Ghost

Ghost

Write or rewrite text so it sounds like the user wrote it, not like an assistant wrote it.

Voice profile

This skill ships in a shared repo, so the profile itself does NOT live in this repo or in any project's memory system (project memory is scoped to one working directory; this skill needs to work from any of them). It lives in one fixed personal file:

~/.claude/ghostwriter-profile.md

• If that file exists: read it and apply it. It has the user's real sentence-length pattern, vocabulary they actually use vs. avoid, how they open paragraphs/messages, and specific tics. Trust it over any generic default below.
• If it doesn't exist yet: ask once for 2-3 samples of things the user has actually written (Slack, email, LinkedIn, whatever's natural for the register in question). Distill a profile (sentence length variance, real vocabulary, opener habits, tics, how they sign off or don't) and write it to ~/.claude/ghostwriter-profile.md so future invocations, in any project, skip this step. Then apply it to the task at hand.
• Never commit or copy this file into a project repo. It's personal, machine-local data, not part of the plugin.
• If the user corrects the tone, register, or a specific word choice more than once in a session, that's a signal the profile has drifted from how they actually write now. Offer to update ~/.claude/ghostwriter-profile.md with what changed — don't rewrite it silently on a single correction, since one atypical piece of feedback isn't a reliable basis for changing a profile built from real samples.

Rules regardless of profile

• No genre templates. A "new job" post, a "thanks for the feedback" reply, a "following up" email all have a default AI shape (tricolon of nice things, nervous-but-excited contrast, CTA question). Break the shape even when individual sentences pass style checks — templated structure is a stronger tell than word choice.
• No setup → tease → punchline joke structure ("Can't say more yet. [reveal]. So [ironic tag line].") This is the single most common way AI fakes personality in marketing copy. It isn't a named tell in check's catalog, so it needs calling out here. Everything else that pattern-family covers — comparative framing ("more X than Y"), aphorism/landing-phrase closers, parallel-subject mirrors, repeated openers, binaries — is already covered by the post-rewrite check audit below; it isn't duplicated here since that's a real invocation now, not a remembered checklist.
• Status lines are plain text, no leading symbol (no >, no bullet, no emoji) — a leading > renders as a blockquote bar in this client, which isn't wanted here.
• Emit each status line as its own text output immediately after that skill call returns, before invoking the next skill — not batched together at the end. The tool calls and the status lines interleave in the actual order things happen: call check, then write its status line, then call polish, then write its status line, and so on. Never collect them and print them all in a block right before the final text.
• If there is existing text to rewrite (not a from-scratch draft), invoke check with the Skill tool on that input text first, and use its findings to target the rewrite at what's actually firing instead of running every move uniformly.
• That input is usually the user's own writing, so read the voice profile before that call and filter what comes back. A tell whose evidence is a documented profile trait is a false positive here: drop it before the findings reach polish, and keep it out of the status line too. The profile is the record of how this person actually writes, so scoring one of its traits as an AI tell and then polishing it out is the rewrite deleting the voice it exists to preserve. Same for anything check's own calibration already weighs toward human, first-person ambivalence and unresolved tension especially.
• What survives that filter is what the rewrite acts on, so the status line reports the count of surviving findings, never a verdict tier - a verdict on the user's own prose is noise, and on a short post it's an unreliable read besides. Checked input - 2 tells to fix, rest is your voice says the useful part. reads AI, 4 tells fired doesn't. If nothing survives, say that (Checked input - nothing to fix, going straight to voice) rather than inventing work for polish to do.
• Then invoke the polish skill with the Skill tool. The arguments passed to that call must literally include check's findings from the previous step (the specific tells and quoted evidence, not just "clean this up") — "directed by what check found" only happens if the findings are actually in the args; don't rely on them being visible earlier in the conversation, put them in the call itself. After it returns, write its status line, e.g. Cleaned draft. Do not apply its moves from memory — memory drifts and skips steps the actual skill text enforces (its pre-output gate, its full banned-word list). Load it, then draft the baseline hygiene pass for real (burstiness, no hedge padding, no banned AI vocabulary, punctuation normalization), targeted at what check found on the input.
• Then layer the voice profile on top, as its own distinct step with its own status line, e.g. Applied voice profile. The profile overrides polish's generic defaults wherever they conflict (e.g. if the user's real writing runs short and fragmented, don't "improve" it toward more complete sentences). This step doesn't call a Skill tool, but it's real work being done to the draft, so it still gets reported like the others — don't let it happen silently just because there's no tool call attached to it.
• The override only covers structural and mechanical habits the profile actually documents — sentence length, fragments, punctuation like a mid-sentence en-dash aside. It is not a blanket exemption for whatever content happens to sit inside a matching construction. A profile trait says "this shape is how the user writes," not "anything wearing this shape is above scrutiny." If a matching construction is carrying content that's independently generic (a stock compliment, an empty-uplift closer, praise with no specific behind it), that's still worth a redraft using the user's own concrete, specific style rather than kept just because the wrapper matched.
• Matching a trait also means matching its scale, not just its presence. If the profile's example of an en-dash aside is two words on each side, stuffing a three-item comma list behind the same dash isn't "the same trait" - it's a heavier construction wearing similar punctuation, and check's clause-stacking tell will fire on it regardless of how well the dash itself matches the profile. Same logic for any other trait with a size to it: "sparing emoji" means matching the documented count, not just using an emoji; "drops the subject pronoun sometimes" means sometimes, not in every sentence. Check the actual proportions of the example against what's being produced, not just whether the same kind of thing is present.
• The profile is built from the user's own first-person writing, so it has nothing to say about phrasing for third-person content where the user isn't the one speaking (a news blurb, an encyclopedia-style paragraph, a quote from someone else). Don't force sentence fragments or personal tics onto that register just because the profile has them. Still report this step — say what actually happened instead of silently skipping the line, e.g. Voice profile - not first-person content, applied general principles only (flat facts over inflated framing, no brochure language, real sentence-length variance still apply regardless of register).
• For "reply to this": match the register of what's being replied to, but keep the user's voice. A reply to a client email isn't a LinkedIn post.
• Output the rewritten/drafted text only, no preamble beyond the status lines above. No trailing result line after the text — the status lines already told the user what happened as it happened, so there's nothing left to summarize.
• Never print check's full report (the Verdict/Confidence/Score block, the signal breakdown table, the evidence log) as part of this flow — only the one-line status per invocation. Give the full report only if the user asks for it afterward.

