Automatic execution policy: Read automation guidance first. Creative rules and review findings guide improvements; scores, review/hash/freeze records, CTA choices and intermediate approval steps do not block authorized work. Preserve actual external-action authorization and stop at the artifact the user requested. This policy supersedes older quality-gate wording below.

Platform guide — the SoT for grammar and specs

A knowledge-only skill holding the writing grammar, video specs, limits, and anti-patterns for the four publishing platforms: Threads · Instagram · Facebook · YouTube. produce (§9 per-platform text), publish (approval-gate review), and the content-reviewer agent all take this skill's playbook as their baseline.

Core principles (summary)

1. "Share the facts, never the sentences" — even with the same material, redesign the speaker, register, sentence endings, and information density for each platform. Automatic cross-posting is banned.
2. Plain language — use the channel audience's knowledge and voice; default to adult non-specialists. No unexplained jargon, no translationese, no over-compressed sentences with the subject dropped. If a necessary term is unfamiliar to that audience, lead with the plain word and gloss the term in parentheses on first appearance only.
3. The first encounter is everything — the first line on Threads, the first 125 characters of an IG caption, the YT title+thumbnail, the first second and then the first 3 seconds of the video. If you don't win there, the rest never gets read. The first second is a measurement, not a feeling: after the build, the first subtitle cue's start time is read off the shipped SRT, and past 1.0s it warns (playbook §7).
4. Golden-hour replies — answer comments in the first 60 minutes after publishing; the same reactions spread further when they land early (grow-threads references/growth-playbook.md §Principles). The one sourced lift figure is Buffer's 10.2M-post Threads set: posts where the author replied to their own post took +42% engagement, the largest of the six platforms it tracks.
5. Korean without AI tells — one sentence of translationese, stock phrasing, or assistant-speak makes the whole post read as an ad and get scrolled past. The rules live in references/korean-style.md; the findings come from references/check-style.py. Narration, subtitles, copy, titles, and comment drafts are all in scope.

Platforms at a glance

Detailed grammar, copy formulas, the anti-pattern checklist, and video specs (safe zones, subtitles, length) are in references/platform-playbook.md — Read it before writing for any platform, without exception.

Style check

After writing or editing any Korean copy, run the checker for its surface. The code lists the findings; a human or agent reads them and fixes the sentences.

PG=${CLAUDE_PLUGIN_ROOT}/skills/platform-guide/references
python3 $PG/check-style.py --surface <surface> <file>
node $PG/extract-text.js ./storyboard/scenes.js <narration|subtitle|screen> | \
  python3 $PG/check-style.py --surface <surface> -

8 surfaces: narration subtitle screen threads ig fb yt reply. exit 0 pass / 1 warning (S2 accumulation) / 2 S1 found / 3 the check did not run (empty input · path error) / 4 skip (the text is not Korean, so the rules have nothing to say about it).

S1 findings are suggestions. Fix the sentence and re-run, or write down why it stays; neither the exit code nor a score decides whether authorized work goes on (owner directive 2026-10-16). exit 3 is not a pass — fix the path and run again. exit 4 is not_applicable, not a pass and not a failure. Write paths against ${CLAUDE_PLUGIN_ROOT} (produce·publish run from data/<channel>/episodes/<topic>/, so relative paths won't resolve).

The YouTube meta file gets a second, structural check on top of the style one — node $PG/check-meta.js output/youtube/meta.md reads the playbook §6 layout, the title limits, the preset hashtags, the summary voice, and a verbatim copy of COMPREHENSION.answer (it finds scenes.js from the episode layout). Same exit codes. Its 0 is "layout and verbatim leak clean" — a paraphrased result is the reviewer's blind read. A short whose cover opens on the result (hookType:"spoiler") keeps this check: the cover may say the answer, the title and the description still withhold it (playbook §2). Only long-form answer-first episodes skip it.

Editorial review guidance for outgoing copy

Read outgoing copy for factual support, natural language and fit with the channel and conversation. A style checker or independent reviewer can help when a concrete issue needs another look. Findings and scores inform revisions; no threshold, review count or intermediate approval blocks already-authorized work. Record missing reviews honestly. Publication and replies still require actual user or standing authorization for that external action.

Additional Resources

Reference Files

• references/platform-playbook.md — per-platform grammar in detail, copy formulas, video specs, anti-pattern checklist (SoT)
• references/korean-style.md — Korean style SoT: AI-tell pattern tables (T·D·C·A·E), severities, per-surface application, the prefer-deletion principle
• references/check-style.py — deterministic style checker (stdlib only; --surface · --json · --doc · --selftest); findings are suggestions, exit codes report severity
• references/check-meta.js — the YouTube meta.md check: the §6 layout, title limits (100 chars · no <> · front-loading warning), the format preset's required hashtags, the summary voice, and a verbatim COMPREHENSION.answer in the title or description (--print title|description|tags · --selftest)
• references/pipeline.js — the step catalog: the stage ladder, the checks and reviews on the attended and unattended paths with what each one decides (only operational contracts stop work; reviews file findings), the reviewer verdict-tail contract, and produce's three build lanes. episode-state.js derives an episode's stage from it; docs/pipeline-manifest.md explains the design
• references/pipeline-lint.js — checks the workflow entrypoints and reviewer table against pipeline.js (7 rules, read-only; --list · --selftest)
• references/extract-text.js — scenes.js → per-surface text extraction (narration · subtitle · screen; handles the window.SCENES global contract)
• references/safezone-landscape.md — the 16:9 safe-zone measurements (2026-08-17): x 96/1920 · top 96/1080 · bottom 285/1080, and which surfaces stay provisional. formats.js, the video template and the storyboard template all cite it
• references/reply-gate.md — reply context and optional editing diagnostics, with actual sending authorization kept separate
• references/formats.js / references/format-resolve.js / references/format-lint.js — the format presets, the resolver the builders call, and the linter that holds the inline mirrors to them
• references/skill-lint.js — the skill-surface check: routing descriptions, tool lanes, reference links and orphans, manifest parity, the tool count in prose (16 rules; --list · --selftest)