Skip to content

References and imports

You have already written most of this down somewhere. References is where you hand those materials over, and the platform reads them and proposes what it found.

A reference is a source, not a rule. Uploading a brand guide does not change your profile. It queues an ingestion job, and what that job extracts arrives as a proposal you approve or reject, field by field. Nothing an import reads becomes canonical without you.

Each reference also carries its own approval state: Pending, Approved, or Rejected. That is your judgement on the source itself, separate from the fields it produced.

The References tab with the site and the uploaded brand notes, and the ingestion that read them.

  1. Drop a file on Add references, or browse.

  2. PNG, JPG, SVG, PDF, DOCX, or PPTX, up to 25 MB.

  3. The file appears in the list and the ingestion starts on its own. Processing shows while jobs are in flight.

What is read depends on the file. A brand guide gives up hex colours, the fonts it names, and the phrase lists under headings like “words to avoid”. An image gives up its palette. A deck gives up both, plus whatever it says about the company on its opening slides.

Alpenglow’s brand notes produced two: the three brand hexes, and the banned-phrase list.

Paste one or more URLs into or paste a link and choose Import. One line per URL, so an About page and a home page can go in together.

A site import reads the page and its stylesheets: your company name, your description, what you say you do and what makes you different, and the colours the page actually uses. It is the fastest way to fill an empty profile, because it starts from what you already publish.

A repository can be read too, and it produces the richest marketing context of the three: an agent clones it, reads it the way a person would, and reports back what the product actually does.

This one is not on the References tab. Ask Eva to import a repository for your profile, and give her the repository. It needs the GitHub App connected and a runner available, because the read happens on a runner rather than in the control plane.

Each card shows its own state:

  • Processing while it runs, alongside the count at the top of the list.
  • Extracted N fields when it finishes, linking to the proposal it raised.
  • Failed with the reason, and a Retry.

A job that finishes having found nothing says Extracted 0 fields and raises no proposal. That is not a failure: it means the file had nothing the extractor recognised, which is common for a design-heavy PDF with no text in it.

Rerun extraction re-reads a reference you have already uploaded, which is worth doing after you replace the file behind it.

Approve or Reject on the card records what you think of the source. Rejecting does not undo a proposal you already approved: the fields are yours now, and changing them means editing the tab or restoring an earlier version.

The overflow menu removes a reference entirely, along with its ingestion history.