Build richer Brain items from typed fields and tagged references, mention a record directly in Flowgen chat, ask AI to digest attached source material into the correct item type, and link a Brand item to a project so its exact identity travels into generation prompts.
Where: Flowgen → Brain; Flowgen chat → type @; project settings → Brand.
Key ideas
- References are a record-level source layer: Every Brain item can keep links, images, videos, and documents with titles, tags, notes, extracted text, ordering, and an In use switch. They are not ordinary typed fields or canvas sockets.
- @ mentions resolve at send time: Flowgen chat lists accessible records under Brain and writes @[Title](record:id). The record is read fresh when the message is sent, then its useful typed fields and a body excerpt are appended as user-authored data without creating a graph wire.
- AI ingest maps before it asks: When the user asks to save attached material, the assistant reads every attachment, chooses a type, maps known values to that type’s real field keys, keeps remaining content in the body, attaches source files and links as references, writes one item, then asks at most three useful follow-up questions.
- A linked Brand becomes project context: Project Brand settings can link an exact Brain Brand item. The project prompt then receives its declared name, mission, audience, tone, emotions, palette, fonts, logos, and tagged references instead of asking the AI to infer them.
Steps
- Open or create a Brain item and choose the specific type before filling it; type controls the available fields and the sockets a Data Record node can expose.
- Fill the typed fields that downstream work must address directly. For a Persona, include voice, strategy, audiences, channel voice, cadence, examples, and explicit writing rules when the evidence supports them.
- Under References, choose Add link, From assets, or Upload files. Give each row a clear title and lowercase purpose/channel tags such as website, video, social, print, ad, linkedin, or instagram.
- Open Notes and details to record what to take from a source, review extracted document text or its style description, and reorder references by importance.
- Turn a reference off when it should be kept but not used; remove it only when the item should no longer retain that source.
- In Flowgen chat, type @ and choose a row from Brain. Send the message; Flowgen reads the record again at send time and supplies its current digest to the assistant.
- To ingest material with AI, attach every relevant file or supply links and ask to save it as a Brand, Persona, Character, Product, Place, Skill, or another specific Brain type.
- Review the result’s fields_empty and reference warnings. Answer only the important gaps, then let the assistant update the same item with merge_fields and append_references so earlier content is preserved.
- For project-wide brand grounding, open project settings → Brand and link the intended Brain Brand item. Review the copied logo, palette, font, and identity values before generating.
- Use list_records to discover exact items, search_records to find passages, and read_record before editing an existing item. Use write_record only when the request authorizes a save or update.
Reference controls
- Tag chips filter a large reference set.
- Notes describe what to take from one reference; an extracted-text preview shows what was read from an uploaded document.
- Move up/down changes order. In use includes the source in prompts; Kept, but not used retains it without applying it.
- The current ingest path reads at most 20 reference inputs in one write and reports skipped, missing, or unresolved rows.
What Flowgen chat sends
- The @ picker is available in the Flowgen chat composer, not arbitrary node prompt fields.
- No connection is created. The visible token is flattened to @Title and the record id, useful typed fields, and up to 3,000 body characters travel in a separate data block.
- A missing or deleted item is reported explicitly instead of silently replaced with the label.
- The picker refreshes stale rows after roughly 30 seconds; the message still performs a fresh read at send time.
Safe AI writes
- type_id chooses the real schema; unknown fields_json keys are rejected because they would not map to sockets.
- An update archives the previous content before overwrite.
- references_json accepts a URL or the exact name of a file attached to that message, plus title, tags, and notes.
- merge_fields and append_references are required for additive follow-up writes. save_in can target org, team, folder, or project.
Brand and Persona depth
- Brand supports raster and SVG logos, favicon, mission, audience, tone, emotions, palette/font roles, and separate web/video CSS, HTML, and motion fields.
- Persona supports writing context, channel voice, cadence, example rows, and weighted writing-rule rows in addition to voice, strategy, audiences, keywords, and discovery tags.
- Only facts supported by the item or its references should be filled; an empty field is preferable to an invented brand or persona claim.
Tips
- Use tags by destination—not vague aesthetics—so the assistant can choose website references for a site and video references for a film.
- Mention a record for one conversation turn; wire a Data Record node when a reusable typed field must participate in the graph.
Limitations and important notes
- Flowgen @ mentions are currently chat-only and do not create wires.
- Mention digests abbreviate the body; the assistant should call read_record when it needs the full item.
- Linking a Brand copies values into project settings at link/sync time; review the link after materially changing the Brain item.
Troubleshooting
A Brain mention is red or the assistant says it could not load the record.
The item may have been deleted, moved, or made inaccessible. Reopen @ and choose the current accessible item, or use list_records.
A follow-up erased earlier fields or references.
Update the same record_id with merge_fields:true and append_references:true. Do not create a second record for the same subject.
A field was refused by write_record.
Read the item type and use only that type’s current field keys. Put non-mapping material under a clear heading in the markdown body.