Insaight — Draft LinkedIn Post

Write posts for the user's own company page (COMPANY_NAME and COMPANY_LINKEDIN from insaight:get_config()) by loading that company's actual published posts first and using them as the style reference. The company's own archive is the only reliable source of their voice — anything drafted from general LinkedIn instincts will sound like everyone else's LinkedIn.

Shared conventions (all Insaight skills):

• Call insaight:get_config() once per session for the user's Notion pages and company name/slug. If it reports unconfigured: true, ask the user to edit the file at the returned path before saving to Notion.
• Read cheaply: list_accounts → list_posts (slim index) → get_posts on the few URNs worth reading (max 20 per call). list_people / list_comments are free; the scrape_* tools call Apify and cost money — only scrape when data is missing or stale.
• Engagement benchmarks (adjust for the niche): < 10 likes low, 10–40 normal, 40+ high signal — read those in full.

Step 1 — Load [COMPANY_NAME]'s past posts (style source of truth)

Check whether [COMPANY_LINKEDIN] posts are already in the DB:

insaight:list_posts(account="[COMPANY_LINKEDIN]", limit=30)

If the account returns no posts, scrape once (this costs Apify credits, so avoid repeating it):

insaight:scrape_profile(url="https://www.linkedin.com/company/[COMPANY_LINKEDIN]/", max_posts=30)
insaight:list_posts(account="[COMPANY_LINKEDIN]", limit=30)

Then pull the full text of the 8–12 most recent or most engaged posts:

insaight:get_posts(urns=["urn:li:activity:...", ...])

Pick posts spanning several formats (story, insight, announcement, question). A sample drawn from one format teaches only that format's voice.

Cache these for the whole conversation. Re-fetching on every iteration burns tokens and can shift the style reference mid-thread.

Step 2 — Extract style patterns

Read the fetched posts and note:

• Hook patterns: How does [COMPANY_NAME] open? Bold statement, question, surprising stat, personal story, provocation? Any recurring openers?
• Structure: Short staccato lines, paragraph blocks, numbered lists, mixed? How aggressive are the line breaks?
• Voice register: Corporate/formal vs conversational/peer-to-peer? First-person plural ("we") or impersonal? Any humour or informality?
• Emoji usage: None, occasional, heavy? Where — hooks, section breaks, bullets, CTAs?
• CTA shape: Does [COMPANY_NAME] close with a question, a link, a soft prompt, or nothing? What is the typical last line?
• Hashtag behaviour: How many? Inline or trailing block? Which recurring tags appear?
• Language: Which language(s) do they post in? Any code-switching between languages inside a single post?
• Typical length: Approximate line count and character count range.
• Topics / frames: Which themes recur (product, team, industry take, customer story, opinion piece)?

Write a 4–6 line internal style note before drafting. Having the pattern stated explicitly is what keeps variant B and every later revision on-voice.

Step 2b — Fetch linked content (when a URL was provided)

If the input contains a URL that is not a LinkedIn profile or post URL (those belong to scrape_profile and the research skills), fetch it before drafting:

WebFetch(url="[pasted URL]", prompt="Return the full article or page text, preserving key facts, quotes, and data points.")

This can run alongside Step 1 — the two do not depend on each other.

From the fetched content, extract:

• Core claim or finding — the one sentence that makes this worth posting.
• Specific data points or quotes to anchor the post.
• Angle for [COMPANY_NAME] — how this connects to their domain, product, or point of view.

If the URL is paywalled or returns nothing useful, say so rather than guessing at the contents:

"I couldn't fetch the full article at [URL]. Paste the key paragraph or stat you want to build the post around and I'll draft from that."

Treat the fetched content as the brief for Step 3 — the URL is the brief, so don't ask the user to restate it.

Step 3 — Understand the brief

The user has given a topic, angle, rough idea, or URL. Before drafting, settle internally:

• Goal: Awareness, thought leadership, lead gen, community, announcement, opinion?
• Audience: Prospects, peers, recruits, partners, or general?
• Angle: What is the non-obvious take that makes this worth posting rather than filler?
• Evidence: Is there a specific story, stat, customer, or moment to anchor it? Concrete anchors are what make a post land, so flag it when the user hasn't supplied one.

If the brief is very thin (a few words, no angle), ask one clarifying question first — drafting twice costs more than asking once:

"Quick question before I draft: is this meant to build [COMPANY_NAME]'s thought leadership on [topic], or is there a specific story / moment / customer angle you want to anchor it to?"

Step 4 — Draft two variants

Produce two variants, both in [COMPANY_NAME]'s extracted voice. Two lets the user choose a direction instead of critiquing a single draft.

Variant A — Hook-led / insight-first

Opens with the sharpest statement, stat, or provocation and reaches the point in line 1; the rest supports or expands it. Suits opinion pieces, contrarian takes, industry observations.

Variant B — Story-led / scene-first

Opens inside a moment, conversation, or situation, with the insight landing at the end. Suits customer stories, team moments, "we learned this the hard way" posts.

Match the reference for each variant:

• Line break density — don't default to dense paragraphs if their posts are staccato, or vice versa.
• Emoji frequency and placement.
• CTA style — if [COMPANY_NAME] never hard-sells, a "DM us" closer breaks the voice; only add one if the sample shows it.
• Hashtag count and placement.
• The language(s) they post in, unless the user asks for another.
• Their typical length range from Step 2.

Step 5 — Visual brief (optional, on by default)

After the drafts, add a short visual brief that can go straight to a designer or into an image generation tool:

🎨 Visual brief

Format: [square / landscape / portrait — match LinkedIn norms]
Style: [photography / illustration / data visual / screenshot / etc.]
Mood: [adjectives — e.g. "clean, optimistic, technical"]
Main element: [what should dominate the frame]
Text overlay: [suggested 3–8 word headline to overlay, if any]
What to avoid: [clichés this topic attracts — e.g. "no stock handshake photos"]

Image prompt (copy-paste ready):
"[full prompt]"

Make the image prompt specific about style, composition, subject, lighting and mood; a generic prompt produces generic stock-looking output.

If a visual would feel bolted on (pure opinion piece, personal reflection), say so and skip the brief.

Step 6 — Show your work briefly

Lead with 2–3 lines before the drafts:

• Which style signals you're matching from the loaded posts.
• Which angle you're leading with, and why.

This lets the user course-correct before reading two full drafts.

After drafting

• Offer to iterate: "Want me to sharpen the hook, change the angle, or try a different format?"
• Produce a third variant or a specific format (carousel, poll, video script) on request.
• Don't save drafts to Notion — save-notion is for research briefs, and drafts aren't research. Only save if the user explicitly asks.
• If the user says they published it, offer to scrape it into the Insaight DB so it joins the style reference for next time.