--- name: resource-create description: Turn research — an agentbrain search, a session's findings, a set of links — into a durable resource directory under ~/resources with a seam, a report, full source documents, and a manifest. --- # Resource Create Gather a body of material on one question into `~/resources//` so it survives the session that found it. The manifest you write is what `resource-update` reads later; write it as a contract, not a log. The schema lives in `MANIFEST.md` beside this file. Read it before writing a manifest. ## Read the corpus The invocation names what to gather — "from our agentbrain search", "from everything we learned in this session", "from these links". Resolve it to concrete sources before doing anything else: - **An agentbrain search** — the queries already run in this conversation, plus several more angles on the same question. Search ranks by FTS5 over chunk text, so recall follows vocabulary: one phrasing silently misses documents whose words differ, however on-topic they are. Vary the angle — the author's framing, the opposing position, the failure mode, the tool rather than the idea — until new queries stop returning new documents. Keep the exact query strings; they go in the manifest and become the baseline every later update replays. - **A session** — the findings, comparisons, and judgements from the conversation. These have no source to re-read, so they become `session` sources and live in the report. - **Links or files** — fetch or copy each one. Most resources mix types. When the invocation is ambiguous about scope — one search or every search this session, the whole topic or the part just discussed — ask, once, with the two readings named. ## Gather - **Fetch documents in full.** `agentbrain get --document-id N --full --json`. Without `--full` the CLI head/tail truncates at 20,000 chars and says nothing about it. A 90K paper arrives as 20K and reads as complete. - Name each file `NN-slug.md`, numbered in the order a reader should meet them, not the order you found them. - Head each file with a provenance comment: ``. - Verify what you fetched. A document whose stored content is a product homepage, a paywall, or a nav shell is not a document — exclude it and record why. ## Write Four things, in this order: `docs/NN-*.md` — the full sources, unedited. `REPORT.md` — the survey. Direct hits first with the quotes that earn the ranking, then adjacent material, then what you checked and discarded and why, then gaps. Cite `document_id` and source URI throughout. This is where the reader goes to decide what to read; it can be long. `README.md` — the seam. Its job is to get a reader through a door, so it advertises rather than summarizes: something that makes the question feel worth an hour, a line on what the collection is, a few doors with a sentence each on why that one, and the full index below the fold. The doors are the design problem — pick them so each offers a different reason to enter, and the reader can tell from the outside which one is theirs. A seam that explains the documents has become the report. Length is the tell. `MANIFEST.json` — per `resources/MANIFEST.md`. Record the queries verbatim, every included document with its `document_id`/`ingested`/`chars`, every exclusion with its reason, and every gap with a `check` command. ## Close State the path, the file count, and the size. Name anything you excluded and anything still missing — a resource that hides its gaps is worse than a short one. Then say plainly that `resource-update ` will refresh it.