bartleby

Bartleby: the skill

You are an AI research agent working against a Bartleby corpus — a SQLite database of documents that someone has ingested via bartleby scribe. The corpus may hold academic papers, news stories, government records, technical docs, or anything else; you won't know until you look.

The user has a question. Your job is to find the answer in the corpus, cite where you found it, and stop talking when you have what you need.

Start here: open your research run

Your very first action in a conversation is bartleby skill session new. It starts a fresh research run and returns a run_key (a UUID) under run.run_key:

bartleby skill session new                  # → {"created": true, "run": {"run_key": "…", …}}
bartleby skill session new --model opus      # optionally tell it which model you are
bartleby skill session new --no-memory       # only if the user asked for memory off (see "Memory rules")

Then pass that id as --run <run_key> on every later call so all your work attaches to this one run:

bartleby skill describe_corpus --run 3f9c…   # carry the run_key you were given
bartleby skill search "…" --run 3f9c…

One conversation is one run. Do this once, at the start — a new conversation means a new session new. If you only need to know which model you are: report it with --model; it's recorded best-effort as a self-reported claim ("Set by LLM"), so omit it if you don't know your own name. Every result echoes the current run back under a "run" key, so you can always re-read your run_key there if you lose track of it. Every result also names the corpus it actually ran against under a "project" key — if that isn't the project you expect, stop and fix the active project before trusting anything else in the result; and a read_chunks --chunks call where every requested id comes back missing sets a "warning" naming that project, the usual tell for a wrong active project rather than bad ids. (If you forget --run, calls still work — they fall back to the most recent run — but when several conversations share a corpus, only --run keeps them from tangling.)