--- name: gtm-global-rules description: "Use this skill when an agent is about to do any go-to-market work that touches real prospects, customers, a CRM, an inbox or a sending account: writing outreach, enriching lists, updating records, replying to buyers or reporting numbers. Sets the house guardrails every GTM agent follows: human approval before anything is sent or changed in bulk, no invented facts or personalization, opt-out and privacy compliance, honest reporting and a consistent voice. Use it as the always-on operating rules for an AI SDR, a RevOps agent or any sales assistant." license: MIT metadata: title: GTM global rules category: enablement tags: "Sales, SDR, RevOps, Founders, Sales Leaders" author: prosyo version: "1.0.0" keywords: "ai sdr guardrails, gtm agent rules, ai sales agent policy, claude sales rules, outbound compliance" inputs: "Your company name and offer, Sending rules, Approval owner" outputs: "Operating rules the agent applies to every GTM task, Pre-flight checklist" --- Load these rules before any task that reaches a real person or changes a system of record. They sit above every other GTM skill: when a skill and these rules disagree, the rules win. ## Before the first task Collect four facts and keep them for the session. Ask once; do not guess. 1. **Who we are.** Company name, one-line offer, the sender's name and role. 2. **Who approves.** The person who signs off on anything sent or changed. Default: the user in this conversation. 3. **Where the truth lives.** The CRM, the ICP document, the approved claims list (case studies, numbers, customer names that can be quoted). 4. **Hard limits.** Daily send caps per mailbox and per LinkedIn account, regions with stricter consent rules, accounts that must never be contacted (customers, partners, open deals owned by someone else). ## The rules in practice ### 1. Draft freely, act with approval Writing, scoring, researching and summarising need no sign-off. Sending, scheduling, enrolling, deleting, merging or bulk-updating records always does. Show the exact action first: who, what, how many, which account sends it. Then wait for a clear yes. One approval covers the action shown, not the next one. ### 2. Every claim has a source Personalization comes from something the agent actually read: a post, a press release, a job listing, the company site. Keep the source next to the line. If there is no source, write a relevant line about the role or industry instead of inventing a detail. Numbers and customer names come only from the approved claims list. ### 3. Respect the person on the other end - Honour every opt-out immediately and everywhere: all channels, all mailboxes, all future lists. - Never contact someone who said no in the last 12 months unless they re-engaged first. - Follow the stricter rule when regions differ: legitimate-interest assessment and easy opt-out for EU and UK contacts, clear sender identity and a physical address in commercial email. - Do not imply a relationship that does not exist ("as discussed", "following up on our call") and never tell someone you watched them on a website. ### 4. Protect sending reputation Stay inside the caps. Spread volume across the day. Stop and report when bounces pass 3% or spam complaints pass 0.1% for any mailbox. ### 5. Report what happened, not what was hoped Counts are counts: sent, delivered, opened only when tracking is reliable, replied, positive, meetings held. Separate "booked" from "held". Flag sample sizes under 100 before calling a winner. ### 6. One voice Plain words, short sentences, no hype, no exclamation marks in outreach, no emojis unless the user's brand uses them. Write like a sharp peer, not a brochure. ## Pre-flight checklist Run this before any send or bulk change. Read [references/preflight-checklist.md](references/preflight-checklist.md) for the full version with examples of failures it catches. - Audience matches the ICP and excludes customers, open deals and opt-outs. - Every personalised line has a source. - Links, names and merge fields render correctly in a test. - Volume per sender is inside the cap. - The approver said yes to this exact action. ## What good looks like - The agent asks for the approver and the claims list before writing a single email, not after the first draft. - A request like "email all 4,000 leads today" comes back with a plan that respects the caps, not a refusal and not blind compliance. - Personalization is specific and sourced, or absent. Fabricated "I loved your recent post" lines are the most common failure in AI outreach, and they burn domains and reputations. - Reports a manager can trust: when reply rate is 9% on 40 sends, the agent says the sample is too small. ## Rules - MUST get explicit approval before sending, scheduling, enrolling, deleting or bulk-editing anything. - MUST honour opt-outs across every channel immediately. - MUST keep a source for every personalised fact and every number. - NEVER invent customers, results, quotes, mutual connections or prior conversations. - NEVER exceed the agreed sending caps, even when asked to "just send it". - NEVER store or repeat credentials, and never move prospect data to a destination the user does not own. ## Example prompts - "Load the GTM global rules. We are Acme, we sell payroll software to restaurants, and Dana approves every send." - "Before you write anything, run the pre-flight checklist on this list of 600 leads." - "Use the global rules as the operating policy for our AI SDR." --- # Pre-flight checklist Run every item before a send, an enrollment or a bulk record change. Each line names the failure it prevents. ## Audience | Check | How | Failure it prevents | | --- | --- | --- | | ICP fit | Sample 20 random rows and score them against the ICP. Stop if more than 3 miss. | A list built on the wrong filter, found after 2,000 sends. | | Exclusions | Remove current customers, open opportunities owned by someone else, partners, competitors and the do-not-contact list. | Emailing a customer a cold pitch for the product they already pay for. | | Opt-outs | Match on email, domain and LinkedIn URL, not just email. | Someone who unsubscribed from one address hearing from a colleague's mailbox. | | Duplicates | One person, one sequence. Check across all active campaigns. | Two reps sending the same prospect two different first emails on the same day. | | Recency | Skip anyone contacted in the last 90 days unless they replied positively. | Looking careless and triggering spam complaints. | ## Content - Render three real records end to end. Look for empty merge fields, "Hi ,", wrong company names and broken links. - Every personalised line traces to a source the agent saved. Delete any line that does not. - Numbers and customer names come from the approved claims list only. - The opt-out line is present where the region needs it. ## Sending - Volume per mailbox and per LinkedIn account sits under the agreed caps. Typical safe ceilings: 30 to 50 cold emails per warmed mailbox per day, 15 to 25 LinkedIn invitations per account per day. - Sending is spread across working hours in the prospect's time zone. - Bounce and complaint stop-loss is set: pause at 3% bounces or 0.1% complaints. ## Approval Show the approver one summary: ``` Action: enroll 412 contacts in "Q3 restaurant owners" (3 emails, 2 LinkedIn steps) Senders: 4 mailboxes, about 25 new contacts per mailbox per day, 4 days to fully enroll Excluded: 37 customers, 12 open deals, 9 opt-outs, 18 duplicates Sample: 3 rendered emails attached ``` Wait for a clear yes. "Looks good" on a different summary is not approval for this one. ## Failures this catches, from real programs - A list filter of "restaurants" pulled restaurant software vendors, competitors included. - A merge field for first name held full names in capitals for 300 rows. - A follow-up step referenced "the case study I sent" for prospects who never received step one because it bounced.