save-to-brain
Every durable artifact goes through this atom. Never keep the only copy in chat or on the local disk, and never tell the marketer about workspaces, documents or imports — say "saved to your brand" or what failed in plain words.
Local mode (no GooseWorks, the marketing-guide skill §5) is the one exception. It calls no gooseworks: tool, and brand_id and coworker_agent_id are not needed:
- Host with files: with the host's own file tools, read
gooseworks-local/<brand-slug>/<path>if it exists, merge by step 1's rules (never-overwrite included), write it, and read it back. Skip steps 2–6. Say "saved in this folder", never "saved to your brand". - Hosted chat without files: nothing can be saved. Give the whole file in chat under its path and say "This isn't saved anywhere: paste it back next time and I'll pick up from it." Skip steps 1–6.
Inputs (the calling playbook supplies them)
| Input | Rule |
|---|---|
brand_id, brand_slug, coworker_agent_id | From the gooseworks:brand_read list answer. |
path | The playbook's own path in the marketing-guide's §4 table (research/market/…, research/competitors/…). Dates YYYY-MM-DD. |
artifact | The artifact name, used in the title and tags (for example Voice-of-customer bank, voc-bank). |
title | <Artifact> — <Brand> — <YYYY-MM-DD>; records with numbered versions add v<N>, and a suffixed file adds the same suffix (step 1). |
tags | [<playbook>, <artifact-slug>], plus any extra key the caller needs to tell copies apart (for example rival:<slug>). All lower-case. The first tag is the document's creator: normally the calling playbook; a shared atom that writes on behalf of another owner puts that owner first (voice-card, claims-gate). |
versioned | supersede — a record that is replaced by its newest version (a bank, a profile, a landscape, an angle bank, a watchlist). keep-history — a dated period artifact (a weekly report, a digest, a dated read); older ones are history and are never deleted. |
import | yes (default) or no — no for working files that must stay out of brand knowledge (unsent reply drafts, raw item lists, pasted private material). |
Brand Brain tools
The Brand Brain tools are brand_knowledge_save, brand_knowledge_search and brand_knowledge_read.
brand_knowledge_save has no document or report target. Its targets are learning, fact, claim, audience, competitor, product, kit fields, operating-plan sections and proposals. So documents keep the route below (workspace file → media_upload → knowledge_import) on every server. What changes when brand_knowledge_* tools are listed:
- Read-back (step 6) uses
gooseworks:brand_knowledge_searchwithsource_types: ["document"], andgooseworks:brand_knowledge_readon the hit to read it in full. Otherwise usegooseworks:knowledge_search. - Status. An imported document is found: never call it confirmed, and it is never an approved claim (the marketing-guide §1.4).
- Indexing lags.
brand_knowledge_searchcan answer "Still indexing: N of M". That means "saved, not yet searchable", not a failure. Do not re-import.
Workspace scope
The scope in the marketing-guide skill §2, on every file call and on the upload's workspace source. With no coworker, tell the human once: "This brand has no coworker yet, so I saved it in your general workspace under its name."
Steps
- Read before you write (the marketing-guide §2).
gooseworks:file_listthe folder, thengooseworks:file_readthe file if it exists. Merge: keep every existing row (ids are stable), add new rows, update changed rows in place, never silently drop a row — mark itstatus: retired (<reason>, <date>). Keep the file's header lines current. In another playbook's file, only add lines under your own request lines. Never overwrite another copy. Akeep-historyfile this run did not write already at the path (another session, an earlier run today) stays: suffix this one's file name — the job slug where the caller has one, then-2,-3— and its title the same way (— <job slug>,(2),(3)), because step 5 deletes a copy with exactly this title. Asupersederecord is meant to be replaced and keeps its name. - Write the file with
gooseworks:file_write(same scope). Ifimport: no, stop here. - Register it as a brand document with
gooseworks:media_upload— scope brand, kind document, source workspace (this path, this scope),title,tags. The inline bytes source does not take text documents; always go through the workspace. - Import it with
gooseworks:knowledge_import(the new media id), thengooseworks:job_get(kindknowledge_ingest, withbrand_id) until it is ready. If any answer shows a price, stop and ask the human before going on. - Retire old copies — only after step 4 is ready. List earlier copies with
gooseworks:media_list(scope brand, kind document) and keep only rows that carry every tag the caller passed (including keys such asrival:<slug>) and whose first tag (the creator) is the caller's first tag — never the copy just imported, never a document whose first tag names someone else. Then:versioned: supersede— delete (gooseworks:media_updatedelete) those whose title matches this title apart from the date (and version).keep-history— delete only a copy with exactly this title (a re-run of the same period, written to the same file); other periods, and same-day copies with a suffixed title, are history and are never deleted. Deleting the upload also removes its knowledge, so search stops returning stale rows. The workspace file is never deleted.
- Verify. Search with a query that should hit the new content, using
source_types: ["document"]:gooseworks:brand_knowledge_searchwhenbrand_knowledge_*tools are listed, otherwisegooseworks:knowledge_search. If it does not come back (or the brain says it is still indexing), say "saved, not yet searchable". Never claim it is in the brain when it is not.
If a tool is missing
Call gooseworks:tools_load once with the group (files, media, brand). If it still fails, give the artifact to the human in chat, say plainly that it is not saved to their brand, and record the failure in the run's hand-over.
