Repurposer (Cross-Platform → LinkedIn — Phase 20)

Turns content from another platform — a tweet/thread, YouTube video, blog post, or newsletter — into a native LinkedIn post: re-hooked for the fold, expanded into the algorithm's verified length sweet spot, with any link moved out of the body, run through the Humanizer, and written to Drafts/ via the same status: draft flow every other draft uses. This is the first skill in this repo that goes other platform's content → LinkedIn — every existing multi-platform writer (write-draft-x, write-draft-substack-article/-note) goes the opposite direction, LinkedIn-pipeline idea → another platform. This skill adds a new input path only; it never adds a new approval path (REQUIREMENTS.md §31).

Arguments

• source-type (required) — one of tweet, thread, youtube, blog, newsletter. Determines which input-handling rule below applies.
• source (required) — either a URL or pasted content (text, transcript excerpt, or summary), per the source-type-specific rules below. Both a URL and pasted text may be given together (e.g. URL for reference/dedup filing, pasted text as the actual content to draft from).

Input handling (per source type)

All four source types sit under REQUIREMENTS.md §27's shared "reading third-party post content" convention: no official API exists anywhere in this repo for reading an arbitrary third-party post/page/video from just a URL; pasted text is the primary, reliable input; a URL is best-effort metadata only; no skill in this repo scrapes, uses session cookies, or otherwise bypasses login/ToS. This skill cross-references §27 rather than restating it — what follows is only the source-specific judgment call on top of it.

• Tweet/thread: per §27, pasted text is primary. A URL may be attempted via WebFetch as a best-effort convenience, but X/Twitter typically requires login for full content server-side — treat any fetch result as unverified. Show the user what was fetched and ask them to confirm it matches what they actually see when they view the post themselves, or ask for the text pasted directly, before proceeding to extraction.
• YouTube video: a mandatory manual-paste case, permanently. No tool in this session's toolset retrieves YouTube captions/transcripts, and WebFetch on a YouTube URL returns the page shell (title, description, metadata), not the transcript. Always ask for a pasted transcript excerpt or a written summary of the key point(s) — never attempt a fetch expecting transcript content, and never claim to have "watched" or reliably retrieved the video's content. This is a deliberate, permanent design stance, not a gap to fill later.
• Blog/newsletter: WebFetch is genuinely plausible here (ordinary HTML renders fine) — attempt it as the primary path. Fall back to asking for pasted text only if the fetch fails outright or returns unusable content (a paywall, a JS-rendered body that comes back empty/garbled, or a result that's mostly navigation/boilerplate rather than the article itself).

If nothing usable is obtained through any of the above — fetch failed, fetch returned unverified/unusable content and the user didn't confirm or paste an alternative, or no source was given at all — stop and ask for a manual paste. Never proceed to extraction or drafting on a guess.

Process

1. Get source content

Per the input-handling rules above. Do not proceed past this step without usable, confirmed-or-pasted content.

2. Extract the core claim(s) and any link(s)

Identify the core argument/claim(s) the source actually makes, and any link(s) worth preserving (the source's own URL, or a link it cites). Do not carry over any statistic, quote, or claim not literally present in what was actually retrieved or pasted — the same never-fabricate discipline every drafting skill in this repo applies to its own grounding material (REQUIREMENTS.md §21).

3. Dedup / fatigue check

Check Content-Learnings/content-index.md for topic/hook/argument overlap with existing drafts or published posts — the same discipline /write-draft's pipeline (via /critique-draft) and /generate-ideas already apply (REQUIREMENTS.md §15/§25.4). If content-index.md doesn't exist yet or is too thin to trust, fall back to scanning Drafts/ and Published-Posts/ directly. Flag — don't silently draft — if this repurposed angle duplicates something recent; a flag doesn't block drafting, but it must be surfaced plainly in the report back.

4. Read the voice guide

Read Content-Learnings/voice-guide.md fresh (never a cached memory of it, since it's a living document) before drafting.

5. Re-hook for the fold

LinkedIn truncates behind a "see more" click after roughly the hook-cutoff thresholds in Content-Learnings/algorithm-rules.md: 210 characters on desktop, 140 on mobile (REQUIREMENTS.md §30, Verified Numeric Threshold). Write a genuinely new first 1-2 sentences that state the concrete claim/tension on their own — this is a rewrite of the opening, not a trim of the source's own opening line, which was written for a different platform's context/audience assumptions (a tweet's first line assumes a very different reader attention pattern than a LinkedIn feed scroller does).

6. Expand to the length sweet spot

Expand into the 900–1,300 character sweet spot (REQUIREMENTS.md §30, Verified Numeric Threshold — cited, not re-derived) via:

• (a) a concrete example or analogy consistent with but not present in the source,
• (b) unpacking an implication the source stated tersely,
• (c) adding the personal-take framing LinkedIn favors ("here's what I'd add...", "the part I keep coming back to...").

Never add a new factual claim beyond what the source actually said — expansion is framing/elaboration only, never new unverifiable facts.

