Prior Work
Literature reviews for the social sciences, built with an AI agent — and screened in VS Code.
日本語
Prior Work keeps a topic-based literature review in a Git workspace. An AI agent (Claude Code or
any agent that reads AGENTS.md) searches Semantic Scholar and OpenAlex, fills in a card for each
paper and writes the review; you make the decisions. This extension puts those decisions in the
sidebar: see every survey, screen the candidates, chase citations and check the report without
typing commands.
- Surveys in the sidebar. Each survey with what is left to screen, its next steps, and its
papers grouped as to screen, maybe, included and excluded.
- A page per paper. Title, authors, journal and its SSCI status, citations, warnings (a DOI
that points to another paper, a working-paper version, a possible book review) and the abstract,
with the survey's inclusion and exclusion criteria at the top. Decide with a click or a key —
I include, M maybe, X exclude (with a reason) — and it moves on to the next one.
- Search and snowball. Search in English (by relevance, boolean bulk search, or SSCI journals
only), then chase the references and citations of the included papers to find what keyword
search missed.
- Checks you can trust. Check the Survey lists author–year citations in the text that match
no registered paper, DOIs that point to other titles and unfilled cards in the Problems panel.
Export and View shows the report as it will be read, marked Draft until it is done.
- Follows your agent. When the agent records a decision or runs a search in the terminal, the
sidebar updates.
Getting started
- Install the extension. You need Python 3.10 or later and git.
- Run Prior Work: Create a Workspace… (or use the button in the Prior Work sidebar). Pick an
empty folder and the language of the reports (English or Japanese). The extension creates a
.venv there and installs the priorwork command from PyPI.
- Open .env and add a free Semantic Scholar API key
(searches without one are often rate-limited). Zotero is optional.
- New Survey…: the topic, an English slug, the depth (quick or full) and, if you like,
the research question.
- Search…, then Screen Candidates.
- Ask your agent to take it from there: the workspace carries
AGENTS.md and skills
(/survey-new, /survey-screen, /survey-snowball, /survey-extract, /survey-check).
Keep the workspace in a private GitHub repository: the reports and state files contain
abstracts, and .env holds your keys (it is kept out of Git).
Commands
| Command |
What it does |
| Create a Workspace… |
Set up a workspace (and its .venv) in a folder |
| New Survey… / Edit the Scope… |
Create a survey; change its research question, period, criteria, depth |
| Search… / Chase Citations / Add Papers by DOI… |
Find candidates |
| Screen Candidates |
Open the paper page at the first unscreened paper |
| Include / Maybe / Exclude… / Back to Candidate |
Also from the sidebar, for several papers at once |
| Get the Full Text |
From Zotero or an open-access PDF (for included papers) |
| Check the Survey |
Findings in the Problems panel |
| Export and View the Report |
The version for reading (HTML) |
| Diagnose the Setup |
API keys, the SSCI list, Git, connections |
| Update AGENTS.md and Skills / Update the Engine |
Keep the workspace in step with the engine |
| Move a lit Workspace to Prior Work |
For workspaces made with lit, Prior Work's former name |
Settings
| Setting |
Default |
|
priorwork.command |
(empty) |
The priorwork command. Empty: the workspace's .venv, then priorwork on PATH |
priorwork.python |
(empty) |
The Python used to create .venv. Empty: python3 (py -3 on Windows) |
priorwork.autoAdvance |
true |
Move on to the next unscreened paper after a decision |
What leaves your machine
Search queries and DOIs go to Semantic Scholar and OpenAlex; with Zotero set up, the extension
reads your Zotero library (read only). Nothing else is sent anywhere.
More
The command-line tool, the workspace layout and the rules the agent follows are described in the
project's README. The extension and the
priorwork command are released together, under one version number
(changelog).
| |
