The Follow-Up Agent: a follow-up written for everyone you meet

Use the Autoploy MCP for every read and save in this task. If it is not connected, stop and tell me instead of improvising. What to build: The Follow-Up Agent: a follow-up written for everyone you meet. When a call is recorded, or you simply say who you just met, write the follow-up note in your saved voice, file the person in your Networking contacts list, and name anyone left waiting. Nothing is sent for you. Runs on your Claude subscription, or on your own AI key. Autoploy never bills you for the AI. This is for founders and non-technical builders. How it behaves: 1. Runs when your notetaker posts a finished call, or when you hand over someone you just met. 2. Read the finished call your notetaker sent, or the details you handed over, and work out who the person is. 3. If enough here to write something real, then go to "Check your contacts list", otherwise go to "Say so honestly". 4. There was not enough to work with, so it says so instead of inventing something. 5. Look in your Networking contacts list for a note already written for this person. 6. If already written to them, then go to "Leave it alone", otherwise go to "Read your business context". 7. A note for this person already exists, so nothing is written twice. 8. Read the saved business fields this automation is locked to: your voice, offer, audience, method, proof, guardrails, and links. 9. If enough context to sound like you, then go to "Read your follow-up rules", otherwise go to "Note the gaps". 10. Note which saved fields are empty or thin, so the note arrives with an honest line instead of pretending. 11. Read the rules in your Follow-up style list: your tone, how to open, what to always include, what never to say, and how to close. Any rule you add is obeyed too. 12. Write a short follow-up in your voice: it names where you met, picks up the one thing that mattered most to them, and offers one concrete thing you can do, never a favour you did not offer. 13. Add one row to the Networking contacts list, keyed by the person, so a replayed call never writes a second note. 14. Read everyone in your list and work out who has been sitting there longer than the number of days you set. 15. If anyone left waiting, then go to "Name who is waiting", otherwise go to "Deliver the note". 16. Name the people who have been waiting longest, so nobody quietly goes cold. 17. Deliver the follow-up note to you, ready for you to send yourself. The tools: - Anthropic (for "What should read the call and write the follow-up?"). I will need you to create the key ANTHROPIC_API_KEY (Anthropic API key). I add the values in Autoploy myself; you only need their names. - Slack (for "Where should the follow-up note land, ready for you to send?"). I will need you to create the key SLACK_BOT_TOKEN (Slack bot token), provide the setting SLACK_CHANNEL (Slack channel). I add the values in Autoploy myself; you only need their names. Shared memory: This automation remembers its work in a list called "Networking contacts" (Everyone you have met and the follow-up note written for them, one row each.) that I can open in the Context section of my dashboard and see every row. Its columns are: who ("Who"), met_where ("Where you met"), what_they_do ("What they do"), looking_for ("What they are looking for"), personal_note ("Worth remembering"), follow_up ("The follow-up note"), status ("Status"), added_on ("Added on"), contact_key ("Contact key"). Before anything runs, make sure the list exists: check with list_collections over my Autoploy MCP, and create it with create_collection (exactly that name, that purpose, and those columns) only if it is missing. Never create a second list under a similar name; reuse this one. The platform never creates a list silently at run time, so a missing list stops the work with a plain refusal. In the automation's code, write rows through ctx.collections.insert with an idempotency key built from the id of the thing the row is about (for example the booking id), so a replayed trigger can never write the same row twice. Read prior rows with ctx.collections.query when the automation must know what it already did. The platform: Connect to my Autoploy MCP server, call get_authoring_guide to read the authoring guide, and build the automation with the Autoploy SDK contract. Model every yes/no fork as a condition step with truthful labeled branches, keep every external side effect idempotent and keyed, and deploy it with deploy_automation. Before you build: Repeat the plan back to me in plain English and ask follow-up questions about anything unclear or ambiguous. Do not guess. Ask me for each key by name when it is time, and never ask me to paste a key anywhere except my own secure setup.