Instagram Extract

Use Instagram as a “find + capture” stream: the user pastes a post link and says something generic like “save it”, and you decide what’s worth capturing, how to summarize it, and where it belongs.

The workflow should be agentic (use judgment) rather than deterministic.

Hard constraints:

Use the logged-in browser relay (browser tool with profile="chrome") when available.
Classify + save into the right pillar so it’s reusable later.
Report back in bullet points (no full transcript unless asked).
After saving, notify the media agent by appending a compact JSON record to the shared inbox file (details below).

Scope constraints (current reality):

This skill is for Instagram + Threads + LinkedIn (logged-in via Browser Relay / existing Chrome/Brave profile).
X/Twitter: don’t try to force it in Brave if it’s blocked; treat as out-of-scope for this skill (we’ll use a different approach/tool later).

Pillars (classification targets)

Pick the single best fit (canonical save) and optionally add a secondary tag when it overlaps. Do not duplicate saves across pillars unless the user explicitly asks.

Current pillars Michael wants to save:

Use these pillar ids in the media-agent inbox payload:

Prompts — creating images (prompts_image)
Tutorials/knowledge — marketing workflows (tutorials_marketing) — frameworks, copywriting, growth, positioning
Design systems / design inspiration (design_inspo) — UI patterns, components, typography, layout
Prompts — web design / development (prompts_webdev) — UI prompts, landing page prompts, component prompts

Also include optional tags like: storytelling, hooks, carousel, landing_page, ui, prompt_template, checklist, case_study.

Lightweight workflow (guidance, not a script)

1) Get access to the real post (relay)

Try to use browser relay takeover (profile="chrome").
If relay isn’t attached, ask the user to attach it (toolbar icon → badge ON).
If token is required, have the user set the gateway token in extension Options.

2) Extract enough to be useful

Prioritize signal over completeness.

For Instagram/Threads:

Caption: prefer og:description / og:title.
Carousel: grab slide text via image alt text (fast); OCR only when the slide text is garbled or crucial.

For LinkedIn:

Expand “see more” if present and extract the full post text.
If it’s a document carousel or images and text extraction is limited, still capture the media URLs + summarize what’s visible.

For all platforms:

Links: collect outbound URLs/domains if present.
Media: collect image URLs when available (good enough for downstream reuse).

3) Decide what matters (summarize)

Write 5–10 bullets focusing on:

the core idea / framework
anything immediately actionable
any reusable templates (hooks, structures, checklists)
any concrete examples worth saving

4) Decide what to do with it (agentic routing)

You’re optimizing for future reuse, not perfect capture.

Default behavior Michael wants:

If the user says "save it" and the post is weak/low-signal (generic motivation, no reusable artifact), ask once: “Not worth saving (too generic). Save anyway?”
If the user says "force save" / "save anyway": save regardless.
If the user says "light save": save a tiny note (title + link + 3 bullets) even if it’s generic.

Use these heuristics (not rules):

If it’s directly reusable instructions (a prompt, a template, a checklist, a swipeable structure) → save it.
If it’s vibes with no payload (generic motivation, no concrete steps) → don’t over-save; create a tiny note only if the framing is uniquely good.
If it’s ambiguous, ask one clarifying question max (e.g., “Do you want this saved as marketing knowledge or as design inspiration?”). Otherwise, pick your best guess and proceed.

Routing:

Prompt content (copy/paste generation instructions) → ingest via prompt-kb.
Tutorials/knowledge (marketing workflows) → local KB note.
Design systems / design inspiration → local KB note focused on patterns.
If it contains both a reusable prompt and a useful framework → save both (local KB summary + prompt-kb ingest).

5) Save (create durable artifacts)

A) Local KB save (tutorials/knowledge/design inspiration)

Save one canonical note in a stable, platform-agnostic structure:

kb/pillars/<pillar_id>/YYYY/MM/<YYYY-MM-DD>-<short-slug>.md

Where pillar_id is one of:

tutorials_marketing
design_inspo (If it’s a prompt-first asset, you can still save a short KB note, but the canonical storage may be prompt-kb.)

Also store a raw capture (optional but recommended when extraction is messy) in:

kb/inbox/<platform>/YYYY/MM/<YYYY-MM-DD>-<short-slug>.md

<platform> is instagram | threads | linkedin.

Note format (keep it tight):

Title (human readable)
Source URL
Platform (instagram|threads|linkedin)
Pillar (primary) + optional tags (secondary)
5–10 bullet summary (your synthesis)
Why it matters (1–2 bullets)
Reusable templates/checklists (if any)
Links found (if any)
Media (image URLs if available)
Optional: extracted text only when it’s actually useful later

Slug guidance:

Prefer something that will be searchable later: <topic>-<creator> or <framework>-<topic>.

B) Prompt vault ingest (prompt-kb)

Use prompt-kb.
Tagging:
- Always include one: prompts | tutorials | resources | ideas
- Add image_gen / video_gen / ui_design when appropriate
- Add tool/model tags if known
Multiple slides/images for the same prompt template → shared promptIngestKey.

C) Media-agent handoff (always)

Append a single JSON line to: kb/_media_agent_inbox/ingest.jsonl

Schema (keep it compact):

{
  "createdAt": "YYYY-MM-DDTHH:mm:ssZ",
  "platform": "instagram|threads|linkedin",
  "url": "...",
  "pillar": "prompts_image|tutorials_marketing|design_inspo|prompts_webdev",
  "tags": ["optional", "tags"],
  "title": "...",
  "summary": ["...", "..."],
  "links": ["..."],
  "media": ["https://..."],
  "repurpose": {
    "x": {"hook": "...", "outline": ["...", "..."]},
    "carousel": {"slideTitles": ["...", "..."]},
    "shortVideo": {"beats": ["hook...", "beat...", "cta..."]},
    "cta": "..."
  },
  "saved": {
    "kbPaths": ["kb/..."],
    "promptKb": {"ingestKey": "optional", "tags": ["..."]}
  }
}

Implementation hint: use a tiny Python one-liner to append JSONL so you don’t need to read/overwrite the file.

6) Telegram report (bullet points only)

Return:

what it is (1 line)
5–10 key takeaways
why it matters (1–2 bullets)
links found (if any)
what got saved + where (file path and/or “ingested to Convex”) + pillar + tags
Repurpose Pack (ALWAYS) — written in Michael’s voice (blunt, marketing-first, specific):
- X post: 1 hook + 3–5 bullet outline
- Carousel: 6–8 slide titles
- Short video: 15–30s script beats (hook → beats → CTA)
- CTA: suggested comment keyword / next step