Auto-comment on ICP Feed

The premise: commenting thoughtfully on a prospect's post before sending an invite raises accept rate by 30-50%. They've seen your name and read something useful from you. This skill scrolls your LinkedIn feed, identifies posts authored by ICP-matching profiles, and posts a value-adding comment on each.

Required prerequisites — check before you start

1. `linkupapi` MCP connected. Verify with linkupapi_list_accounts.

• If the list is empty, tell the user they have two options to connect a LinkedIn account, then stop and wait:
• Hosted UI — open https://app.linkupapi.com/account-connection (fastest, handles checkpoints in-browser).
• MCP login — run linkupapi_login directly (platform=linkedin, with email+password OR a login_token). On checkpoint_required → run linkupapi_checkpoint.

1. Pick the sending account before Stage 1. After confirming at least one status = connected account exists, present the connected accounts via AskUserQuestion (single-select). Each option label is the account display name; description shows email + country. The chosen account_id is the one whose feed will be scrolled and that will post the comments.
2. The user must already know their ICP — this skill is NOT a discovery tool, it filters an existing feed against a known ICP.

Daily LinkedIn safety caps (MANDATORY — enforced)

• 100 profile gets / day — used for borderline ICP-match verification
• 15 searches / day — generally not consumed by this skill (feed ≠ search)
• Comment volume: LinkedIn soft-caps around 30-50 comments/day. This skill defaults to 15 comments / day max to stay safely below.

Before Stage 1, run linkupapi_get_logs for last 24h on the chosen account, count today's linkedin_profiles/get and linkedin_content/comment (or whichever action posts a comment — confirm via tool schema). Compute remaining budget per category. Never override.

Stage 0 — ICP + comment style

AskUserQuestion:

1. ICP — same shape as linkedin-outreach Stage 0. Offer to reuse ./icp/*.json if present.
2. Topic filter — optional keywords the post body must contain (e.g. "outbound", "hiring", "AI sales"). Empty = no topic constraint, just author match.
3. Comment style:

• Auto-generate (recommended) — agent reads each post + writes a 2-3 sentence value-adding comment in real time
• Templates with variables — user provides 3-5 templates with {post_topic} / {author_first_name} placeholders, agent picks one per post
• Manual approval (HITL) — agent drafts each, user approves/edits/skips before posting

1. Volume cap — max comments today (hard ceiling = 15)
2. Time window — only posts authored in the last X hours (default 48h; older posts have low engagement ROI)

Echo the brief with budget math + the daily cap remaining, wait for "yes".

Stage 1 — Pull the feed

Tool: linkedin_content. Load schema with ToolSearch query="select:mcp__linkupapi__linkedin_content". Likely action: get_feed or get_user_feed.

{"account_id": "...", "action": "get_feed", "params": {"limit": 100}}

Pull 50-100 recent posts. Each result should include post_url, author_profile_url, author_name, author_headline, post_text (or excerpt), posted_at, engagement_count.

Drop immediately:

• Posts older than the time window
• Reposts where the user authored neither the original nor the repost commentary
• Promoted/sponsored posts (they're ads, not organic)
• Posts the user has already commented on (check viewer_has_commented if available)

Stage 2 — Pre-filter on visible author signal

For each post, score the AUTHOR against ICP using only visible fields (author_headline, author_name):

• ✅ Headline matches ICP roles/industry → KEEP
• ❌ Clear mismatch → DROP
• ⚠️ Ambiguous → keep for Stage 3 enrichment IF budget permits

Then apply the topic filter on post_text:

• KEEP if post_text contains any topic keyword (case-insensitive substring or semantic match)
• If user provided no topic filter, skip this check

No tool calls at this stage — pure LLM judgment.

Stage 3 — Enrich borderline authors (budget-aware)

For ⚠️ ambiguous authors only, run linkedin_profiles get to verify current role.

Cap this stage at 1/3 of today's remaining profile budget. Never exhaust the 100/day on Stage 3 — leave room for the user's other workflows (linkedin-outreach, linkedin-enrich).

Drop authors whose enriched current role doesn't match the ICP.

Stage 4 — Generate comments

For each kept post, generate a comment per the chosen style.

Auto-generated comment principles:

• 2-3 sentences max (≤300 chars ideal)
• Add value — don't just agree. Worst possible comment: "Great post!"
• Reference a specific point from post_text (proves you read it, not just scrolled)
• End with an open question or insight, not a CTA / pitch
• No emojis unless the post itself uses them
• No links (LinkedIn ranks comments with links lower)
• No name-drop of the user's company unless directly relevant

Example shape:

"The point about {specific_thing_from_post} resonates — we've seen the same pattern with {related_observation}. Curious whether {open_question}?"

If user picked HITL, surface each draft with post_excerpt + author + comment_draft and wait for approve / edit / skip per post. Don't batch the approvals.

Stage 5 — Post comments (paced, cap-aware)

Use linkedin_content action that posts a comment (comment / post_comment — confirm via schema).

{"account_id": "...", "action": "comment", "params": {"post_url": "...", "comment_text": "..."}}

Pacing rules (non-negotiable):

• Sleep 30-60 seconds between comments. Comment cadence is the most-watched anti-bot signal — 15 comments in 5 minutes will get flagged.
• Vary the sleep duration randomly within that range.
• Stop immediately if today's 15-comment cap is reached.
• Stop immediately if any call returns a rate-limit error or 429.

If the run starts producing errors mid-batch, stop and surface to the user — don't retry blindly.

Stage 6 — Persist & report (mandatory)

Write ./campaigns/{YYYY-MM-DD}-feed-engage.md:

1. ICP definition + topic filter
2. Feed scan stats: posts pulled / matched author / matched topic / commented
3. Comment style used
4. Per-comment table: # | post_url | author | author_company | comment_posted | timestamp
5. Daily caps remaining
6. Skipped posts and why (already commented / topic mismatch / borderline dropped for budget)

Concise on-screen summary.

Pre-handoff to outreach

When linkedin-outreach runs later this week and the prospect list overlaps with authors commented on this week, the outreach skill should reference the comment in the invite note — accept rate jumps significantly. The campaign log here is the source of truth for that cross-reference. Make sure author_profile_url is captured in the log so the outreach skill can match.

Common pitfalls

• Comment cadence too fast: 15 in an hour = flag. Default 30-60s + random jitter.
• Generic praise comments: "Great post!" / "Love this!" — kill credibility AND get demoted in feed ranking. Always reference something specific.
• Templated phrasing: LinkedIn fingerprints repeated phrases across an account. Vary every comment, don't reuse opener words.
• Controversial content: skip posts about politics, layoffs, personal news unless the user explicitly opts in.
• Old posts: comments on posts >72h old get low visibility. Filter to last 48h by default.
• Author replies expected: when the author replies to your comment, the user should engage back manually within 24h (no automation here — too risky).

Tool quick reference