Responsible AI Use in Sessionplan

If you have ever asked ChatGPT to design a complete workshop agenda, you know what happens next: the list is good, but it is stuck in the chat window. Someone still has to move it into a planning tool, enter the timings, and create the blocks. That final step can be automated. The question is how much control you want to hand over.

Sessionplan deliberately has no built-in LLM. There is no chat window in the app and no background process that sends your notes to an AI provider. The reason is simple: Sessionplan stores your data locally in the browser by default, without a server. A built-in LLM would undermine that principle as soon as your agenda content had to be sent to a cloud service for processing.

AI support is still possible. It just runs outside Sessionplan, in the AI tool of your choice. Here are three ways to do it, in ascending order of automation.

Way 1: The manual prompt

The most direct option is in the Sessionplan settings: a pre-written prompt that you copy into any LLM. Claude, ChatGPT, Gemini, a local model – whichever tool you use and trust.

Here is how it works:

1. Open the Create session with AI setting and copy the prompt.
2. Describe your workshop in the chatbot of your choice – duration, group size, goal, and any special requirements.
3. When you are ready, paste the prompt into the LLM. It returns a complete Sessionplan JSON object.
4. Import the JSON into Sessionplan and use, share, or continue editing the session there.

The advantage is the complete absence of an automatic process between you and the result. You see the JSON before it goes anywhere. You can read it, edit it, or reject it. No tool calls an API in the background, and no connection to Sessionplan is involved during the prompt process.

That makes this the right choice for anyone who wants maximum control or needs to document exactly which data goes where for professional reasons – for example, in organisations with strict compliance requirements, or simply for people who prefer to see every step before it happens.

The downside is obvious: copying and pasting in both directions takes time. For larger changes – “move the break ten minutes later” – you have to repeat the process from scratch.

Way 2: The skill for your own LLM

The second option reduces the copying without creating a permanent connection to Sessionplan. It uses the Sessionplan API skill file, a small text file that you load into an LLM capable of running code or making HTTP requests, such as Claude Code, Cursor, or a similar coding assistant. For tools that can import OpenAPI definitions, the API’s OpenAPI specification is available as well.

The Skill file describes two lightweight REST endpoints:

• data/encode.php turns a complete agenda JSON object into a Sessionplan link
• data/decode.php reads an existing link back into JSON

Once you have loaded the file into your tool, your LLM knows how to call both endpoints. It can build an agenda from a description, create the link, and give you a clickable Sessionplan link directly – or take an existing link, modify it, and return a new one.

The important part is that Sessionplan does not store anything. Every call to the endpoints is stateless: the request arrives, the server processes it in memory, the response goes out, and the process is over. There is no account, no server-side history, and no way for Sessionplan to reconstruct your previous agendas.

Who is this for? Anyone who wants to use an LLM for Sessionplan links regularly but does not want to copy and paste for every small change. The trade-off is a one-time setup, and your AI tool must be able to make HTTP requests. A plain chat without tool support cannot use this approach.

Way 3: The MCP server as a complete solution

The third option requires the least setup and works directly in an increasing number of AI tools. MCP, the Model Context Protocol, has become the common standard for letting an LLM talk to external tools. Instead of loading a text file that your LLM has to interpret, you enter just one URL:

https://sessionplan.de/mcp/

Tools such as Claude Desktop, MCP Inspector, and coding environments with MCP support can automatically discover two tools there: one to create a link from an agenda and one to read an existing link. No manual description and no copy-pasting instructions are needed; the protocol handles the communication.

The same privacy promise applies here as with the other two approaches: the MCP endpoint does not store data permanently, run accounts, or collect a history. Each request is handled as a single operation and ends when that operation is complete.

The difference from Way 2 is mainly the setup effort and which tools support the connection. A coding assistant without MCP support can still use the skill approach. A chat client such as Claude Desktop, on the other hand, needs MCP because it does not run code itself.

Which way is right for whom?

