Useful agentic network prompts begin with a real objective and ask for evidence, not invented familiarity. Tell the assistant which people or companies matter, what sources it may use and what a useful output should contain.
Use these original prompts in Claude or Codex with information you are authorized to share. If Finta MCP is connected, first confirm the relevant tools are available. Without that connection, use a small supplied dataset and ask the assistant not to imply it searched your private network.
1. Find relevant experience, not just matching titles
My goal is [specific decision]. I need to learn from people with experience in [role, sector or situation]. From [authorized sources or supplied records], identify plausible matches. For each, show the source, the evidence of relevant experience and what needs verification. Do not assume a job title proves responsibility. Return no more than five candidates. Do not contact anyone or change records.
Output check: each candidate needs evidence tied to the requested experience. A matching company name alone is insufficient. If none of the reviewed records supports the role, a useful answer says so.
2. Resolve an ambiguous person before making an ask
These records may refer to the same person: [authorized records]. Compare the available name, company, role, source and stable IDs. Separate facts that match from conflicts or missing fields. Recommend what I should verify before merging or using the records. Do not infer private contact details, merge records or choose an identity just because the name matches.
Output check: ambiguity should remain visible. “Alex at Cedar” is not enough if the records show two different people or time periods.
3. Turn a vague introduction request into a useful brief
Improve this introduction request: [draft]. Use only these facts: [verified purpose, target role, problem, relationship context and requested conversation]. Return a short brief that explains why the conversation could help both sides, names the smallest reasonable ask and offers an easy decline. Flag unsupported claims. Do not imply the connector has agreed or that the recipient is interested. Draft only.
Output check: the note should help the connector decide, not pressure them to endorse you. It should not invent traction, friendship or urgency.
4. Find an unfinished commitment
Review [authorized conversation notes or supported records] for commitments involving [person or project]. Return who promised what, the supporting source, any actual due date and what remains unresolved. Distinguish a proposal from an agreement. Suggest one next action for my review. Do not send reminders, create tasks or update the CRM.
Output check: the assistant must cite the promise. “We could meet next month” is not the same as a scheduled meeting.
5. Audit a network gap around one decision
For [specific business decision], these perspectives would be useful: [roles or experience]. Review only [authorized sources]. Map supported matches to each perspective. Label uncertain roles, stale context and not-returned results. Do not infer sensitive personal attributes or claim the search covers my entire network. Suggest one appropriate way to learn from a missing perspective. Do not contact anyone.
Output check: an incomplete search should not become a claim that a relationship does not exist.
A small fixture for checking the prompts
Fictional evidence supplied by the reader
| Record | What it supports | What it does not support |
|---|---|---|
| Mina, former operations lead; role last confirmed two years ago | Relevant historical experience | Current buying authority |
| Two Alex records at different companies | An identity question | A safe automatic merge |
| A note says “I will send the brief on Friday” | An explicit commitment | Proof that the brief was sent |
Expected answers should preserve these distinctions. The fixture is an illustrative editorial check, not a transcript of a live Claude or Codex tool session.
Bring the source behind the answer
Finta MCP can connect supported tools to an authorized client. Use the tools reference and permission boundary, rather than assuming a prompt creates access. Setup remains in the Claude guide and Codex guide.
For the human side of the work, use the introduction playbook. A strong prompt makes the evidence and next decision easier to inspect. It does not manufacture trust.
