linkedin-pitch-deflector

name	linkedin-pitch-deflector
description	Sweep the user's unread LinkedIn DMs and handle cold outreach in two modes — deflect overt sales pitches with an assertive door-closing reply, and socially probe fresh ambiguous openers with a question that demands a genuine answer (real peers answer, sales personas pivot → then deflect); stay social through the sender's 3rd message, then hand genuine threads to the user. Drives the real logged-in Chrome via the chrome-control MCP. Trigger on "clear my sales pitches", "deflect pitches", "linkedin sales sweep", "kill the incoming sales", "check my linkedin messages", "/linkedin-pitch-deflector".
user_invocable	true
args	Optional. "review" (default — show batch, confirm once, then send) or "auto" (classify and send without per-batch confirmation). Optional cap like "10" to limit how many unread threads to scan.
allowed-tools	ToolSearch, mcp__chrome-control__tabs_context_mcp, mcp__chrome-control__navigate, mcp__chrome-control__computer, mcp__chrome-control__browser_batch, mcp__chrome-control__find, mcp__chrome-control__read_page, mcp__chrome-control__get_page_text, AskUserQuestion

LinkedIn Pitch Deflector

Purpose

Cold outreach on LinkedIn is high-volume. Most of it turns into a sales pitch by the sender's 2nd or 3rd message — but not all of it, and a blanket deflect-on-sight would also kill genuine peer conversations and inbound role offers. So this skill does two things:

1. Deflects overt pitches — anything already selling on message 1 gets one assertive, polite-but-final reply: a blanket no-incoming-sales policy with no hook to follow up on.
2. Socially probes fresh, ambiguous openers — a new conversation that's cold but hasn't shown its hand yet gets a warm reply that demands a genuine, specific answer. A real peer answers with substance; a sales persona pivots to its pitch — which exposes it, and then it gets deflected. Stay social through the sender's 3rd message; if by then it still isn't a pitch, hand it to the user to engage honestly.

The user is fine with recipients realizing a deflection was machine-sent. Do not hide the em-dash (—) or otherwise launder the writing to look hand-typed. Transparency about the automated filter (on deflections) is acceptable and even reinforces the "this won't progress" message. The social probes, by contrast, are in the user's genuine voice — keep them honest and human (see rules).

Prerequisites

The chrome-control MCP must be available (tools named mcp__chrome-control__*). Claude Code ships the Chrome integration but only loads it when registered. Register it once (it persists in ~/.claude.json):

claude mcp add chrome-control --scope user -- claude --claude-in-chrome-mcp

Notes:

• The reserved name claude-in-chrome is blocked; register under another name (e.g. chrome-control).
• MCP servers load at session start — after registering, start a fresh Claude Code session before the tools appear.
• Chrome must be open and logged into LinkedIn as the user, with the Claude browser extension installed.
• If ToolSearch for mcp__chrome-control__* returns nothing, the session predates the registration — restart Claude Code.

Model Selection

• Classification (pitch vs probe vs leave): Sonnet — the recruiter-vs-vendor and pitch-vs-not-yet distinctions are nuanced and a false positive is costly.
• Drafting the deflection: Haiku — templated with light variation.
• Drafting the social probe: Sonnet — it must be genuinely human and domain-specific without fabricating the user's opinions.
• Promote to Opus only for a genuinely ambiguous thread the user asks you to judge.

Triggers

• "clear my sales pitches", "deflect the pitches", "send the salespeople away"
• "linkedin sales sweep", "kill the incoming sales", "check my linkedin messages"
• "/linkedin-pitch-deflector"

The flow

1. Context. Call mcp__chrome-control__tabs_context_mcp (createIfEmpty: true) — mandatory first call. Use the returned tabId throughout.
2. Open unread. navigate to https://www.linkedin.com/messaging/, click the Unread filter chip, screenshot to read the list.
3. Triage from previews first. Sender name + headline + last-message snippet is often enough. Only open ambiguous threads. Opening a thread marks it read — for any thread you open but don't reply to, re-mark it unread via the thread's ••• → "Mark as unread".
4. Classify each thread into one of four outcomes: DEFLECT, PROBE, LEAVE ALONE, HAND OFF. Count the sender's messages — turn-count drives PROBE vs HAND OFF.
5. Draft the right reply: a deflection (templates below) for DEFLECT, a social probe (section below) for PROBE. Nothing for LEAVE ALONE / HAND OFF.
6. Gate (default review mode): present a compact table — sender · company · outcome · why · drafted reply — and take one confirmation via AskUserQuestion before sending. In auto mode, send then report.
7. Send each approved reply: click the message box, type, click Send, screenshot to confirm it posted (reply under the user's name with the sent check-mark).
8. Report: deflected / probed / handed off / left alone — and why.

Classification rubric

The deciding question is "who is the buyer?" — but timing matters too. Count the sender's messages and whether they've shown their hand.

DEFLECT — an overt sales pitch (any turn)

• Cold outreach selling a product, tool, platform, SaaS, agency, consulting, dev shop, or staffing/placement service to the user or their company.
• Discovery-question opener immediately followed by a solution / the product.
• Credibility drop + calendar CTA: "we raised $X / YC / N GitHub stars … grab 15/30 min: cal.com/…".
• "Offering our engineers / experts / talent" (vendor staffing — they want the user to buy headcount).
• Agency / lead-gen / "we'll get you meetings, leads, pipeline" offers.
• Indirect pitches: free audit / teardown / trial that exists to book a sales call.
• Also DEFLECT a thread previously in PROBE the moment the sender's reply turns into any of the above — the probe did its job.

