Sounds Like Me

Sounds Like Me

Audit an existing file (script, teleprompter, blog draft, social post) so every line sounds like Ryan. This does NOT write new content — it finds off-voice lines, shows them to Ryan, and fixes the ones he approves. Findings first, fixes on approval.

Step 0 — Load the voice cards (REQUIRED, every time)

The cards are the single source of truth. Read them live — never rely on a copy in this skill, and never hardcode the rules here (they'd drift). They ship in sibling skills of this plugin, so read them relative to this skill's directory:

${CLAUDE_SKILL_DIR}/../youtube-script/references/style-card.md    # core voice — ALWAYS read
${CLAUDE_SKILL_DIR}/../wnd-script/references/wnd-series-card.md   # WND delta — read for WND files

When the plugin is loaded in place (symlinked from the source repo), these resolve into git and refinement-log updates persist. If the resolved path is under ~/.claude/plugins/cache, still read it, but warn in Step 7 that any refinement-log entries written there are lost on the next /plugin update and must be ported to the source repo.

For written content (blog posts, tutorials, recipes, workshop sections, docs), also read the STE card — it defines the action zones this skill must leave alone:

${CLAUDE_SKILL_DIR}/../ste-pass/references/ste-card.md            # zones — read for written tutorials/docs

1. Always read style-card.md in full — the core voice, including Anti-AI tells and Hard bans. Its refinement log lives in a separate refinement-log.md; don't read it, every entry is already distilled into the card.
2. If the target is a What's New for Developers (WND) file, also read wnd-series-card.md in full (open/close/identity deltas). Detect WND from the filename (WND, What's New for Developers) or content (branded greeting open, "have a good one" close). Other series may have their own card — read it if one exists.
3. If unsure which series a file belongs to, ask before passing.

Step 1 — Read the target(s)

Read the current file(s) — the user may have edited them by hand since they were written. Never work from memory or an earlier version; the file on disk is the source. If pointed at a folder (e.g. a WND month folder), find every script/teleprompter in it and treat them as a set.

Step 2 — Find (don't fix yet)

Go line by line through the spoken/written content only (skip [SCREEN] directions, > notes, headers).

Skip STE action zones in written content. Numbered steps, prerequisites, **WARNING:** / **CAUTION:** / **NOTE:** blocks, UI click paths, and troubleshooting tables follow ste-card.md, not the voice card. Plain, imperative, joke-free wording there is intentional — never flag it as off-voice or rewrite it into chattier prose. If an action zone has a clarity problem, suggest running ste-pass instead. Scripts and teleprompters have no action zones.

Flag anything that trips a card rule:

• Anti-AI tells and Hard bans from style-card.md (dive in / delve / supercharge, label-colon openers, "real talk," relentless parallelism, press-release rhythm, etc.).
• Off-voice phrasing the cards flag — corporate/PM framing ("the action item changed from X to Y"), announce-the-feeling ("here's the part I love"), meme/internet slang, cost/price framing for effort, and over-use of a single emphasis word (e.g. "whole" more than ~2–3 times across a file).
• Series conventions (WND): branded greeting open (NOT a cold open), "…have a good one" close (NOT "friends"), identity = developer advocate at Automattic.
• The gut check: if a line wouldn't survive Ryan reading it out loud, it's a finding.

Step 3 — Present findings for approval (DO NOT edit yet)

Show Ryan a numbered list. For each finding give: the file + rough location, the offending line, the card rule it breaks (name it), and a proposed rewrite in his voice. Group by file. Keep it skimmable. Then ask which to apply — "all," specific numbers, or skip. Note any borderline ones separately (you're unsure it's really off-voice) so he can rule on them. Make no edits in this step.

Step 4 — Fix the approved findings in place

Apply only the approved rewrites with Edit. Preserve the user's manual edits and the file's format (read-copy teleprompter vs. **VO:** script). Don't touch facts, URLs, or > ✅ / ⚠️ notes.

Step 5 — Keep versions in sync

If a script and a teleprompter exist for the same piece, apply each approved fix to both so they don't drift. Sync only the spoken wording — respect each format's conventions (teleprompter spells out URLs, softens code identifiers; script keeps [SCREEN] lanes and > notes).

Step 6 — Log new corrections

When an approved fix reflects a new rule Ryan would want enforced next time, append a dated bullet to the correct refinement log (don't duplicate an existing rule):

• Core-voice correction (true on any channel/series) → the relevant section of style-card.md, plus a dated bullet in ../youtube-script/references/refinement-log.md.
• Series-specific correction → the relevant section of that series' card (e.g. wnd-series-card.md), plus a dated bullet in that card's refinement-log.md. Routing to the shared card is what keeps the voices from drifting — when unsure whether it's core or series-specific, prefer the shared style-card.md.

Step 7 — Report

Briefly recap: what you changed (tie each to a card rule), which files you synced, any refinement-log entries added, and any findings Ryan skipped or borderline lines still open.

Non-negotiables

• Findings first, fixes on approval — never edit in the same step you surface findings.
• Read the cards live every run; never inline their rules here.
• Never voice-edit STE action zones in written content — they belong to ste-card.md.
• Audit the on-disk file, not memory — honor manual edits.
• Keep script + teleprompter in sync on approved fixes.
• Route new rules to the right refinement log (core → shared style card).