Make Agent Fast

View saved version →

---
name: fix-agent-answers
description: Find questions a Make Agent Fast agent answered wrongly or could not answer, agree the correct answer with the user, save it, and confirm the agent now uses it.
---

Use this when the user says the agent gave a wrong answer, or wants to fix unanswered questions.

1. Call `list_agents` to get the `agent_id`.
2. If the user quotes a wrong answer, find that conversation with `list_conversations` (use `search` with words from the question) and read it with `get_conversation`. Otherwise call `list_unanswered_questions`.
3. For each question, ask the user for the correct answer, or draft one from facts they give you. Never invent prices, hours, addresses or policies.
4. Show the final wording. Explain that saving supersedes the previous answer and can send replies to customers waiting for it on connected channels. Use `save_answer` only after the user approves those effects and the wording.
5. Confirm the fix with `test_agent`, asking the same question, and show the user the agent's new answer and its source.
6. If the answer needs a whole page of information rather than one reply, offer `add_knowledge` with the page address or the text.

Each `save_answer` call creates a new knowledge source and uses embedding credits; do not retry a successful save. `test_agent` uses credits, so test each fixed question once.

View saved version →

---
name: get-started
description: Give a first overview of the user's Make Agent Fast agents — which are live on the website, recent conversations and leads — and suggest the next useful step.
---

Use this when the user has just connected Make Agent Fast or asks what it can do.

1. Call `list_agents`. If the company has no agents, offer to build one: ask for the business website address and call `create_agent_from_website`, or, if the business has no website (only Telegram or Instagram, for example), ask for its products, prices, hours, location and how customers order, and call `create_agent_from_notes`.
2. For each agent (at most three), call `get_report` with `days: "30"` and summarize in one short paragraph: conversations, leads, whether the website widget is on, and the most asked questions.
3. Call `list_unanswered_questions` for the busiest agent. If there are any, say how many and offer to fix them together (see the `fix-agent-answers` skill).
4. End with two or three concrete next steps the user can ask for, for example "read the latest WhatsApp conversation" or "list this week's leads".

Rules:
- Every agent tool needs the `agent_id` from `list_agents`; never guess one.
- Report only what the tools return. If a number is zero, say so plainly.
- Ask before any tool that changes data or reaches a customer: `create_agent_from_website`, `create_agent_from_notes`, `add_knowledge`, `save_answer`, `reply_to_customer`, `update_lead`, `set_website_widget`, and `test_agent` (it uses credits).

View saved version →

---
name: weekly-customer-review
description: Review the past week for a Make Agent Fast agent — conversations by channel, new leads, and questions the agent could not answer — and propose fixes.
---

Use this when the user asks how the week went, what customers asked, or for a weekly review.

1. Call `list_agents` and pick the agent the user means. If there are several and it is unclear, ask.
2. Call `get_report` with `days: "7"`. Summarize conversations by channel, leads, bookings, handoffs, and the change from the previous week.
3. Call `list_unanswered_questions` with `days: "7"`. List each question with how often it was asked.
4. Call `list_leads` and show the new leads from the past week: name, what they asked, and whether they are qualified.
5. Call `list_conversations` with `only_unanswered: true` and open the most recent one with `get_conversation` so the user sees a real example of what went wrong.
6. Propose short answers for the unanswered questions. Explain that `save_answer` supersedes the previous answer and can reply to customers waiting for it on connected channels. Save only after the user approves the wording and those effects; do not retry a successful save.

Keep the review to one screen: totals first, then the list of gaps, then the leads. Never contact a customer or change a lead without the user asking.