Social Fetch Agent Skill

/social-fetch — Pull any social post by URL

Normalized fetcher for social posts across platforms. Detects platform from URL, tries strategies in order, returns the same JSON shape regardless of source.

Step 1 — Detect platform

| URL pattern | Platform |

|---|---|

| x.com/<user>/status/<id> or twitter.com/<user>/status/<id> | x (Twitter) |

| linkedin.com/posts/<slug> or linkedin.com/feed/update/urn:li:activity:<id> | linkedin |

| linkedin.com/in/<handle> (profile, recent activity) | linkedin-profile |

| instagram.com/p/<id> or instagram.com/reel/<id> | instagram |

| tiktok.com/@<user>/video/<id> | tiktok |

| bsky.app/profile/<handle>/post/<rkey> | bluesky |

| reddit.com/r/<sub>/comments/<id>/... | reddit |

| <mastodon-instance>/@<user>/<id> (e.g. mastodon.social, hachyderm.io) | mastodon |

| threads.net/@<user>/post/<id> | threads |

| news.ycombinator.com/item?id=<id> | hn |

| youtube.com/watch?v=<id> or youtu.be/<id> | → defer to watch-video |

If the URL doesn't match any pattern, ask the user what platform it is.

Step 2 — Pick strategy chain

Read references/strategies.md for the per-platform strategy chain. Each platform has 2–5 strategies tried in order.

Key principles:

• Free strategies first (direct APIs, agent-browser)
• Paid only as fallback (ScrapeCreators / Apify) — and only if the env key is set
• Bluesky / Mastodon / HN / Reddit are free + reliable (public APIs)
• X / LinkedIn / Instagram / TikTok / Threads need paid or scraping fallback for full data

Step 3 — Execute strategy

For each strategy in the chain:

• Try it
• If success: parse → normalize → return
• If failure (404, 402, auth wall, empty response): note the failure and try the next strategy

After exhausting the chain, return a clear error: which strategies were tried, why each failed, and what's needed to unlock (e.g., "Add $SCRAPECREATORS_API_KEY for X — see references/auth-keys.md").

Step 4 — Normalize output

Return this shape regardless of platform (see references/output-schema.md for the full spec + platform-specific examples):

{
  "platform": "x",
  "url": "https://x.com/example/status/1234567890",
  "fetched_at": "2026-06-17T14:35:00Z",
  "raw_source": "scrapecreators",
  "author": {
    "handle": "@example",
    "name": "the user Ganim",
    "verified": true
  },
  "posted_at": "2026-06-17T16:53:00Z",
  "text": "The 80/20 of a useful AI second brain: ...",
  "media": [],
  "engagement": {
    "likes": 51,
    "reposts": 13,
    "replies": 9,
    "bookmarks": 7,
    "views": 32700
  },
  "is_thread": true,
  "thread": [],
  "replies": []
}

Fields with no equivalent on a platform (e.g., bookmarks on Mastodon) get null, not 0. Missing data is different from zero data.

Step 5 — Optional enrichments

Based on flags / asks:

| Flag | Behavior |

|---|---|

| --with-replies | Fetch top-level replies (1 hop). Costs extra API quota. |

| --thread | If the post is part of a thread by the same author, fetch the whole thread. |

| --raw | Include the raw API/scrape response in the output (for debugging) |

| --media | Download media files (images/videos) to ~/Documents/social-fetches/<platform>-<id>/ |

Default: just the post itself, no replies, no media download (just URLs).

Step 6 — Cache (optional)

If ~/Documents/social-fetches/_cache/ exists, cache successful fetches there by {platform}-{id}.json for 24h. Saves API quota when the same post is referenced repeatedly across skills.

Skip cache if --no-cache flag is set or for --with-replies / --thread (likely-stale).

Composes with

• deep-research — cite specific posts in research briefs. When research surfaces a relevant tweet/post URL, fetch and include in the brief.
• jab-hook — pull recent posts from inspiration accounts for deeper format analysis (currently uses agent-browser inline; should call this skill instead).
• business-brainstorm — pull competitor / operator commentary as evidence during scoring.
• second-brain — capture a post into raw/ with the tweet- / bookmark- prefix; the structured output makes for cleaner raw files than a screenshot or copy-paste.
• watch-video — for YouTube URLs (or any video — Loom, Vimeo, Riverside, MP4), route there instead.

Known limits

• X: free strategies return tweet preview only (text, author, basic engagement). Full thread + replies need $SCRAPECREATORS_API_KEY or $APIFY_API_TOKEN.
• LinkedIn: agent-browser works for profile recent-activity (after dismissing the modal). Specific post URLs (linkedin.com/posts/...) often need paid fallback.
• Instagram / TikTok / Threads: heavy anti-bot. Paid fallback strongly recommended.
• Bluesky / Mastodon / HN / Reddit: free + reliable.
• Private / deleted posts: nothing helps. Try Wayback Machine for deleted content.

If a platform consistently fails on free strategies and the user uses it often, prompt to set up the paid key (see references/auth-keys.md).

Notes on quality

• Strategy chain, not single-source. Every platform has a fallback ladder (native oEmbed → agent-browser → SCS API → Apify). If one step fails, degrade gracefully to the next. Never fail hard on the first attempt.
• Structured output over screenshots. Downstream skills (jab-hook, deep-research, second-brain) need JSON with author + text + engagement fields, not an image. Even when the underlying strategy is a screenshot, extract text before returning.
• Cache aggressively, invalidate honestly. 24h TTL on ~/Documents/social-fetches/_cache/ prevents API burn when the same post is referenced across multiple skills in a session. --with-replies / --thread skip cache because replies age fast.
• Respect paid-key economics. ScrapeCreators / Apify calls cost real money. Prompt before hitting paid strategies if the user hasn't confirmed they want depth. Free strategies first, always.
• Media download is opt-in. Default is post text only; --media downloads images/videos. Silent media downloads eat disk quickly.
• Private / deleted content is a hard stop. No strategy chain rescues private accounts or deleted posts. Suggest Wayback Machine for deleted content and stop.
• Rate-limits are per-platform. X free strategies hit rate limits fast; LinkedIn agent-browser burns session fingerprints. Space out calls in loops or the workflow degrades to worse-than-manual.