Many software companies now offer a way to connect their product to Claude or ChatGPT. The connection takes minutes, and then the customer opens a new chat and faces an empty box with no idea what to ask. On 12 November 2026 Livestorm, a webinar platform, launched the Webinar Prompt Library to deal with that. This playbook shows how to build the same thing for your own product: a small set of tested prompts that give a new user a useful answer from their own data on their first try.

The idea and why it works

A connector lets an AI assistant read, and sometimes change, the data inside your product. Most connectors are built on the Model Context Protocol (MCP), an open standard for linking AI apps to outside systems. Livestorm's version is a remote MCP server with OAuth sign-in that works with Claude, ChatGPT, Microsoft 365 Copilot and other assistants, the company says.

Connecting is quick. Livestorm says an admin does it once, "in about two minutes". What happens next is the hard part. In the launch post, Livestorm's growth manager, Brillixa Herdhiana, says that once everything is connected, "the hardest part is knowing what the MCP can do for you. Which use cases? Where do you start? You hit a 'now what?' moment, like staring at a blank page."

The blank chat box causes a second problem too. A broad question gets a broad answer. Livestorm's example: ask "which of my webinars did best?" and the assistant counts your internal all-hands as a webinar, counts your own team as attendees, and puts a three-person test event with perfect attendance at the top of the ranking. The answer comes back in a neat table that looks confident and is wrong. A user who gets an answer like that once will reasonably decide the connector is a toy.

A prompt library fixes both problems. Each prompt names one job, already knows the traps in your data model, and runs on the customer's own records, so the first answer is specific to them and far less likely to be wrong. Livestorm's head of growth, Benoît Leggieri, is plain about the commercial aim: "help more of our customers adopt the MCP, get more out of Livestorm". Customers who get answers from your product inside the assistant they already use every day have one more reason to keep paying.

There is a second payoff. Each prompt gets its own public page, which describes a real use case in the words customers search for. Livestorm's library listed 26 prompt pages when we counted on 19 November, sorted into five stages: plan, promote, animate, repurpose and analyse.

Who it's for

This play fits any small SaaS team that has shipped an MCP server or AI connector, or plans to soon, and whose product holds data the customer cares about: bookings, invoices, tickets, orders, events, subscribers, projects. It works best when the obvious questions have traps in them (test records, internal accounts, refunds, duplicates) that a general-purpose assistant will not know about. If your connector only does one simple thing, a few example prompts in your setup docs may be enough, and a full library would be more than you need.

What you need

  • Tools: your working connector, accounts on Claude and ChatGPT for testing, two or three real customer workspaces you have permission to test on (or your own production account), and a web page template for prompt pages.

  • Time: about two hours per prompt to write, test on three datasets and fix, so roughly 20 hours for a first set of ten (estimate). Add a day for the pages and the launch links.

  • Cost: close to nothing beyond staff time and existing assistant subscriptions.

  • Tracking: server-side logs of connector tool calls per account, with timestamps, so you can see who asked a second question.

Step by step

1. Pick the ten jobs customers do in their first week

Use your support tickets, onboarding calls and connector logs to list the questions new users actually ask. Choose ten that are frequent, produce a decision, and touch data the user already has on day one. Livestorm grouped its prompts by stage of the job (plan, promote, animate, repurpose, analyse), which also gives you navigation for free. Avoid prompts that only work after months of history; a new account would get an empty answer.

2. Write each prompt as a numbered procedure

A good library prompt reads like a short brief to a careful analyst. It says which records to include, which to skip, how to calculate, what to compare, and what shape the answer should take. Livestorm's "plan a webinar" prompt is a good model. It looks at public webinars over a period you set (12 months by default) with at least 15 registered participants, skips all-hands, sales calls and test events, sets aside and reports sessions with more than three times the usual registrations, counts each recurring webinar once using its median, and marks any group with fewer than three rows as a small sample. Put the parts the user changes in square brackets, such as [TOPIC] or [LAST 12 MONTHS], and keep everything else fixed.

3. Build the guardrails into the text

  • Exclude test, internal and staff records by name and by pattern.

  • Show the numbers behind any conclusion, including how many records sit behind each row.

  • Say when the data is too thin to answer, and say what is missing.

  • Before any write (creating, editing, sending, deleting), show exactly what will change and wait for the user to approve. Livestorm creates new events as drafts so no registration email goes out until the customer publishes.

4. Test on real data, from the customer's chair

Run every prompt on at least three real workspaces: a busy one, a new one, and a messy one with test records and odd outliers. Run it in both Claude and ChatGPT. For each run, check four things. Did the user get a specific answer about their own data within one message? Were the traps handled? Did a thin dataset produce an honest "not enough data" instead of a confident guess? Did any write wait for approval? Livestorm says every prompt in its library was "tested on real Livestorm data" before it went in. Fail a prompt if any check fails, fix the wording, and run it again.

5. Give each prompt a page and a one-click launch

Each page should say what the prompt is for, what it needs, and what the user gets back, then offer three buttons: copy, open in Claude, open in ChatGPT. Livestorm's buttons open a new chat with the prompt already filled in, using links of the form claude.ai/new?q= and chatgpt.com/?q= followed by the URL-encoded prompt. Write the page title in the words a customer would search, for example "score webinar leads for sales follow-up".

6. Put the library where connected users land

Link it from the connector setup success screen, the setup guide, the in-app integrations page and the first email after someone connects. Livestorm's library page tells existing customers to "Connect the MCP and try your first prompt". The moment right after connecting is when the empty box hurts most, so that is where the first prompt link belongs.

