linkedin

name	linkedin
description	Draft, post, and review Dan's LinkedIn posts (handle saattrupdan) via the `linkedin` CLI, which drives the real LinkedIn web UI with agent-browser. Use when the user wants to write/post a LinkedIn post, save or view a draft, or fetch their recent posts with engagement stats. Also use whenever drafting LinkedIn content, to match Dan's voice and formatting.
last-updated	2026-07-01T00:00:00.000Z

linkedin

Manage Dan's LinkedIn presence (profile handle: saattrupdan) from the terminal. LinkedIn has no public API for drafts, scheduling, or reading your own posts, so this skill drives the real LinkedIn web UI via agent-browser, wrapped in a linkedin CLI so the browser steps are deterministic.

CLI

linkedin <command> [options]

Prerequisites

which agent-browser || npm i -g agent-browser && agent-browser install

The linkedin command is an editable install pointing at linkedin_skill/main.py, so edits to the source take effect immediately. The CLI shells out to agent-browser and runs the browser headless — no window appears. It does send keystrokes to the editor (via agent-browser type, which targets the editor element), but these are in-page key events, never OS-level chords, so nothing leaks to the OS. Standard library only.

All browser commands use --session-name linkedin for persistent cookies/storage between invocations, so login state is preserved automatically.

Authentication

Credentials live in the macOS Keychain (service linkedin), never in a file and never echoed. Read/set them with:

security find-generic-password -s linkedin -a username -w      # read
security add-generic-password  -s linkedin -a username -w <email>
security add-generic-password  -s linkedin -a password -w <password>

Then:

linkedin login

The login flow (all automated except the one 2FA step):

1. The CLI opens the login page and fills the credentials from Keychain. After a logout LinkedIn shows a "Welcome back" page that remembers the account and asks only for the password — the CLI handles both that and the full email+password form (the email field is optional). It clicks the real "Sign in" button, carefully skipping the "Sign in with Google/Apple" SSO buttons.
2. LinkedIn then issues a verification challenge — this is the one manual step:

   • App push: approve it on your phone. linkedin login polls for up to --wait-push seconds (default 60) and completes automatically.

   • Authenticator code (LinkedIn often demands this instead of a push): the browser daemon stays alive on the code page, so supply the code with:

     linkedin login --code 123456
     

On success the session persists and future runs skip login. linkedin login --verify checks the current state; --force re-runs even if a session exists (a still-valid session redirects to the feed and is treated as already logged in).

Session fragility: rapid/interleaved browser activity (e.g. mixing raw agent-browser commands with linkedin CLI calls, or killing/restarting the daemon) can get the session logged out, forcing another 2FA round. Prefer the CLI commands, and don't restart the daemon mid-task unless it has hung.

Commands

Command	What it does	Verified
linkedin login [--code N] [--force] [--verify] [--wait-push N]	Log in / finish 2FA; saves session	✅
linkedin posts [-n N] [--full] [--include-reposts] [--json]	Fetch recent posts with engagement stats (default 10)	✅
linkedin post "<text>" [--yes]	Publish now (dry run unless --yes)	✅ type; publish via --yes
linkedin draft "<text>"	Save the text as a draft (empty composer only)	✅
linkedin drafts	Show the current draft (non-destructive)	✅

Scheduling is intentionally not supported: LinkedIn's schedule dialog (date field + calendar + time picker) was too brittle to automate reliably. Schedule manually in the browser when needed.

Fetching posts

linkedin posts                 # last 10, first line + 👍/💬/🔁 stats + URL
linkedin posts -n 20 --full    # 20 posts, full text
linkedin posts --json          # machine-readable

Reposts: bare reposts (reshares with no commentary of your own) are excluded by default — per Dan, a repost only counts when he added his own thoughts. Use --include-reposts to include them; reposts-with-commentary are always included and tagged [repost +commentary].

