LinkedIn Power Audit

Runs on: local-exec - runs a local Python script against a local export folder. On a cloud surface I read the produced audit.json, I do not run the script.

The deep counterpart to linkedin-network-scan. The scan ranks who to act on; the audit reads the whole shape of your network and writes a structured audit.json the rest of the pack builds on. Python standard library only, no pip install, no AI in the extract, message content never read.

Known limitation: the export is title-only and point-in-time, has no location field, no firmographics, email mostly blank, and no engagement data on other people. The audit is a strong read of the network you have, not a database of strangers.

Step 1 - the export gate (do this FIRST)

Ask: "Do you have your LinkedIn data export downloaded and unzipped?"

This skill needs the UNZIPPED folder, not the .zip (unlike the scan, which takes the zip). If they have not requested it:

LinkedIn -> Settings -> Data privacy -> Get a copy of your data -> Download larger data archive -> Request archive.

LinkedIn delivers in two installments:

• Basic (within ~10 minutes): profile, connections, positions, education, skills only.
• Complete (~24 hours later): everything else - messages, posts, reactions, comments, invitations, follows.

The script detects which one it reads and prints a banner. Run on the Complete export for the full audit; re-run when it arrives if you only have the Basic one. If the zip unzipped to another zip (LinkedIn sometimes wraps twice), unzip again until you see Profile.csv and Connections.csv directly.

Step 2 - run the audit

python skills/linkedin-power-audit/power_audit.py <unzipped-export-folder> <output-folder> [--goal-buckets a,b]

• <unzipped-export-folder> is the folder with Profile.csv and Connections.csv in it (the script descends one level automatically if it is nested in a single subfolder, and resolves member-id-suffixed activity files like Shares_<id>.csv).
• <output-folder> MUST be outside any git repo - audit.json holds real names. A folder in their home directory is fine.
• --goal-buckets is optional. Pass the stakeholder buckets the user's goal needs (comma-separated) to add a network.gap read - see Step 4. Valid bucket keys are in taxonomy.json under stakeholder_priority (e.g. founder_owner_entrepreneur, tech_product_innovation, sales_partnerships_bd, hr_talent_recruitment, operations_general_management).
• --as-of YYYY-MM-DD is an internal determinism input. linkedin-start/run.py supplies today. Tests pass a fixed date.

If python is not found, try python3. The script refuses a .zip path with a clear unzip message.

Step 3 - read audit.json and give the read

Read audit.json from the output folder. It is structured and compact - safe to read (it does not dump 2,000 raw rows; it dumps rolled-up counts plus the top 50 message counterparties and top 30 companies and founders). Give the user the read that matters for their goal:

• Network composition: the role clusters and stakeholder buckets - what the network is weighted toward.
• Industry and company concentration: where their connections cluster.
• Founder pool: how many founders/owners, by use and by industry.
• Message warmth: how many hot / warm / dormant / outbound-only / light relationships, from metadata only. Every relationship label requires at least one inbound reply; outbound-only activity is never revival evidence.
• Content lanes: posting themes and follow themes - what they already play in (feeds linkedin-brand-direction).
• Export honesty: if _meta.export_type is basic or partial, say so - the audit is thinner than it looks and a re-run on the Complete export is worth it.

Frame it as a read of where they stand, not a file dump.

Step 4 - network-gap (over/under-representation)

When the user has a goal, pass --goal-buckets with the buckets that goal needs, and surface audit.json's network.gap:

• For each goal bucket: its share of the network and a status - under (thin, below 10%), balanced, or over (above 25%).
• The biggest_gap is the goal-relevant role the network is thinnest on.
• Every line traces to a real count ("operations leaders: 4 of 220 connections, 1.8% - under"). Do not assert a gap the counts do not show.

This is the "you sell to ops leaders but only 4% of your network holds an ops title" read. It is evidence-cited, never invented.

What it produces

• audit.json - the deterministic structured audit. The prerequisite for linkedin-warm-revival (reads the dormant counterparties) and linkedin-brand-direction (reads the composition and content lanes). Schema documented in audit.schema.json.

Privacy

• Message content is never read - only counterparty, direction, count, and date.
• The script makes no API calls and uses no scrapers. The raw export stays on your machine. What your Claude session reads from audit.json, real names included, goes to Anthropic under your plan like anything else the session reads.
• audit.json holds real names and should stay local - keep the output folder out of any repo.

Honest limitations (state these)

• Title-only and point-in-time. Seniority and role are inferred from one title string.
• No location field in the export. Region is never asserted from this data.
• Warmth sees export metadata only - a thread that moved to email looks cold here.
• The Basic export is thin. Half the audit needs the Complete installment.

Files

• power_audit.py - the deterministic extractor + classifier + network-gap (stdlib only).
• taxonomy.json - role / stakeholder / industry / founder-pool keyword taxonomy (generic, editable).
• theme-keywords.json - content and follow theme keywords (generic, editable).
• audit.schema.json - the audit.json contract downstream skills depend on.