Skip to main content

Start here

Quick start

Follow these steps in the chatbot you already use. Each step has a short instruction and a prompt to copy.

New to SMEme? Read why business rules don’t belong in prompts.

1. Connect SMEme

Copy the connect prompt into your chatbot. It walks you through adding SMEme in the app’s settings, using the values below. In ChatGPT, Claude, and Microsoft Copilot you click through the settings yourself. Coding agents such as Cursor and Claude Code can usually add the connector for you.

MCP connector values

The connect prompt includes these. Your app’s connector settings ask for them.

MCP endpoint URL

https://www.smeme.ai/api/v1/mcp

OAuth Client ID

When your client asks for a Client ID, leave the secret blank (PKCE public client).

NRdsdBvrio0DW9yo

Connect prompt

Help me add SMEme as a remote MCP connector.

MCP endpoint URL: https://www.smeme.ai/api/v1/mcp
OAuth Client ID (only if the app asks; leave the client secret blank): NRdsdBvrio0DW9yo

Tell me where connectors live in this app's settings, step by step. If you can edit your own MCP config, add a Streamable HTTP connector named SMEme yourself and tell me what you changed. When the SMEme sign-in tab opens, I'll sign in and come back here.

Once you’re signed in, check the connection.

Check prompt

Check that SMEme is working.

1. Call SMEme's capabilities tool. Tell me which SMEme tools you can see, and whether the tools for building decision trees are among them.
2. Load SMEme's design guidance for decision trees and tell me its version. You'll follow it when we build.
3. List my SMEme decision trees. A new account may show a sample tree. That's expected.

If any step fails, quote the error and tell me what to do next.

If your chatbot can’t see SMEme tools, start a new conversation. Most apps load connectors only when a chat starts.

2. Choose and build a decision tree

Encode a decision when it matters, when it repeats, or both. Strong candidates are costly to get wrong, made often enough that consistency counts, have to be explained afterward, or are complex enough that people regularly miss a detail. Once a decision is encoded, your agent can do more than run it. While it runs, SMEme asks only the questions that can still change the outcome. Afterward, your agent can tell you which facts decided the outcome and what would change it.

Each answer comes from evidence: a document, a record, a field in a system, a calculation, a lookup, or a person who knows. While you build the tree, tell your agent where each question’s answer should come from. That might be a tool to call, a file or database to check, or a person to ask. Also tell it where not to look, such as an outdated folder or a draft policy. SMEme saves these as hints on each question. When you run the tree later, your agent reads them as it gathers evidence, so it looks in the right places instead of guessing.

Pick the prompt that fits where you’re starting. Each one takes your agent from finding the decision to saving a draft tree. Along the way it stops for your approval three times: first the outcomes, then the questions, then the full outline. Nothing is saved until you approve.

  • Attach or link the source. Don’t summarize it.
  • Name the edge cases you know about, such as “contractors count as vendors.”
  • Say what each outcome means in practice: who does what next.
  • Bring past cases with known outcomes, including one that went the less obvious way. They become test cases that SMEme checks every time you Deploy.

You have a specific decision procedure in mind.

It’s likely named and described in a policy, runbook, or standard, either your own or a public one.

Known procedure prompt

I want to encode this procedure in SMEme: [name it].
Source: [attached file / pasted text / a URL / the name of a public standard or regulation]. If it's public and I haven't attached it, find the current official version and tell me where you got it.

Start by telling me:
- the decision it governs, in one line
- the possible outcomes
- the questions that decide between them, with the answer options for each
- where the evidence for each answer comes from, including any lookup, calculation, research, or tool it depends on, and anything the source says not to rely on
- what the source leaves open that I'll have to decide

Then build it as a SMEme decision tree. Follow SMEme's design guidance and pause for my approval at each stage it calls for. Record where each answer should come from and where not to look, and ask me when you don't know. Ask me for past cases with known outcomes and turn them into test cases. Where the source is silent or ambiguous, ask me. Don't fill the gap yourself. When I approve, validate the tree, save it as a draft, and tell me how to Deploy it.

Have your agent observe your work.

The procedure lives in your head. Start a new chat and work through a real case, or a few cases if you can.

Observe your work prompt

I'm going to work through a real [kind of case, e.g. a vendor approval] with you. The procedure isn't written down, so watch how I decide and learn it.

As we work, ask me why whenever I make a call that isn't obvious, and note where I get each fact: a file, a tool, a calculation, or someone I ask. Don't steer my answers.

When I say "done," tell me:
- the decision in one line, and the outcomes I was choosing between
- the questions that decided it, with the answer options for each
- where the evidence for each answer came from, and anywhere I said not to look
- any call that seemed to contradict an earlier one
- whether a public standard or common practice covers this, and which one

Then build it as a SMEme decision tree. Follow SMEme's design guidance and pause for my approval at each stage it calls for. Record where each answer should come from and where not to look, and ask me when you don't know. Use the cases we just worked as test cases, and ask me for more. Where something is unclear, ask me. Don't fill the gap yourself. When I approve, validate the tree, save it as a draft, and tell me how to Deploy it.

You’re not sure which decision procedure to encode.

Brainstorm from what your chatbot already knows about you, and have it interview you.

Brainstorm prompt

Help me find a decision procedure worth encoding in SMEme.

Start with what you know about me: memory, past conversations, project instructions, and any connected apps you can read. Tell me what you looked at.

Then interview me, one question at a time, to fill the gaps. Ask about my role, the calls I'm responsible for, and which ones are costly to get wrong, come up often, have to be explained afterward, or are complex enough that details get missed. Ask before relying on anything you're unsure about.

When you have enough, suggest up to five decision procedures. For each one, give me:
- the decision in one line, and why it's worth encoding
- the likely outcomes
- the questions that would decide it, and where the evidence for each would come from

Rank them by value and recommend one. When I pick one, build it as a SMEme decision tree. Follow SMEme's design guidance and pause for my approval at each stage it calls for. Record where each answer should come from and where not to look, and ask me when you don't know. Ask me for past cases with known outcomes and turn them into test cases. Where something is unclear, ask me. Don't fill the gap yourself. When I approve, validate the tree, save it as a draft, and tell me how to Deploy it.

3. Deploy, List, and try it

When your agent saves the draft, it gives you a link to the SMEme editor. Open it, or find the tree on your dashboard, and choose Deploy. SMEme runs your test cases first and stops if any of them reaches a different outcome than you said it should. Then switch Listed on so your agents can find the tree. After you edit a deployed tree, choose Redeploy. Until you do, the dashboard marks it Stale and agents keep running the last deployed version.

Then test the tree on a case where you already know the right answer. Pick one that isn’t among your test cases. Start a new chat, so the agent that runs the tree hasn’t seen how it was built. Don’t tell it the answer.

Try-it prompt

Run my [decision tree name] SMEme decision tree on this case: [describe it, or attach the file].

Gather the evidence for each question using the hints in the tree, and cite where each answer came from. Ask me when you can't find an answer. Then tell me the outcome and which answers decided it.

Next: Run a tree →. After a run, Ask more shows how to find out which facts decided the outcome and what would change it.