Posting / drafting

• post is dry-run by default — it types the content into the composer and shows you what it entered but does nothing irreversible. Add --yes to actually publish. Always show Dan the composed text and get the OK first.
• Opening the composer auto-loads the most recent draft. LinkedIn keeps a single server-side draft, and the editor cannot be cleared in place (see below), so post/draft refuse to type over a non-empty composer.
• Replacing an existing draft is not automated — the discard-then-resave chain proved too flaky and can lose the draft. To overwrite: open the draft in the browser, Discard it there, then run linkedin draft "<new text>" on the now-empty composer. (linkedin draft on an empty composer is reliable.)
• Trailing whitespace: pass text via "$(cat file)" — command substitution strips trailing newlines. Reading a file into a variable that keeps a trailing \n will save a draft with a stray blank line.

Writing in Dan's voice

When drafting any post, match Dan's established style and show him the draft for approval before publishing.

• Hook opener: a question or "Did you know…", often in bold-unicode (e.g. 𝗗𝗼 𝘆𝗼𝘂…). One short hook line, then a blank line.
• Voice: conversational but technical; explains clearly without dumbing down. "we" for Alexandra Instituttet research work; "I" for talks/personal actions.
• Structure: a few labelled sections, each led by a single emoji marker (📊 🧭 🔒 🌐 🔍 📈 📴), section title often in bold-unicode, then a sentence or two.
• Emojis: sparing and purposeful — section markers and the occasional reaction (🎉 😬 🤔). Not decorative spam.
• Links: trailing CTA lines with the 👉 arrow (Read more 👉 <link>). Write full URLs — LinkedIn auto-shortens to lnkd.in. Do not use markdown-style links ([text](url) doesn't work in LinkedIn posts) — paste raw URLs instead.
• Collaborators: thank coauthors/colleagues by name when relevant.
• Closing: a short list of lowercase hashtags, commonly #nlp #evaluation #llm #opensource #machinelearning #research.
• Posts in English even when the underlying material is Danish.
• Audience mostly already knows what LLM evals are — lead with what's new/novel, don't justify why evaluation matters. Recurring topic: EuroEval, multilingual LLM evaluation, hallucination/bias/values benchmarks.

Post length & engagement

LinkedIn allows up to 3,000 characters. Engagement patterns by length:

Length	Engagement Pattern
< 500 chars	Lower engagement
500–1,000 chars	Moderate engagement
1,000–1,500 chars	Good engagement
1,800–2,100 chars	Highest engagement
2,100–3,000 chars	Strong for deep content

Aim for 1,800–2,100 characters for optimal engagement, or 2,100–3,000 for technical deep-dives.

(This mirrors the linkedin-post-style memory; keep the two in sync.)

How it works / robustness notes

The composer is a Lexical contenteditable inside an iframe, which drives most of the gotchas below. The CLI reads the page with agent-browser snapshot -i (pierces frames) and acts via ref-based clicks; page-level eval/keyboard don't reach into the iframe. These behaviours were established the hard way — do not "simplify" them away:

• Enter text with type, not fill. agent-browser fill only mutates the DOM: it does not reliably flip Lexical's "dirty" flag (so the save prompt never appears) and can drop the caret mid-text, misplacing characters. type fires real per-char key events on the editor element and preserves order. type_into_editor uses type; keep it that way.
• The "Save this post as a draft?" prompt only appears when the composer has unsaved changes, and is triggered by pressing Escape on the focused editor. Do not click the composer's "Dismiss"/X control — LinkedIn's a11y tree exposes hidden decoy "Dismiss" buttons that ref-clicks silently no-op on.
• Never take a snapshot between entering text and pressing Escape — it loses the trigger. That's why cmd_draft/cmd_post log the input text rather than reading the editor back before closing.
• Choosing Discard in that prompt deletes the single draft entirely (it doesn't just revert changes).
• Elements are matched by accessible role + name with multiple candidates (English + Danish) and exact-vs-contains rules, so small LinkedIn relabels usually don't break it. If LinkedIn renames a control, add a candidate to the *_NAMES/BTN_* tuples near the top of linkedin_skill/main.py.
• "Start a post" is rendered as a link in the current layout (previously a button); open_composer accepts either role and polls for it, since the heavy feed takes a few seconds to render.
• Composer helpers poll and retry (open, save/discard, draft auto-load) because the live UI is slow and occasionally flaky. Sustained snapshots on the heavy feed can hang the agent-browser daemon — the polling keeps snapshot volume modest; if the daemon hangs, killing it usually also logs the session out.
• If a command reports a session problem, re-run linkedin login.