ai-prospecting

The atomic skill for OnFire’s Phoenix prospecting engine. Either ranks prospects at a target company (action="run") or scores one named person within a company context (action="get_prospect").

When the agent picks this skill

• “Find buyers at Acme.” / “Who should I talk to at this company?”
• “Score Jane Doe at Acme — is she worth a meeting?”
• A batch run across many companies.
• Testing a draft schema with activate_shadow_run=True (handed off from manage-ai-prospecting).

What the skill teaches the agent

• Both action paths — when to use run vs. get_prospect.
• The dataset-backed response shape: top_picks, priority_summary, preview_rows, facets, field_index, and the field_glossary_resource_uri. Lead with priority_summary as a one-line lede, then render top_picks. Don’t flat-dump preview_rows — the full list lives in the dataset.
• The polling pattern for long runs: on a still_running retry, prefer passing run_ids back instead of re-POSTing.
• The $$-onfire-bold-$$ sentinel conversion in ai_reasoning.
• When to call ai_prospecting_field_glossary (once per conversation, on clients that don’t auto-load MCP resources).

Underlying MCP tools

• ai_prospecting — the main call.
• ai_prospecting_field_glossary — the field contract.
• Dataset companions: describe_dataset, query_datasets, download_dataset, list_datasets.