Warm intro finder

Warm-intro and referral path finder

Cold is not the only way in. Given a target account, find the warm path hiding in the network. Given the whole network, find which accounts are already reachable.

Before anything else

Ask for the LinkedIn export if it is not already provided: Settings → Data privacy → Get a copy of your data → Connections. The file arrives as Connections.csv with a 3-line privacy notice above the real header. scripts/match.py handles that automatically.

Workflow

1. Profile the network first, always

python3 scripts/match.py <path-to-Connections.csv> --profile

Read the output before doing anything else, and tell the user what it means for them. Two numbers decide how much the rest is worth:

• with_email: usually near zero. LinkedIn only exports an email

when that person opted in. If it is low, say so, because it removes the strongest available relationship signal.

• connections_by_year: if the last 12 months dominate the file,

most of the network is inbound from content, not relationships. Warn the user plainly rather than scoring inbound follows as warm ties.

likely_own_employers is inferred from dense, long-lived clusters. Read it back to the user for confirmation, since it is a guess and it drives exclusions in reverse mode.

2. Run the mode that matches the question

Target mode when they named an account:

python3 scripts/match.py <csv> --target "Acme"

Reverse mode when they want to know where to spend the week:

python3 scripts/match.py <csv> --reverse --min-score 55 --limit 40

The two modes score differently on purpose. See references/scoring.md.

3. Apply the judgment pass

The script is deterministic and stops where data stops. Only after it returns, reason over the shortlist. This is the part that cannot be scripted:

• Would this person actually know the buyer? A BDR at a

4,000-person company does not know the CRO. A Chief of Staff at a 40-person company knows everyone. Weigh seniority against company size, which the script cannot see.

• Does the account fit the ICP? Reverse mode ranks by tie quality

alone. Deep ties often sit in the user's previous life, in an industry they no longer sell to. Filter for fit and say when a well-connected account is simply the wrong account.

• Is any path good enough? If not, say so and hand the account back

to cold outbound. Never manufacture a tenuous connection. A skill that refuses a bad intro is worth more than one that always finds something, because a bad intro-ask spends relationship capital that does not come back.

4. Draft the ask

Produce all three artifacts, never just the first. See references/templates.md for the patterns and the reasoning behind them.

1. The ask to the connector, with an explicit easy out.
2. The forwardable blurb, a separate short paragraph the connector

can paste without editing a word. This is what makes double opt-in intros work and it is the piece everyone skips.

1. The fallback, in case the connector says no or goes quiet.

Output format

Lead with the recommendation, not the data dump. For each recommended path: who, why them specifically, what to say, and what the risk is. Cap it at the top 3 paths per account. A ranked list of 15 names is a research artifact, not a decision.

Always state the two scores separately and never average them. Tie strength answers "will they say yes". Path relevance answers "is the intro useful". A strong tie into a useless contact and a weak tie into the right VP need opposite messages, and one blended number hides that.

Known limits, state them rather than working around them

• A LinkedIn export holds current employer only. Alumni paths and

second-degree bridges are not derivable from this file. If the user has an older export, diffing the two recovers job history, which is the single highest-leverage upgrade to this skill.

• No mutual-connection data, no message history, no interaction data.
• Company names are self-reported free text. Normalization handles legal

suffixes, parentheticals and emoji, but not everything.