Reconnect

Help the user build useful professional relationships around their goals. Start from a topic, field, professional role, event, community, existing contacts, or roster. A prior relationship, course, employer, or private archive is not required.

Read network-plan.md when starting a new project or changing its audience. Resolve a private project directory for its preferences, sources, candidate records, batch reservations, and outcome ledger. Use the user's current context to fill known preferences rather than conducting a long intake.

Start with the requested mode

• Find: read discovery.md. Search and record candidates with source URLs, dates, supporting and conflicting evidence.
• Prepare: read tracking.md. Refresh source state, import candidates, and reserve a stable batch. Show unresolved identity or eligibility issues. An evidence score orders review; it is not a probability.
• Connect: read linkedin.md. Resolve the specific recipients covered by the user's instruction, check each live profile, perform permitted authorised actions, and verify the resulting state.
• Reconcile/resume: audit uncertain outcomes before retrying. Update the project's canonical contact tracker as well as the batch ledger. Never equate opening a profile, pressing a button, or a timeout with a successfully sent invitation.

The core accepts generic candidate JSON. Keep source-specific adapters and private project instructions in the user's project, outside the distributable skill. An adapter may read a permitted CRM export, address book, event list, or community database and emit this format; it must preserve provenance and existing exclusions.

Decisions that must survive sessions

Keep identity evidence, relevance to the user's goal, review priority, contact relationship, and group membership separate. A confirmed identity may be ineligible for this campaign. One profile can map to multiple archival identities: deduplicate by canonical destination and preserve all underlying source IDs.

Respect the actual scope of user approval. An explicit instruction to send an identified batch persists across interruptions; do not repeatedly ask for the same approval. If recipients or purpose change, resolve that change first. Do not infer outreach authority from approval to research, export, or install this skill.

Use existing browser skills and supported browser tools in the current harness. Keep authenticated LinkedIn work sequential in one tab. Do not invent a universally safe numerical sending limit, simulate human behaviour to evade detection, bypass challenges, or schedule an unrestricted outreach campaign. Stop on platform warnings, limits, restrictions, or identity challenges and report the concrete state.

Preserve each project's choice of invitation note, tone, and sender identity. If a note is requested, use a specific truthful reason to connect and the user's voice. Never imply a prior meeting, shared affiliation, endorsement, or familiarity without evidence. Never submit an old private email to satisfy an identity challenge without specific authorisation and evidence that it is appropriate.

Included helper

scripts/reconnect.py uses Python's standard library. It imports evidence records, deduplicates destinations, reserves batches transactionally, records observations, and exports reviewable JSON. It makes no network calls and does not operate LinkedIn. The agent performs research and permitted browser actions; the helper preserves their evidence and outcomes. See tracking.md for commands.

Discovery can use public websites and professional platforms. The bundled ledger currently tracks LinkedIn person-profile destinations; other channels require their own supported tools and project integration. Do not claim a universal social API.

Keep generated databases and batch outputs in the private project, never in an installed skill or its distributable snapshots. Do not commit, push, or publish personal data as a side effect of using this skill.