researcher

Researcher

A tireless research assistant for the entire academic pipeline, for Claude Code, Cowork, and OpenAI Codex. It does not sleep, so you finally can.

Built by a doctoral student, for doctoral students (and anyone else drowning in research and manuscripts at 3 AM).

The academic writing workflow is too long, too fragmented, and too lonely. You have a research question on Monday, a literature review due on Friday, a methods section that won't write itself, reviewer comments that question your existence, and a formatting guide that was clearly written by someone who hates you.

Researcher is the assistant that absolutely does not understand what you're going through (it is, after all, just an AI that has never once cried in the library stacks), but it will still help you claw your way out.

But let's be real: this tool is not going to do your research for you. It's not going to write your paper while you sleep (though, and this is rather the entire point, you might finally get some sleep). It won't fabricate citations, invent data, or pretend to understand your domain better than you do. What it will do is handle the brutal, fragmented, soul-crushing logistics of academic publishing so you can focus on the part that actually matters: your ideas.

Think of it as a very capable assistant that never sleeps, never forgets a citation format, never sighs when you restructure your introduction for the ninth time, and, crucially, does not need you to stay awake to keep it company.

Full documentation and a step-by-step cookbook live at sokolmarek.github.io/researcher. Prefer to read the source? Worked examples with real, DOI-verified output (and rendered figures) are in examples/.

What This Is

A Claude Code / Cowork plugin with 35 specialized skills, 9 agents, and 15 slash commands covering the full research pipeline:

Brainstorm -> Literature Search -> Research Gaps -> Experiment Design
    -> Statistical Planning -> Implementation -> Manuscript Drafting
    -> Visualization -> Citation Management -> Peer Review -> Revision
    -> Journal Selection -> Formatting -> Submission

It also runs on OpenAI Codex: all 35 skills install with one script (python scripts/install-codex-skills.py), because Codex reads the same open agent-skills format. See OpenAI Codex under Installation for what carries over.

Underneath the skills sits the evidence kernel (core/, new in 0.3.0): a deterministic Python package that queries 8 scholarly indexes directly, deduplicates the results, and verifies every reference on four independent axes. It is optional, it ships with its benchmarks, and it is designed so that the one thing it must never do (tell you a real citation of yours is fabricated) it measurably does not do.

On top of the kernel sits the evidence-lineage compiler (new in 0.4.0): every claim and every number in a draft traces to a qualified span in an external source or to an internal experiment run, and researcher compile fails the gate on anything that does not. It names six defect classes (orphan claims, altered numbers, stale evidence, qualifier mismatches, retractions, and artifact-code drift), and researcher passport exports the whole lineage as an RO-Crate or W3C PROV record. A source that errors during a re-check is inconclusive, never a defect, so a downed index can never fail your build.

On top of the lineage graph sits the systematic-review vertical (new in 0.5.0): a workflow you could defend to a methodologist. It locks a protocol before any screening starts (a post-lock edit without a recorded amendment is caught as a hash mismatch), runs dual independent screening with genuinely blinded adjudication (the adjudicator sees only the record and the eligibility profile, never the other screener's vote, and that blinding is verified end to end through the real CLI), organizes RoB 2 and GRADE appraisal worksheets that a human completes, and keeps reviews living as new records arrive. The whole PRISMA 2020 flow and checklist are DERIVED by aggregating the event ledger, not stored: delete one screening event and the derived flow changes, which is the proof.