7. Move any link to the first-comment field

Per REQUIREMENTS.md §30's documented 40-60% external-link-in-body reach penalty (Verified Numeric Threshold), any link worth preserving does not go in ## Post Text. Store it in the Draft Note's source_link frontmatter field instead (see step 10), and note explicitly in the "Report back" that it needs to be posted as the first comment manually after the LinkedIn post itself goes live — this repo has no mechanism to auto-post a follow-up comment (same deliberate-manual precedent as visuals, REQUIREMENTS.md §7, and Substack publishing, §25.3).

8. Hashtags

3-5, directly tied to topic/category, no stuffing — same convention /write-draft uses.

9. Humanize the draft

Call /humanize-draft's documented Input/Output Contract (.claude/skills/humanize-draft/SKILL.md) as the final style pass:

• text = the drafted body from steps 5-8.
• platform = "linkedin".
• draft_id = left unset on this first call (the Draft Note doesn't exist yet — it's created in step 10, immediately after).

Use the returned revised_text as the final post copy. Do not re-implement any of Humanizer's own vocabulary/em-dash/pattern-density detection logic here — call the contract, use its output.

10. Write the Draft Note

Copy _Templates/Draft-Note.md into Drafts/ as YYYY-MM-DD--kebab-slug.md (today's date — a repurposed post has no target_date to inherit from an Idea Note). Fill in:

• platform: linkedin, status: draft.
• category, format, hook_style — same judgment call /write-draft applies, based on the repurposed content's actual topic/shape.
• hashtags from step 8.
• source_type — the source-type argument (tweet | thread | youtube | blog | newsletter).
• source_link — the link from step 7, or left blank if the source had none or none was worth preserving.
• idea_id and spine_id both left empty — this draft has neither; it did not come from an Idea Note or a Story Bank Post Spine. Its grounding is the source_type/source_link fields instead; scripts/validate_vault.py treats a non-empty, non-"none" source_type as a third valid grounding for the vault schema's idea/spine check, so this is expected to validate cleanly, not an error to work around. Never fabricate an idea_id or spine_id to satisfy a check that doesn't apply to this path.
• sources: the vault schema requires this list non-empty on every Draft Note (same as the idea/spine paths), and there's no research-note or Story Bank row id to cite here — populate it instead with the source reference itself (the URL if one exists, e.g. "https://x.com/...", or a short descriptive string like "pasted tweet, no URL given" / "pasted YouTube transcript excerpt, no URL given" if none does). This keeps the citation-retention discipline (REQUIREMENTS.md §17) intact even though the grounding is external content, not an internal Research Note.
• hook_formula, engagement_goal, founders_angle: leave blank — this skill doesn't select a canonical hook formula or founders angle; its hook comes from the re-hook step (5), not the shared formula taxonomy.
• history: [{action: created, date: today, note: "Repurposed via /repurpose-post from <source_type>: <one-line description or URL of the actual source>"}].
• ## Post Text: the Humanizer's revised_text from step 9.
• ## Sources (internal — not part of the post): the source content's own reference (URL if given, or "pasted <source_type> content, no URL given") — for traceability, not for publication.

Leave viral_score at 0 and visual_ids empty — scoring and visuals are /critique-draft and /generate-visual's jobs, not this skill's.

Report back

• The full drafted post text and its character count (confirm it lands in the 900-1,300 sweet spot, or say by how much it doesn't and why).
• The hashtags used.
• The first-comment link text, if any, and an explicit reminder that posting it is a manual step after the LinkedIn post goes live.
• Which piece of source content backed each claim/example in the draft — traceable back to what was actually fetched/pasted.
• The content-index.md dedup check result (clean, or what it flagged).
• The Humanizer pass's caveats string, surfaced verbatim, plus its score_report and changes_made if non-trivial.
• A plain statement that this is status: draft, waiting for /critique-draft next — nothing has been approved, scheduled, or published, and no comment has been auto-posted anywhere.

Hard rules

• Never carry over a stat, quote, or claim from the source that isn't literally present in the fetched-and-confirmed or pasted content.
• Never fabricate what a tweet/thread/video/article said if the fetch attempt failed or returned unusable content — stop and ask for a manual paste instead of guessing or paraphrasing from a title/preview alone.
• Never claim a YouTube transcript was "watched" or reliably retrieved — this is always a manual-paste case, permanently, not a temporary limitation.
• Never treat an unconfirmed WebFetch result (tweet/thread) as ground truth — get the user's explicit confirmation, or a direct paste, first.
• Never auto-publish or auto-post the first comment. Output is always status: draft, flowing through the existing /critique-draft → /audit-draft → /review-drafts → /schedule-approved pipeline unchanged.
• Expansion (step 6) adds framing, examples, and elaboration only — never a new unverifiable fact, statistic, or quote beyond what the source actually said.
• Never leave a link in ## Post Text — it always moves to source_link once one exists worth preserving.
• Never silently draft a near-duplicate of a recent post — the content-index.md check must be run and its result reported, even when clean.