 2 — togglable checklist. Present references/checklist.md's groups via AskUserQuestion (multiSelect per group): symbols, vocabulary, structure, openers/closers, formatting, hedging/tone, and typos. State the protected-content carve-out (code, quotations/citations, URLs, numbers, proper nouns, domain jargon are never touched) once, up front — don't repeat it per toggle. Typos default off; everything else defaults on.

3d. Custom instructions. Ask if there's anything else to add as free text (e.g. "always use British spelling," "never start a sentence with 'So'"). Persists by default; can be overridden for a single later request without changing the saved profile.

3e. Preview toggle. Ask whether the user wants a before/after preview shown each time text is naturalized, or just the rewritten result. Store as Preview: enabled.

3f. Synthesize and confirm. For each personalized register, turn its sample into a short characteristic summary — sentence rhythm, formality markers, greeting/sign-off style, contraction/emoji use. Skip synthesis for any generic register — there's nothing to summarize. Show the whole assembled profile back to the user in plain language (not the raw file format), and if either register ended up generic, restate that plainly in the summary — e.g. "your formal writing will use general de-AI-ifying rules only, not a voice matched to you, since no sample was given." Let them correct anything before saving.

3g. Save. Write the confirmed profile to ~/.naturalize/profile.md using the structure in references/profile-template.md.

Step 4 — Partial recalibration (any time, no full redo)

Handle these without re-running all of Step 3:

• Add/replace a sample for one register → re-run just 3f's synthesis for that register, leave the other register and all rules untouched. Adding a sample to a generic register switches it to personalized.
• Toggle one rule ("turn off typos," "stop flagging passive voice") → flip that single entry in the Rules section.
• Edit custom instructions → replace or append to that field.
• Change default voice or preview setting → update that single field.

Always write the change back to ~/.naturalize/profile.md immediately — don't batch edits across a session. (This is distinct from Step 1's explicit_generic override, which touches nothing on disk.)

Step 5 — Resolve register for this request

1. Explicit wins — "as an email," "for Slack," "make it formal" (or informal) directly sets the register.
2. Otherwise infer from the content/context being naturalized (greeting style, recipient cues, subject line) and state the assumption briefly so the user can correct it.
3. Otherwise fall back:
   • profile mode → the profile's default_voice.
   • explicit_generic mode → no profile to fall back to; default to informal as a neutral standalone default.

Step 6 — Apply the rules

Determine this request's effective mode for the resolved register:

• generic if the request is running in explicit_generic mode (Step 1), or if profile mode is active but the resolved register's saved mode is generic (no sample on file for it).
• personalized otherwise — profile mode, resolved register's saved mode is personalized.

Rewrite the input text:

• Vocabulary: common words over rare synonyms, corporate filler stripped, domain jargon left alone — per whichever toggles are on.
• Symbols: ASCII-typeable forms over Unicode lookalikes, per the table in references/checklist.md, honoring any symbols_exceptions the profile records (personalized effective mode only — generic uses the hard default with no exceptions, since the profile isn't consulted).
• Structure / openers-closers / formatting / hedging: apply whichever toggles are on, per references/checklist.md's descriptions.
• Typos: only when the resolved register is informal and the toggle is on, at the stored frequency (rare/occasional), never in protected content. (generic effective mode: typos off, per the checklist default.)
• Custom instructions: personalized effective mode only — apply the profile's stored text plus any one-off instruction given for this specific request. generic effective mode ignores stored custom instructions (there's no profile in play) but still honors a one-off instruction given directly in this request.

Which toggle values to use:

• effective mode == personalized → use the profile's saved Rules section.
• effective mode == generic → use references/checklist.md's stated defaults (everything on except typos), regardless of what any saved profile has toggled — a generic pass never reflects saved customizations, only the standing default checklist.

If effective mode == generic, don't attempt to match tone, rhythm, or greeting style — apply only the toggle-driven rules above. This is the "no personalization layer" path: functionally a general de-AI-ifier/ humanizer pass rather than a voice match.

Never touch protected content — code blocks, quoted/cited text, URLs, numbers, proper nouns, domain terminology — regardless of which toggles are on.

Don't over-enforce. An "on" toggle applied with 100% consistency (e.g. literally zero dashes, perfectly even sentence lengths) reads as its own kind of artificial pattern. Allow rare, natural-looking exceptions.

Step 7 — Deliver the result

• profile mode, and the profile's Preview setting is on → show a before/after (or a short list of what changed) and let the user confirm or ask for adjustments before treating it as final. This applies regardless of whether the resolved register's effective mode ended up personalized or generic — Preview is a standing profile setting, not tied to one register.
• explicit_generic mode, or profile mode with Preview off → output the rewritten text directly.

Add one brief trailing note whenever effective mode is generic, so the user isn't left assuming this was matched to their own voice when it wasn't. Keep it to one line; don't repeat the full explanation from calibration. Pick the wording that matches the actual reason:

• Request-level override → "(generic pass, per your request — your profile wasn't consulted.)"
• No sample on file for this register → "(generic de-AI-ify pass — no voice sample on file for this register; add one anytime to personalize.)"

Error handling

• No profile exists, profile mode is active, and the request isn't a calibration request → run Step 3 first; don't attempt to naturalize text against a profile that doesn't exist yet. (Doesn't apply in explicit_generic mode — that path never needs a profile.)
• Profile file exists but doesn't match the expected structure, and profile mode is active → say so plainly and offer to recalibrate rather than guessing at a partial read. If the user actually wanted explicit_generic for this request, that still works fine regardless of profile-file health, since it's never read.
• Register genuinely ambiguous and there's no fallback that obviously fits (e.g. a message that reads as neither clearly formal nor informal, and either there's no default_voice in play or it doesn't fit either) → ask once, don't guess silently.