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

# Prompts

> Manage tracked questions and regenerate an ICP-driven prompt set.

Prompts define the questions used to measure AI answer visibility. Each prompt records text, intent, topic, language, importance, state, and lineage.

## Prompt management

You can add, edit, archive, restore, and list prompts from Settings. User-authored prompts remain distinct from generated prompt lineage.

| Field      | Current constraint                      |
| ---------- | --------------------------------------- |
| Text       | 3 to 2,000 characters                   |
| Topic      | 1 to 120 characters                     |
| Importance | 0 through 1                             |
| Intent     | One of the six registered intent values |
| Language   | Validated language identifier           |

The registered intents are:

| Intent                         | Use                                                      |
| ------------------------------ | -------------------------------------------------------- |
| <code>discovery</code>         | Explore a category or set of options                     |
| <code>problem\_solution</code> | Find an answer to a stated problem                       |
| <code>use\_case</code>         | Find a product or approach for a specific situation      |
| <code>comparison</code>        | Compare named options or ask for alternatives            |
| <code>expert</code>            | Ask for specialist guidance or evaluation criteria       |
| <code>brand\_research</code>   | Research the tracked brand under its separate brand rule |

Archiving removes a prompt from active measurement without deleting its history. Restoring makes it eligible again.

## ICP-driven generation

Current regeneration reads the workspace ICPs and the latest eligible research artifacts. It does not invent a replacement for stale or missing research.

<Steps>
  <Step title="Refresh research">
    Complete a research run no more than seven days before regeneration.
  </Step>

  <Step title="Freeze the source selection">
    The generator stores the exact artifacts and source digests used by the job.
  </Step>

  <Step title="Generate candidates">
    Candidate questions cover the workspace market, ICP patterns, languages, and configured target.
  </Step>

  <Step title="Apply bias guards">
    Deterministic checks and a critic reject leading, favorable, duplicate, or unsupported questions.
  </Step>

  <Step title="Store lineage">
    Accepted prompts and rejected candidates retain job, provider, and evidence references.
  </Step>
</Steps>

## Bias guards

Research-driven generation rejects a candidate when it:

* Names the tracked brand
* Requests favorable treatment or assumes a preferred answer
* Refers to evidence that cannot be traced
* Manufactures a search query or unresolved reference
* Does not match its generation stage
* Uses vendor-leading language
* Duplicates another normalized prompt
* Reaches the configured critic bias threshold

<Note>
  The older CLI curation path has a separate <code>brand\_research</code> rule that can require a tracked-brand mention. The current research-driven web regeneration path applies the stricter no-brand rule to its candidates.
</Note>

## Regeneration behavior

Regeneration runs as its own job and consumes provider credits. Rejections are kept for inspection. User prompts are not overwritten by generated curation.

If regeneration fails:

| Error                          | Resolution                            |
| ------------------------------ | ------------------------------------- |
| No completed research          | Run the research pipeline             |
| Research older than seven days | Refresh research                      |
| Provider unavailable           | Check operator OpenRouter access      |
| Credits stopped                | Add credits and start a new operation |
| Candidate rejected             | Inspect the stored rejection reason   |

## Measurement impact

Only active prompts enter a measurement run. Preset caps and per-engine caps can reduce the number sent. Importance affects downstream selection and reporting but does not bypass a preset cap.