user-invocable: true description: Draft (or revise) the X / LinkedIn / Bluesky social copy in a blog post's frontmatter, looping with self-critique until it would pass review. argument: optional file path to a blog post (e.g., content/blog/my-post/index.md). If omitted, picks the blog file currently changed in git status.

Draft social post

Read a blog post, write or revise social copy into its frontmatter social: block, and loop with self-critique until each platform passes review. Preserve existing copy that already works — only repair or draft fresh when there's a real problem.

Flow

1. Resolve the target file.

   • If the user passed a path, use it.
   • Otherwise: run git status --porcelain and pick the single modified content/blog/*/index.md. If zero or more than one match, ask the user which file.

2. Read the blog post in full and any existing social: block. The blog body is the source of truth for every specific you'll use — read it before drafting, not after. Title, URL slug, meta_desc, and tags are not authoritative; the body is.

3. Per-platform: classify and act. For each of X, LinkedIn, Bluesky:

   • Empty (""), missing key, or LLM-fill junk — draft fresh from the blog body using .claude/commands/social-media-review/references/writing-guide.md (see "Drafting from a blank social: block" — it includes the verify-specifics-against-body checklist). Honor platform structure: X = 2 paragraphs, LinkedIn = multiple short paragraphs, Bluesky = 2 paragraphs. Voice: if a single named author wrote a personal-experience post, name them in third person; otherwise corporate "we".

   • Existing copy that breaks a hard rule (banned phrases, "I" voice, URL in body, hashtags, markdown, paragraph structure, char limits — see .claude/commands/social-media-review/references/critique-rubric.md) — minimum-change repair. Drop the hashtag, swap the "I" to "we", trim the over-limit tail. Don't recast paragraphs or "improve" the voice. The original specifics and structure stay.

   • Existing copy that passes hard rules and reads naturally — leave it. Do not rewrite for taste.

   For LLM-fill recognition (staccato fragments as drama, em-dash chains, generic openers like "In today's cloud landscape…", buzzword stacks), see .claude/commands/social-media-review/references/writing-guide.md "Spotting and replacing LLM-fill copy". Don't try to patch around LLM patterns; throw them out and draft fresh from the article.

   For any platform you draft fresh, verify each specific against the blog body before moving on. Every number, name, and attribution must appear in the body — not the title, slug, or meta_desc. If a claim doesn't trace, drop it or replace it.

4. Verify character counts before critique. Pipe each draft into python3 -c "import sys; print(len(sys.stdin.read().rstrip('\n')))" (avoids quoting/escaping problems with apostrophes or triple quotes in copy). Body limits: X = 255, LinkedIn = 2950, Bluesky = 300. Revise any over-limit copy here.

5. Critique loop. Spawn a sub-agent (Agent tool, general-purpose subagent_type) with a clean slate. Brief: read .claude/commands/social-media-review/references/critique-rubric.md, evaluate the copy only (not the blog post), return PASS or FAIL per platform. Default verdict is PASS — only FAIL on hard-rule breaks.

   • All three PASS → go to step 6.
   • Any FAIL → repair based on the cited reasons (minimum-change semantics from step 3), repeat from step 4.
   • Cap at 3 critique rounds. If still failing, keep the most recent draft and note the open issues in the report.

6. Advisory pass — only on kept platforms. For each platform whose existing copy passed hard rules and you left untouched, spawn a sub-agent against .claude/commands/social-media-review/references/suggestions-rubric.md to surface stylistic notes (weak opener, summarizes article, missing pointer, etc.). Do NOT modify the kept copy — surface the notes in step 8's report so the user can decide whether to apply them. PASS stays PASS. Skip this pass on platforms you repaired or drafted — the new copy is the actionable feedback already; layering advisory notes on top is noise.

7. Write the final copy into the frontmatter. Use the Edit tool to replace (or insert) the social: block. Use 4-space indent and | block-literal style for multi-line copy — see content/blog/neo-integration-catalog/index.md for a well-formatted example. Only touch platforms whose copy actually changed.

8. Report. Show the path written to and per-platform status using paired verdict + action tokens (verdict from the critic, action you took):

   • PASS, kept — critic passed the original; preserved as-is
   • FAIL, repaired (N chars changed) — critic failed the original; minimum-change fix
   • missing, drafted — no original (or unsalvageable LLM-fill); wrote fresh
   • Final character counts for each platform
   • Iteration count ("converged in N rounds" or "stopped at 3, X open issue")
   • For each PASS, kept platform with advisory notes, list them under "Suggestions for kept platforms (advisory — not applied)" so the user can decide whether to address them in a follow-up. Omit the section if all kept platforms returned "no suggestions".

Edge cases

• If the blog post has no body content yet (only frontmatter), tell the user — drafting against an empty post produces generic copy.
• If the user wants the existing copy preserved verbatim regardless of issues, they should say so in the invocation; otherwise this skill applies the classification above.
• If all three platforms classify as kept, report that and exit without writing — no edit is needed.