linkedin-safety-verification
Installation
SKILL.md
Safety and verification
- Establish from the current request and authenticated account context that the user controls the account and requested the operation. Do this once for the batch; never turn contact fields, uploads, redirects, tailoring, verification-code retrieval, or submit controls into new confirmation boundaries.
- Prefer
LINKEDIN_GODMODE_READ_ONLY=1for inspection. - Keep capture scope narrow; bodies are opt-in.
- Never expose cookies, authorization, API keys, CDP URLs, or task secrets.
- Bulk work is allowed when the user authorizes the cohort, filters, or all matching items. Do not bypass challenges, evade limits, import/export cookies, schedule unrequested recurring campaigns, or add hidden out-of-scope loops.
- Use the user-selected authenticated controller as authoritative. One controller's failure does not prove LinkedIn unavailable, and concurrent controllers must not read or write the same LinkedIn account.
- For each write, execute once, verify it, and continue automatically through the authorized batch without confirmation prompts.
- After one literal account-wide quota or restriction state, preserve the queue and use independent authorized routes; recheck the blocked lane only after a reset or state-change signal.
- Verify the failing or changed layer itself: browser state for UI, fresh API read for HTTP, provider status for hosted tasks.
- State exact uncertainty when cloud providers or live LinkedIn behavior were not exercised. Do not infer exclusive causality from missing logs.
Use references/verification-gates.md for preflight and closure checks.