LinkedIn Engagement Automation

A Python + Playwright automation that reads the LinkedIn home feed, likes the most interesting posts, and drafts human-sounding comments with an AI assistant.

• Level achieved: Level 2 (Read & React + Comment Thoughtfully)
• Actual time spent: ~1.5 hours

---

What it does

Level 1 — Read and React Reuses a warmed Chrome profile, reads the home feed, likes the top N interesting posts, and prints author + first 200 chars + outcome for each.

Level 2 — Comment Thoughtfully Picks the 2–3 most substantial of those posts and drafts an authentic, human-sounding comment for each with the Anthropic (Claude) API. Drafts are printed, never posted (per the assignment).

---

Requirements

• Python 3.10+
• Google Chrome installed (the automation drives your real Chrome; it falls back to Playwright's bundled Chromium if Chrome isn't found)
• A LinkedIn account
• (Level 2 only) an Anthropic API key

---

Setup

# 1. (recommended) create a virtual environment
python -m venv venv
venv\Scripts\Activate.ps1            # Windows PowerShell
# source venv/bin/activate           # macOS/Linux

# 2. install dependencies + the browser
pip install -r requirements.txt
playwright install chromium

# 3. one-time: log in and warm the Chrome profile
python save_login.py
#    -> a Chrome window opens. Log in to LinkedIn by hand (do any 2FA/captcha),
#       SCROLL the feed and open a couple of posts, then press Enter in the
#       terminal. Your session is saved in ./chrome-profile/ and reused.

# 4. (optional, for Level 2) configure the AI key
copy .env.example .env               # Windows   (cp on macOS/Linux)
#    -> edit .env and set ANTHROPIC_API_KEY=sk-ant-...

---

Run

# safe first run — reads the feed and drafts comments, but does NOT like anything
python main.py --dry-run

# real run — likes posts and drafts comments
python main.py --max-likes 5

# capture the run to a file for review (UTF-8; use --out, not PowerShell ">")
python main.py --max-likes 10 --out sample_output.txt

Flag	Default	Description
--max-likes	10	How many interesting posts to like
--comments	3	How many comments to draft (Level 2)
--dry-run	off	Read + draft, but don't actually like anything
--out FILE	—	Also write the run to FILE in UTF-8
--headless	off	Run without a visible window (less reliable here)

Expected output

LEVEL 1 — READ & REACT
[1] Jane Developer
    Spent the week migrating our background jobs from cron to a proper queue...
    -> liked
...
LEVEL 2 — DRAFT 3 THOUGHTFUL COMMENTS (not posted)
[3] Jane Developer  (picked: 1274 chars)
    Post: Spent the week migrating our background jobs...
    Draft comment: We hit the same thing — half our "split the service" debates
    evaporated once we could actually see where time went...
Done.

---

Troubleshooting

• "Feed didn't load (no posts found)" — LinkedIn is likely showing a checkpoint/captcha. Solve it in the open browser, or re-run python save_login.py and make sure real posts are visible before pressing Enter.
• Several likes show failed (TimeoutError) — LinkedIn throttles automation after a few quick likes. Lower --max-likes, and rerun later. This is expected; see Obstacles.
• AI unavailable (...) in a draft — a transient Anthropic API error (e.g. 529 Overloaded). The run still completes; rerun to get a real draft.
• Garbled non-ASCII in a saved file — you used PowerShell > (UTF-16). Use --out FILE instead, which writes UTF-8.

---

Design decisions

• Persistent real-Chrome profile, not automated login. save_login.py opens your real Chrome (channel="chrome") with a persistent on-disk profile (chrome-profile/). You log in and scroll once by hand; main.py reuses it. This avoids automating the login form (the most bot-watched step), stores no password, and is what makes LinkedIn actually render the feed (see Obstacles).
• Anti-automation launch flags. --disable-blink-features=AutomationControlled hides the main automation fingerprint; chromium_sandbox=True avoids the --no-sandbox banner. Headed by default (headless is easily detected).
• Class-independent selectors. LinkedIn ships the feed with hashed CSS class names and no data-urn, so we anchor on the accessibility attributes it must keep for screen readers: the "Open control menu for post by NAME" button (author), the "Reaction button state" button (Like), and the post card as the nearest ancestor of both.
• Body text parsing. The body has no stable element, so we read the card's full text and parse the body out of it (after the "time •" marker or "Promoted"), then strip footer noise.
• Resilience. Retries + graceful fallbacks on both liking (LinkedIn throttling) and comment drafting (transient API errors), so one flaky call never crashes a run.
• Human-ish pacing + ad filtering, and sync Playwright (a single sequential scrape — no concurrency to exploit).

---

Obstacles encountered

1. LinkedIn blocked the automation outright. A fresh Playwright Chromium loaded the feed (logged in, correct title) but served an empty shell — no posts — plus reCAPTCHA Enterprise and PerimeterX frames. LinkedIn had detected the automated browser and withheld the feed.
2. Fix: a warmed, persistent real-Chrome profile with anti-automation flags. Logging in and browsing once by hand got the feed to render.
3. Hashed CSS classes. Even rendering, every class is a random hash and there's no data-urn — an anti-scraping measure. Solved by anchoring on ARIA attributes instead.
4. No stable body element — the body is parsed out of the card's text.
5. Throttled likes. LinkedIn makes the Like button non-actionable after a few quick likes; mitigated with slow pacing + retries, but not 100% avoidable.
6. Likes are real — use --dry-run to exercise everything without touching LinkedIn.

---

Files

main.py            Orchestration: Level 1 + Level 2, CLI flags, UTF-8 output
save_login.py      One-time manual login + profile warm-up
linkedin/
  browser.py       Persistent real-Chrome launch config (the anti-bot core)
  models.py        Post dataclass
  feed.py          Feed reading, ARIA-based selectors, body parsing, liking
  comments.py      Anthropic-based comment drafting (with retry/fallback)
requirements.txt   Dependencies
.env.example       Config template (ANTHROPIC_API_KEY)