The three options mainly differ in how much you still have to handle yourself before the link is created.

Way	Automation level	Technical requirement	Your control
1. Manual prompt	Low	Any LLM with a chat window	Maximum – you see every JSON object before importing it
2. Skill file	Medium	An LLM that can run code or make HTTP requests	Medium – you review the generated link
3. MCP server	High	An MCP-capable tool (e.g. Claude Desktop)	Lower – the protocol handles the communication

Read between the rows and a pattern emerges: automation level and manual control move in opposite directions – none of the three ways maximizes both at once.

A real-world example: a Scrum Master automates weekly retros

A Scrum Master facilitating weekly retrospectives for three teams set up the MCP server once in their coding assistant. Instead of drafting a fresh agenda every week, they now describe the format they want – say, “Start-Stop-Continue, 45 minutes, 8 participants” – and get a finished Sessionplan link back directly. The real time savings show up not in the first draft but in the small tweaks afterward: “move the check-out five minutes earlier” happens in a single sentence instead of another round of copy-pasting. They still never type confidential feedback content into the LLM – only the bare structure of the agenda.

If you want to see exactly what happens at every step, use the prompt in the settings. If you use a coding tool every day and want to add Sessionplan links as one more capability, install the skill. If you want the least setup and would rather enter a single URL, use the MCP server.

None of these approaches is inherently better. A trainer who has good reason not to put client data into any LLM will choose Way 1 and write the agenda without real names or sensitive internal details. Someone who plans daily retrospectives for an internal team and trusts their AI setup may quickly settle on Way 3.

What all three have in common is the result: a normal Sessionplan link, the same kind of link you get when planning without AI. One click opens the timeline, where everything remains editable regardless of how the plan was created.

What AI cannot replace

An automatically generated agenda is a first draft, not a finished concept. Good facilitation depends on things no LLM can infer from a single prompt: What is the team’s mood right now? How much energy will be left after lunch? Which conflicts are simmering beneath the surface and could derail a discussion?

Common pitfalls with AI-assisted planning

• Typing real names or internal details into the prompt: entering client names, internal project names, or confidential feedback into an LLM means losing control of where that data ends up. Describe the flow structurally, without real names attached.
• Importing the generated JSON without checking it: an LLM can swap block order or suggest unrealistic timings. A quick scroll through the imported timeline before you share it further can save an awkward surprise during the live session.
• Setting up Way 2 or 3 without checking compliance requirements first: in organisations with clear AI policies, it's worth confirming upfront whether HTTP access or MCP connections are even approved.
• Treating the first draft as the finished agenda: an AI-generated structure never replaces a read on the team's actual current mood. The draft is a starting point, not a finished solution.

AI in Sessionplan FAQ

Does any of the three ways cost anything?
No. Sessionplan itself is free and needs no account, and none of the three ways requires a Sessionplan subscription. For Ways 1 and 2, all you need is access to an LLM of your choice – many offer a free usage tier.

Which LLM works best for Way 1?
Practically any current language model can return structured JSON when the prompt explicitly asks for it. The real differences show up in how reliably a given model returns plain JSON without extra surrounding text.

Does Sessionplan store the AI-generated plans on a server?
No. All three ways end in a link that carries the data the usual way, in the URL hash or locally in the browser – exactly like any other Sessionplan agenda.

Can I switch between the three ways?
Yes, any time. Since all three ways end in the same kind of Sessionplan link, an agenda created via prompt can later be edited through the MCP server without any friction, and vice versa.

Does the MCP server need a constant internet connection?
Yes, every request goes through the MCP URL – without a connection, the tool won't respond. Offline, Way 1 with a locally running model is the only practical option.

Sessionplan remains the place where the draft becomes visible, no matter which of the three approaches created it – and where you adapt it to the reality of your workshop before using it.

For AI support beyond agenda drafting – live documentation, question sparring, follow-up – see our post on AI co-facilitation.