> ## Documentation Index
> Fetch the complete documentation index at: https://alphacastio.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Writing better prompts

> How to frame requests so the Alphacast MCP returns exactly the data, pipeline, or diagnostic you need — five guidelines, a broad-to-precise example, copy-paste templates, and the safety pattern for write actions.

You don't need a rigid formula. The clearer the frame, the more direct and useful the result. You set the goal; the assistant reasons; the MCP does the heavy lifting against Alphacast. This page shows how to phrase requests so you spend less time re-asking and more time using the data.

## Five guidelines for better results

<CardGroup cols={2}>
  <Card title="1 · Objective" icon="bullseye">
    Say what you want to do: **find**, **compare**, **quality-check**, **download**, or **build a process**. The verb sets the whole task.
  </Card>

  <Card title="2 · Indicator" icon="tag">
    Pin down **country, variable, source, and unit**. If you know the dataset name or ID, even better — it removes ambiguity between similar series.
  </Card>

  <Card title="3 · Frequency" icon="clock">
    State whether you need **daily, monthly, quarterly, or annual** data. Mixing frequencies is a common cause of surprising results.
  </Card>

  <Card title="4 · Period & scope" icon="calendar">
    Give the **dates** and say **where to look**: your repositories, the general catalog, or one specific provider.
  </Card>
</CardGroup>

<Card title="5 · Format" icon="table-list">
  Ask for the shape you want back: a **table**, a **download**, a **diagnostic**, a **chart**, or a **pipeline proposal**. For a large task, split the message into **context, task, and acceptance criteria**; if it has several stages, number them and validate the intermediate results before moving on.
</Card>

## From a broad ask to a precise one

The same question, sharpened in three steps:

<Steps>
  <Step title="Too broad">
    > "How's inflation doing?"

    Missing country, source, frequency, period, and expected output — the model has to guess on every axis.
  </Step>

  <Step title="Better">
    > "Give me Argentina's monthly inflation for the last year."

    Now country and frequency are defined, but there may still be different sources and versions.
  </Step>

  <Step title="Precise">
    > "In the Alphacast catalog, identify the monthly INDEC National CPI dataset. Give me the month-over-month change for 2024 in a table, and flag any missing observations or base changes. Include source, IDs, unit, and the latest available date."

    Indicator, source, frequency, period, format, and quality checks are all defined — so is what "done" looks like.
  </Step>
</Steps>

## Templates to copy

<CodeGroup>
  ```text Data query theme={null}
  In the Alphacast catalog, find the monthly EMAE dataset from INDEC.
  1. Show me the candidates and confirm which one is the seasonally adjusted general level.
  2. Tell me the latest date it reaches.
  3. Give me 2023 and 2024 in a table with date and value.
  4. Include source, IDs, unit, and any transformation applied.
  ```

  ```text Pipeline change theme={null}
  On pipeline 4319, I want to add a step that multiplies the variables by 100.
  1. Show me the current steps and their order.
  2. Propose where the new step goes and explain its impact.
  3. Preview the output.
  4. Don't save, run, or publish until I confirm.
  ```
</CodeGroup>

## Traceability: ask where the data came from

Every result can carry what you need to verify, cite, and reuse it. Ask for it explicitly and the model will include it:

* Source and dataset name
* Dataset and variable IDs
* Unit, frequency, and period
* Latest available observation
* Formulas and transformations applied
* Missing values, assumptions, and warnings

<Tip>
  A ready-to-use closer for any data prompt:

  > "Include source, IDs, unit, frequency, latest available date, and the transformations applied."
</Tip>

## Preview before you commit

<Warning>
  For any action that changes Alphacast — creating or editing a pipeline, saving, running, or publishing — ask the model to **show what it will change, preview the result, and wait for your confirmation** before it writes anything.
</Warning>

A phrase that works well in write-capable clients:

> "Show me the proposed change and a preview first. Don't save, run, or publish until I say so."

The MCP surface is read-only except for the four pipeline write tools, and destructive actions (deleting repositories, datasets, or pipelines) are never exposed to an agent — but the preview-first habit still keeps edits predictable. See [Safety boundaries](/mcp/tools/overview#safety-boundaries) for the full list.

## Who does what

Understanding the division of labor helps you phrase requests — and know when something is genuinely out of scope rather than a bad prompt.

| Actor                        | Role                                                                                               |
| ---------------------------- | -------------------------------------------------------------------------------------------------- |
| 🧠 **The AI assistant**      | Understands the request, organizes the task, and analyzes the results.                             |
| 🔌 **The Alphacast MCP**     | Gives the assistant the tools to search, read, and operate on Alphacast and its enabled providers. |
| α **The Alphacast platform** | Manages permissions, stores datasets and charts, and runs or publishes pipelines.                  |

Every task falls into one of three buckets:

<CardGroup cols={3}>
  <Card title="Yes — inside the MCP" icon="circle-check">
    Search, download, transform, and operate on Alphacast within your permissions.
  </Card>

  <Card title="Depends on the assistant" icon="circle-half-stroke">
    Reading an attached file, creating an Excel or document, generating an image, or browsing the web.
  </Card>

  <Card title="No — out of scope" icon="circle-xmark">
    Bypassing permissions, modifying the original source, or inventing data that doesn't exist.
  </Card>
</CardGroup>

<Note>
  A few limits are worth keeping in mind:

  * **It works with your permissions.** The MCP can only reach resources your account is authorized for.
  * **Freshness depends on the source.** It retrieves the latest available observation but can't produce a data point that hasn't been published yet.
  * **Detecting isn't always explaining.** It can flag a jump or a pattern change, but pinning down the cause may require checking metadata or official documentation.
</Note>

## If the answer isn't what you expected

You can course-correct inside the same conversation — no need to start over:

* **Refine** — "same series, but seasonally adjusted."
* **Narrow** — "only from 2020, monthly frequency."
* **Ask for alternatives** — "show me the candidate series before picking one."
* **Ask for traceability** — "tell me which dataset it came from and what transformation you applied."
* **Split the task** — search and validate first; transform or publish afterward.
* **Stop a change** — if something isn't clear, tell it not to run and to show the proposal again.

## Next steps

<CardGroup cols={2}>
  <Card title="Examples" icon="lightbulb" href="/mcp/examples">
    End-to-end walkthroughs that combine several tools to solve realistic tasks.
  </Card>

  <Card title="Tools overview" icon="wrench" href="/mcp/tools/overview">
    The full list of MCP tools, grouped by capability, with input schemas.
  </Card>
</CardGroup>