<!-- SYNCED COPY — do not edit here. Canonical: common-docs/skills/early-user-discovery/SKILL.md This file is distributed to every consuming repo by common-docs/meta/scripts/sync_skills.py. Edit the canonical, run the sync, and commit each repo. Edits made here are overwritten and lost. -->

Early user discovery

Human activity and an external relationship are separate facts. OAuth, a plausible name, nonzero spend, no admin flag and a solo organization never establish that someone is outside Arman's circle. Tahir was misclassified because the analyst skipped organization connections.

Resolve relationships before ranking

1. Read the owner's confirmed category, notes, contact history and restrictions first. Owner-confirmed employee, former employee, friend/family, test or real-user facts override guesses.
2. Check all organization memberships, roles, inviter and invitation history. Compare by org ID against the owner's organizations, not just company email domains or org names. Do not narrow this check to the currently selected organization.
3. Resolve the existing crm.party through claimed_by, then check CRM employment/affiliations and available HR links. Former membership or employment still explains a known connection.
4. Record the evidence and its source. Shared organization means connected / needs review, not automatically employee. An unread or empty relationship source means unknown; do not equate a failed lookup with no relationship. Never merge accounts from names alone.
5. Keep known people eligible for usage analysis while excluding them from automatic external discovery and unsolicited trial invitations. No bulk outreach from inferred categories.

Date and usage window

This program starts 2026-06-01T00:00:00Z, by Arman's September 30 instruction. Ignore earlier activity for selection, growth and offers. An old signup can qualify through later activity; an old signup or pre-June conversation alone cannot qualify. Preserve old relationship facts for exclusion. Label every request and cost window. Never call account counts people counts; deduplicate only from verified identity evidence.

Use the last seven days and the preceding seven days for changes, recent signups for discovery, and at least two separate active days for a sustained-use signal. Human requests, child-agent requests and workflow executions are different counts. Acquisition and page visits are evidence only when actually captured; do not invent them from a conversation title.

Recommend an allowance

Read live billing.plan and billing.plan_limit, not marketing constants. Compare usage inside actual billing periods, including AI points, storage and resource limits. All-time cost alone cannot prove a monthly tier. Distinguish measured plan fit from a provisional offer. A three-month paid-tier trial is not three months of an already free plan. Do not promise unlimited spend, guaranteed custom work, or silently grant entitlements. Owner approval names recipient, plan, duration and script.

Approved personal outreach only

• Read current owner notes and prior contact first. Preserve blocks and contact holds. Owner-reported contact is not permission to infer delivery IDs or contact channel.
• Draft email and DM separately. Begin in Arman's requested voice when appropriate: “Hi [first name]. This is Arman, not one of the agents. lol. I'm the founder of AI Matrx and the lead engineer.” Attribute it to him only through his approved sending identity.
• Personalize from coarse verified feature/page activity: “I noticed you tried our workspace.” Never quote private prompts, file contents, conversation titles, emails, personal topics or precise tracking timestamps. If page evidence is absent, use a general welcome.
• Approval binds exact recipient, channel, sender, offer and final text. Any edit revokes it. A template approval alone is not approval for every future recipient or personalization.
• Use existing Gmail and DM paths only after verifying the sending identity and replay guarantees. Respect canonical suppression, restrictions and stop requests. Store provider message/thread IDs and stable replay keys. After an ambiguous send, reconcile sent records before retrying; do not send blindly.
• No invitation to blocked, opted-out or known-internal accounts. A reply stops follow-ups. Do not schedule a reminder merely because the send returned successfully.
• Read only relevant reply threads, preserve inbound cursors and report new replies with their source link, brief meaning and next action. Do not change the approved offer yourself.

Daily run

Use one cheap execution model. Claim the day once, load owner facts, reconcile relationships, collect bounded new-signup/usage evidence and draft only new candidates. Saved drafts reserve the account; they are not approvals. Read replies only from verified existing outreach threads. Sending remains a separate approved action through the existing app, with a recorded receipt. Do not treat draft JSON as an executable queue or claim an unattended sender is implemented. Keep quiet when nothing actionable changes. Report new qualified candidates, replies, failed sends and uncertain delivery. Never claim a scheduled task is active until the scheduler confirms it. Schedule activation needs the owner's approval of the script and proposed time.

Proof scenarios and outcomes: evals.md.