Open source · academic use
Research skills for Claude Code and OpenAI Codex, covering the social-science pipeline from a first idea to a submission-ready manuscript — with verification gates that decide what survives, and data safety enforced by the machine rather than by good intentions.
Each skill is a plain-markdown instruction folder your coding agent loads on demand. Ask for
/scholar-design and you get power analysis, a variable specification and a
pre-analysis plan; ask for /scholar-verify and a four-agent panel checks every
number in your manuscript against the output that produced it. Nothing is a black box — the
skills are text you can read, fork, and argue with.
This is not autonomous research. The suite is built to keep you in the loop:
run the skills individually, review what each one produces, and stay the author of the research
question, the argument, and every interpretation. The orchestrator it does ship,
scholar-auto-research, has a mandatory human-in-the-loop mode that pauses for your
approval between phases.
Deliberately not included in the open release: scholar-full-paper,
scholar-grant, scholar-teach and scholar-book — withheld
to discourage fully automated paper generation without meaningful researcher involvement.
Not every project needs every skill. Skip what does not apply; repeat what does —
/scholar-write → /scholar-verify → revise → repeat is the loop most
papers actually run.
/scholar-init builds the project layout, scans every raw file, and makes you resolve each one before an agent may read it./scholar-idea · /scholar-brainstorm · /scholar-lit-review · /scholar-hypothesis — from a vague topic to a formal, testable claim with the literature behind it./scholar-design · /scholar-causal · /scholar-data — identification strategy, power, variables, and a pre-analysis plan locked before you look./scholar-eda · /scholar-analyze · /scholar-compute · /scholar-annotate — publication-quality tables and figures, plus LLM measurement held to a κ ≥ 0.70 gate./scholar-write · /scholar-citation · /scholar-verify · /scholar-code-review — drafting, reference verification against CrossRef, and multi-agent review that catches results-changing defects./scholar-journal · /scholar-open · /scholar-replication · /scholar-ethics · /scholar-respond — formatting, a replication package, an AI-use disclosure, and a simulated review before the real one.Thirty-five research skills plus one utility, backed by twenty specialist agents.
Reading a data file through a coding agent transmits its contents to an API. That is harmless for a public CSV and a data-use-agreement violation for restricted records. The suite refuses to leave that to memory.
Five explicit statuses — CLEARED, LOCAL_MODE, ANONYMIZED, OVERRIDE, HALTED — and a local-mode contract that computes on data without ever printing rows.
/scholar-init runs a local PII and HIPAA scan on every file entering the project, then walks you through each flagged one until it has an explicit decision.
A PreToolUse hook intercepts every read, in Claude Code and Codex alike, and refuses anything still unreviewed. A lockdown level adds an OS-level sandbox deny — a wall, not a request.
The suite has a handbook, a workshop, and a peer-reviewed paper behind it.
A tutorial you work through at your own pace — from “Claude is not yet installed” to a complete, verified paper draft on real CFPS data. Every command and finding came from an actual run.
Six parts as a walkable book: Day 0 is the whole thing in one sitting, then the model, the agent, the harness and the pipeline, one layer at a time. Passcode-gated for workshop participants.
“Can AI Agents with Skills Replace or Augment Social Scientists?” — Chinese Sociological Review (2026). The argument behind the suite: where delegation works, and where it quietly fails.
Eleven minutes, condensed from a multi-hour session: project setup and safety scan, literature review, analysis, a multi-agent code review that catches real defects, then drafting and verification.