Best practices with MCP and the app

Before you start

• SwiftCNS installed and connected to your AI tool.

Steps

1. Work in two windows

   Ask in your AI tool, because it already knows what you're working on. Keep SwiftCNS open beside it: every record your tool creates appears there live, the Now tab shows the next step, and you decide with the whole lineage in view.

2. Describe the idea, not the method

   Say what you want to do and why, with the facts you have. Name SwiftCNS so your tool uses it.

   Use SwiftCNS to test this: [the idea]. Context: [numbers, quotes, constraints].

   Don't ask your tool to list assumptions or design an experiment itself. SwiftCNS does that, scores the options and keeps them on the record.

3. Make each choice explicitly

   Tell your tool what you chose; it submits the choice for you.

   Select the assumption about [topic] and the hypothesis tested at [moment]. Show me the experiment options; I'll choose.

   Choose the [design] experiment, assign it to me, due [date], and approve it.

4. Bring evidence from where it lives

   Paste rows, point to a CSV, or ask your tool to pull results from a source it can read.

   Add these results to the SwiftCNS tracker, one row per [unit]: [rows]. Then show me the totals against the success criteria.

   When the sample is in: "Complete the tracker."

5. Keep the two calls that matter

   The result and the decision are yours. Make them in the app, or dictate them:

   Set the Learning result to [Met / Not met / Unclear], write: "[what I learned]", and publish it. Then prepare a decision review for [question], citing the Learning, without a recommendation.

   Record my decision: [Move forward / Change course / Test again / Drop it]. Why and next: "[reasons and next step]".

6. Pick up later

   Days later, ask: "Where does my SwiftCNS test on [topic] stand?" Your tool reads the current step and carries on from there.

What you get

Your AI tool does the typing; SwiftCNS does the drafting and keeps the record; you make the calls. In the running example, the whole cycle took about 20 minutes of tool time, and every step appeared in the app as it happened.

Done when

• You ran a step from your AI tool and saw it land in SwiftCNS.
• Your tool's instructions include the SwiftCNS rule.

Common mistakes

• "Just run the whole thing." Your tool will stop at each choice, or make choices you meant to make. Say which steps it may take for you.
• Asking for a recommendation before you decide. Make your own call first; it keeps your judgment and your record honest.
• Losing the thread. Keep one question per conversation; start a new one for the next idea.

For agents: MCP tools and gotchas

• Read get_conversation_progress before acting: activeActionRequests, actionOperations, permittedTransitionIds, continuationMode.
• send_message (with a stable requestKey) returns an assistantMessageId; poll get_message until execution.status is SUCCEEDED, then reload progress.
• Pass continueConversation: true with a branch-leaf parentMessageId only when the user asked to continue.
• Submit only what the user chose or explicitly delegated. Never set a Learning result or record a Decision on your own judgment.

Next

Permissions: what your tool may do, and how to revoke it.