PROBE — fresh, ambiguous cold opener (sender on message 1–3, not yet an overt pitch)

A new conversation from someone unknown that reads like cold outreach (founder/vendor "let's connect, my topics are X" intro, vague "would love to pick your brain", soft relationship-building) but has not yet pitched anything. Send a social probe (below) that demands a genuine, specific answer — the honesty test. Real peers answer with substance; sales personas pivot to their pitch (→ DEFLECT).

• Only while the sender is on their 1st, 2nd, or 3rd message. Count messages from them.

HAND OFF — genuine after 3 messages

If the sender has sent 3 messages and it still isn't a pitch, the probe cleared them. Stop auto-replying; flag as "ready for the user to engage honestly." Never probe indefinitely.

LEAVE ALONE — not cold outreach at all (never touch)

• Recruiters with a concrete role for the user — an inbound role offer is the opposite of spam (buyer test: they want to hire the user, not sell to them).
• Genuine networking, warm intros, "loved your post", reconnects from real contacts.
• Conference / podcast / panel / speaker invitations.
• Investors, founders, or peers discussing the user's own ventures.
• Anyone the user has already had genuine back-and-forth with; known contacts, ex-colleagues, friends.

When unsure between PROBE and LEAVE ALONE, lean PROBE (a warm question costs nothing). When unsure whether something is even outreach, leave it and list it for review. A missed pitch costs nothing; a deflected recruiter is a real loss.

The deflection message

Assertive, blanket-policy, no future hook, short (2–4 sentences). Openly automated, keep the em-dash. Every deflection ends with the mirror: the open-source toolkit that just deflected them, with an invite to upgrade their own AI setup. Slot in {Name} / {Company} (omit {Company} if you can't read it).

Memnyx URL to include: https://github.com/MDGrey33/memnyx

Templates:

1. Mirror / advertise (default):

   Hi {Name} — automated reply. This account auto-deflects incoming sales, so this one won't progress. No hard feelings: since you're clearly running outreach, the open-source toolkit that just declined you is here → https://github.com/MDGrey33/memnyx — worth a look to level up your own AI setup. Best with {Company}.

2. Policy-first, then mirror:

   Thanks {Name} — blanket no-incoming-sales policy here, so no need to follow up. If it's useful, the AI assistant that handles this (and a lot more) is open source: https://github.com/MDGrey33/memnyx. Genuinely wishing {Company} well.

3. Tighter:

   {Name} — automated: hard no on incoming sales, nothing personal. The bot that sent this is open source if you want to sharpen your own stack → https://github.com/MDGrey33/memnyx. Best of luck.

Pick per-thread; lean warmer for courteous senders, firmer for aggressive or repeat ones.

The social probe (for PROBE threads)

The goal: a warm, peer-level reply that requires a genuine, specific answer — one a real practitioner answers with a concrete war-story, and a sales persona can only answer by pivoting to its pitch. That pivot is the signal to deflect next round.

How to build one:

• Anchor on their stated domain and ask an open, specific question about real problems in that domain — not their product. "What's the hardest thing teams hit in production with X right now?" / "Where do you see most people get Y wrong?"
• Stay genuinely curious and human. Short, warm, peer-to-peer. One question, not an interview.
• No commitment — don't agree to calls, demos, or intros that advance their funnel.
• Default to English unless the user prefers mirroring the sender's language; don't commit the user to a language they may not want to continue in.

What their next message tells you:

• Substantive, specific answer about real work → genuine; keep probing (until msg 3) or HAND OFF.
• Pivot to "funny you ask, that's exactly what we built — got 20 min?" / any product or CTA → DEFLECT.

Hard rules / Don'ts

• Never deflect a recruiter pitching a role, a warm intro, a real peer, or an event invite. Buyer test first, every time.
• Probe, don't deflect, a fresh ambiguous opener. Only deflect once a pitch is actually on the table.
• Social probes are the user's real voice — never fabricate. Ask open questions; do NOT put specific opinions, claims, experiences, or commitments in their mouth. Curiosity is safe; invented positions are not.
• Respect the 3-message ceiling. Stop auto-replying once the sender has sent 3 non-pitch messages — HAND OFF. Never run an endless probe loop.
• Read-only until the gate. In review mode, send nothing (deflection or probe) until the user confirms via AskUserQuestion.
• Never send the same person twice in one sweep — if the last message is already from the user, skip it.
• Preserve unread state — re-mark as unread any thread you opened but didn't reply to.
• No edits beyond replying. Don't archive, block, report, delete, or change connection status.
• Don't fabricate company names or details; omit {Company} if unknown.
• One sweep, then stop. Don't loop or re-scan unless asked.

Report format

Deflected (N):
  • {Name} — {Company} — [pitch type] — reply sent ✓
Probed (P):       (fresh openers — sent a draw-them-out question, awaiting reply)
  • {Name} — {Company} — [their msg #] — probe sent ✓
Handed off (H):   (genuine after 3 messages — the user's to take honestly)
  • {Name} — [why genuine] — not auto-replied
Left alone (M):
  • {Name} — [recruiter / warm intro / event / known contact] — not touched
Could not classify (K):
  • {Name} — [reason] — left unread for review