Write down the moment the customer would ask for help. A brand voice assistant starts with a draft that sounds wrong. A tutor starts with a student stuck in practice. A calculator starts when chat needs a number, not another paragraph.
Docs
Follow the path from one repeated question to a paid AI app.
A creator usually starts with the same customer question showing up again and again. Prostir turns that question into a small app: first the source material, then the behavior, then tools and state, then a remote MCP endpoint with access, quotas, and billing.
Ask an AI client
Docs
Start from a concrete case.
The fastest path is to pick a case study, copy its structure, and replace the example knowledge with your own.
01
1. Start with the scene
Write down the moment the customer would ask for help. A brand voice assistant starts with a draft that sounds wrong. A tutor starts with a student stuck in practice. A calculator starts when chat needs a number, not another paragraph.
01 1. Start with the scene
02
2. Give the app a memory of your work
Upload the examples, policies, lessons, scripts, formulas, and checklists that make the answer yours. The app should not sound confident because the prompt is clever; it should answer because the approved source exists.
02 2. Give the app a memory of your work
Use approved examples Separate public and private material Check answers against source coverage
03
3. Add a tool when the story needs an action
Some moments cannot be solved by retrieval. Pricing, scoring, eligibility, refund logic, and plan limits need deterministic tools. The assistant asks for missing inputs, runs the named JavaScript or API tool, and explains the assumption behind the result.
03 3. Add a tool when the story needs an action
3. Add a tool when the story needs an action
Ask for missing inputs Run the named tool
04
4. Model the workflow as states
When a user must move through steps, describe it the classical way: current state, event, guard, action, and next state. Prostir keeps that current state and user memory inside Orleans-backed app, session, user, and access-grant boundaries.
04 4. Model the workflow as states
- Current state
- Event or trigger
- Guard or can-fire check
- Action and next state
05
5. Publish the app as a product
The private prototype becomes real when it has a stable remote MCP endpoint. Access rules decide who can use it, Stripe decides which plan unlocks it, and quotas protect the runtime when customers start using it repeatedly.
05 5. Publish the app as a product
06
6. Know the runtime underneath
Prostir is built by Managed Code on .NET, Orleans, MCP, and ManagedCode packages. Microsoft.Extensions.AI is the .NET abstraction reference for chat clients, embeddings, function invocation, telemetry, and middleware.
06 6. Know the runtime underneath
Microsoft.Extensions.AI ManagedCode.MCPGateway ManagedCode.MarkdownLd.Kb ManagedCode.Storage
07
7. Trace one customer request
A real request is a journey: the MCP client calls the endpoint, Prostir resolves the customer access grant, checks quotas, chooses knowledge or a tool, runs inside the configured policy, and returns an answer the app is allowed to give.
07 7. Trace one customer request
7. Trace one customer request
MCP client request Access grant and quota check
08
8. Rehearse the difficult scenes
Before customers rely on the app, test the parts that break the experience: wrong-user questions, missing tool inputs, unsupported statements, quota exhaustion, provider failures, and the handoff path. That is where an AI demo becomes an AI product.
08 8. Rehearse the difficult scenes
- Wrong tenant or user
- Missing inputs
- Quota exceeded
- Unsupported answer refusal
Docs
Start from a concrete case.
The fastest path is to pick a case study, copy its structure, and replace the example knowledge with your own.