Prompting Guide

Ground first

Tell the assistant what universe it's in before asking for action:

The MemberPass enum catalog is at memberpass://enums/subscription-status. Load it, then tell me which of my subscriptions are past due.

This is faster and safer than letting the assistant guess a status name it doesn't know.

Be explicit about which project

Almost every tool takes a project_id or project_handle. Prompts that name the project once up front are far more reliable:

Work on the project with handle research-premium for the rest of this conversation. First, list its active plans.

This prevents Claude from silently switching projects mid-chat, which is the most common source of confusion.

Scope the token narrowly

An agent token limited to project:view-any, project:view, project-subscription-plan:view-any, and project-user:view-any can answer 90% of analytics questions. Don't hand it write abilities unless the prompt explicitly involves writes — mint a second token instead.

Use the assistant for aggregations

The server surface is low-level. The assistant's value is combining tools:

For each active plan on research-premium, count subscribers whose most recent payment succeeded in the last 30 days.

That becomes: list_plans → iterate → list_recent_payments(plan_id=..., since=...) → tally → render.

Avoid unbounded loops

When you ask "find every X", cap it:

List up to 50 subscribers who match…

Otherwise an overly-literal agent paginates through 100,000 rows before answering.

Drive writes through previews

When write tools land, run them through a dry-run pattern:

Plan 10 new access codes for research-premium. Show me the exact payload you'd send, but don't execute yet.

Then:

OK, go ahead.

This is cheaper than re-running a large batch operation because the prompt was unclear.