Necessary, not sufficient

Invoke check a second time, now on the drafted output, before returning it - load it fresh each time, don't self-score from memory of a past run. This is a different check than the pre-rewrite pass above: that one diagnosed the input to steer the rewrite, this one verifies the output actually improved and didn't introduce anything new. After it returns, output one short status line, e.g. Verified output - reads human. Report just the verdict tier here too, never a raw density number or a made-up percentage (the density is a weighted tell count, not a probability — turning it into "96% human" would be inventing precision the checklist doesn't support).

If the verdict isn't "Reads human" or "Mostly human," don't stop there - go back to polish, same rule as the first call: put this verification pass's specific findings literally in the args, not a vague "try again" (status line: Redraft 1 - reapplying polish), then run check again and report its status line (Verified output - mixed signals). Cap this at two redraft cycles. If it's still not in one of those top two tiers after that, say so plainly in that last status line (e.g. Verified output - still mixed signals after 2 redrafts) instead of looping forever or quietly passing a weaker verdict through.

check covers word-choice tells, sentence-rhythm tells, document-structure tells, and the harder sentence-level rhetorical moves (chiasmus, mirrored subjects, reframe pivots) - plus chatbot-residue and formatting-abuse tells that come from the same broader body of observed AI-writing patterns this whole field draws on.

A clean check result means no known, rule-based tell fired. It is not proof the text will pass a real trained classifier (Pangram, GPTZero). check is a self-graded checklist applied by the same model that may have drafted the text, and short posts (the common case here: LinkedIn posts, replies) simply have fewer words for tells to fire against, so they get a thinner check by construction, not a more lenient one. A real "new job" or "we're hiring" post that read clean by a checklist like this has still been caught by a real classifier as 100% AI before. Don't report a clean check pass as "this is safe" or "this will pass detection." Say what it actually means: no rule violations found here, risk from anything this checklist doesn't cover still possible, especially under 150 words.