| name | prior |
| description | Assess the state of the art on a research question, grounded in primary literature. Use when the user asks "has this been done?", "what's the evidence on X?", "find the contradictions in the literature on X", or wants a cited, honest answer (including an honest "no") instead of an ungrounded summary. Builds a queryable atlas of claims from OpenAlex + arXiv and reasons over it. |
Prior — literature-assessment skill
Prior turns primary literature into a queryable atlas of claims and answers
questions from it, with citations and an honest abstention path. You (the agent)
drive it through its CLI tools — that is what makes this an agent capability
rather than a one-shot prompt.
When to use
- "Has anyone shown X?" / "What's the state of the evidence on X?"
- "What are the contradictions in the literature on X?"
- The user wants a grounded, cited answer and an honest "no" when the
literature doesn't support the claim — not a confident guess.
Tools (run from the repo root)
Set a backend once: export ANTHROPIC_API_KEY=... or
export PRIOR_LLM_BACKEND=claude-code (runs on the Claude Code login, no API key).
| Step | Command | What it does |
|---|
| build | python -m prior.cli build "<topic>" --max-papers 15 | ingest → read → map → data/atlas/atlas.json |
| ask | python -m prior.cli ask "<question>" | forward: supporting / contradicting / open, all cited; honest not_found |
| origin | python -m prior.cli origin "<concept>" | backward: trace a concept to its origin paper |
| contributions | python -m prior.cli contributions | each primary paper's self-declared, standalone contributions (full text) |
| view | python -m prior.cli view [--contributions] | render the atlas (or contributions) to an interactive HTML graph |
| info | python -m prior.cli info | one-line atlas summary |
(If installed with pip install -e ., prior <cmd> works too.)
Workflow for "assess the state of the art on a question"
- Pick a focused topic string from the user's question.
build the atlas for that topic (skip if data/atlas/atlas.json already
covers it — check with info).
ask the user's exact question. Report the verdict, the supporting and
contradicting evidence with their citations, and any open questions.
- If the verdict is
not_found, relay the honest "no": the closest work and the
gap — do not fill the gap with ungrounded knowledge.
- Offer
view for an interactive graph, or origin to trace where an idea began.
Rules
- Never answer from your own parametric knowledge when Prior returns claims —
ground every statement in cited claims from the atlas.
- Preserve Prior's abstention: a graceful "no" is a correct answer, not a failure.
- One build per topic is enough; the atlas is cached and reusable.