self

/idiolect:self — your voice, defended

The roster's voices are fictional. This one is real: a running stylometric picture of how YOU actually write, so drafts in "your voice" are measured against you, not against a compliment. It's also the anti-homogenization move — co-writing with RLHF-tuned models measurably pulls writers toward one style (Padmakumar & He 2024; CHI 2025 showed the pull is strongest on non-Western writers). A maintained self profile is how your idiolect survives the tooling.

Setup

ROOT="${CLAUDE_PLUGIN_ROOT:-${CODEX_PLUGIN_ROOT:-$IDIOLECT_ROOT}}"
# unset everywhere (opencode)? ROOT = two directories up from this SKILL.md
IDIO="python3 $ROOT/scripts/idiolect.py"

How capture works (tell the user this the first time)

On Claude Code, a UserPromptSubmit hook feeds each of your prompts through a filter: prose-like messages ≥ 15 words are kept; commands, code, paths, and logs are rejected. Kept text lands in a LOCAL corpus ($IDIO path → self_corpus) that never leaves the machine, never enters the repo, and caps itself (samples outlive prompts when it compacts). self off stops capture instantly; self clear deletes the corpus. On Codex/opencode there's no per-prompt hook — feed it explicitly with add.

Prompt text is a REGISTER of you (terse, imperative, workish), not the whole you. Explicit samples of real writing (your posts, emails, comments) are tagged sample and weigh more in the build. The more samples, the truer the profile.

Actions

• status — $IDIO self status: corpus size, sample/prompt split, capture on/off, ready-for-build.
• add — user pastes text or gives files: Write to a scratch file, $IDIO self add --file <f> (chunks and tags as samples). This is the highest-leverage thing a user can do for profile quality; nudge for 3–5 real posts/emails.
• on / off / clear — pass through to the CLI; confirm before clear (destructive).
• build (or refresh) — the main event:
  1. $IDIO self status --json — if not ready_for_build, say what's missing (~800 words or 5 samples) and offer add.
  2. $IDIO fingerprint --self --json > fp.json — the measured stylometrics, with the sample/prompt provenance split.
  3. $IDIO synth-scaffold --slug self --from-fingerprint fp.json — skeleton with THEIR real numbers.
  4. Co-author, don't fabricate. The stylo block comes from measurement; the human block comes from the human. Show what you observed in plain terms ("you average 11-word sentences with high variance; you drop capitals in short lines; 'honestly' and 'basically' are load-bearing"), then ASK for the parts no corpus reveals: what they do, what they'll argue about, what they refuse to sound like (that's their banned list — start it with whatever corporate words they hate). Exemplars: prefer 3 REAL posts from their samples (lightly trimmed) over synthetic ones; synthetic exemplars only with their sign-off.
  5. Save to $IDIOLECT_HOME/voices/self.md, run $IDIO validate self (custom voices get the relaxed exemplar threshold), fix until ok.
  6. From now on /idiolect:write self <platform> — <brief> writes as them, conform-checked against their own parameters.
• drift — rerun fingerprint --self, diff the numbers against the profile's stylo block, report what changed ("your sentences got 20% shorter since March; em dashes crept in — keep or update?"). Update the profile only with consent; drift is sometimes growth and sometimes contamination, and only the owner knows which.

Privacy lines (hard)

Corpus and profile are local files under the user's home; nothing is uploaded, and this skill never pastes raw corpus entries into chat beyond short quoted fragments needed for the build conversation. If the user asks what's stored: $IDIO self corpus --tail 10 shows them, self clear erases it, self off stops it. Their voice is their asset — treat the profile file as sensitive as an SSH key.