7. Measure the second question and prune

Your main number is the share of connected accounts that make a second connector request within seven days. One request shows curiosity. A second one suggests the first answer was useful. Review it monthly by prompt, drop the prompts nobody runs twice, and add the questions support keeps getting.

Template and example prompts

Here is a template you can adapt. Replace the bracketed parts and the product nouns.

Using my [PRODUCT] data, [JOB IN ONE LINE] for [PERIOD].

1. Include only [REAL RECORD TYPE] with at least [MINIMUM SIZE]. Skip internal, test and staff records, including anything named "test" or created by my team members. List anything you skip that looks unusual and say why.

2. For each record, note [FIELDS].

3. Compare [GROUPS]. Mark any group with fewer than [N] records as a small sample and show how many records sit behind each row.

4. Recommend [DECISION], with a runner-up and how confident you are.

5. If there is too little data to answer, say so and tell me what is missing instead of guessing.

6. Do not create, change or send anything. If I ask you to, show me exactly what will change and wait for my OK.

Answer with the recommendation first, then the table, then what you left out and why.

Three examples for different products:

  • Invoicing tool: "Using my invoice data, list the five clients most likely to pay late this quarter, based on their average days to pay over the last 12 months. Skip test clients and invoices under [AMOUNT]. Show each client's on-time rate and how many invoices it is based on."

  • Booking app: "Using my booking data, find my three least-filled time slots over the last 90 days. Exclude cancelled bookings and staff test bookings. Suggest which slots to drop, and say so if any slot has fewer than ten bookings to judge from."

  • Newsletter platform: "Using my subscriber data, show which sign-up sources brought the readers who are still opening after 60 days. Exclude my own and my team's addresses. Draft a segment of inactive readers, but create it only after I approve."

A worked model

Illustrative maths, not Livestorm data. Livestorm has not published any results from the library. The numbers below are assumptions to show how to size the opportunity for your own product; replace each with your own figures.

Assumption (illustrative)

Without library

With library

Accounts that connect each month

1,000

1,000

Ask a second question within 7 days

25% = 250

40% = 400

Retained at 90 days if they asked a second question

70%

70%

Retained at 90 days if they did not

40%

40%

Accounts retained at 90 days

250 × 0.7 + 750 × 0.4 = 475

400 × 0.7 + 600 × 0.4 = 520

The library adds 150 accounts that have seen a real answer from their own data, and on these assumptions 45 more accounts still paying at 90 days from each monthly cohort. Part of the retention gap between the two groups is selection: keen customers ask more questions and also stay longer. To allow for that, credit the library with only half the difference, about 22 accounts per cohort. At an illustrative $50 a month per account, that is roughly $1,100 of extra monthly recurring revenue from each month's cohort, against about 20 hours of one-off writing and testing. If your second-question rate is already above 40%, the library will do less for you and the effort may be better spent elsewhere.

Metrics to watch

  • Connector activation: the share of connected accounts that make at least one successful request within 24 hours of connecting.

  • Second question within 7 days: your main measure, split by the prompt that started the session.

  • Prompt page to run: of the people who open a prompt page, how many trigger a connector request in the next hour.

  • "Not enough data" rate: how often each prompt reports thin data. A high rate on new accounts means the prompt belongs later in the journey.

  • Write approvals: how often a proposed change is approved, edited or rejected. Frequent rejections mean the prompt is guessing at intent.

  • 90-day retention, connected versus not: watch the gap over time, keeping the selection caveat above in mind.

  • Organic visits to prompt pages: a sign the pages are working as acquisition too.

Pitfalls

  • Shipping untested prompts. One confident wrong answer does more damage than an empty box. Test on messy data before anything goes live.

  • Writes without approval. An assistant that emails a customer list or deletes records by mistake will end the connector for that account, and possibly the account itself. Default to drafts and confirmation.

  • Prompts that need history. A new account with two weeks of data should still get an answer, or an honest note on what it needs.

  • Marketing prompts. A prompt that steers the user towards an upgrade or another product breaks trust. Each prompt should serve the job the user is already paying for.

  • Letting it go stale. When your data model or connector tools change, prompts break quietly. Re-run the test set after every connector release.

  • Counting opens as success. Page views and copies say little. Measure the second connector request.

Variations

  • Built-in prompts. The MCP specification lets a server expose user-controlled prompt templates that a client can show, for example, as slash commands. Shipping your best three this way puts them inside the assistant, with no website visit needed.

  • Team skills. Livestorm says it plans to let customers "package your favorite prompts as reusable skills your whole team can run", listed as coming soon. For team products, a shared set of saved prompts spreads usage from the admin who connected to the rest of the team.

  • Customer-submitted prompts. Livestorm says the library grew from use cases and prompt ideas that customers, partners and its own teams shared after the MCP launched. Invite submissions, test them, and credit the author on the page.

  • Starter prompt on connect. Show one prompt on the success screen, chosen from what the account has the most data for. It pairs well with a proactive first message and the same instinct behind letting people get an anonymous first task done before anything else.

  • Clarify before running. For prompts where the right answer depends on context, have the prompt ask one or two questions first, as covered in clarifying questions before the first build.

The short version: list the ten questions new users should ask, write each one so it handles your data's traps, test it on real accounts, and make it open in one click. Then watch whether people come back for a second answer.