research-kit

research-kit

Spec-Driven Development for research papers, as slash commands for your AI coding agent — Claude Code, Codex CLI, or GitHub Copilot CLI. Instead of one giant "write my paper" prompt, you get a pipeline of small commands: each reads what came before, takes your input, and writes one artifact under ./.research/. Just Markdown commands + templates and a default research constitution — no Python CLI, no build step.

Pipeline

constitution → proposal → relatedwork → feasibility → tasks → (design + eval + paper, in parallel) → analyze → review

After feasibility, tasks fans out into three parallel lanes that co-evolve: design builds the system (code), eval evaluates it, paper writes it up. analyze is the sync checker that catches drift between them and tells you what to re-run. The design lane is paper-type aware - heavy for systems/defense, skipped for measurement / SoK. Plus auxiliary commands: rebuttal (post-submission) and ae (artifact evaluation). Run any subset, re-run any stage as your work evolves; commands only touch their own artifacts and never overwrite silently.

📐 Workflow diagram + per-command inputs/outputs →

Quickstart

Claude Code — plugin (recommended, no script):

/plugin marketplace add jiancui-research/research-kit
/plugin install research-kit@research-kit

Plugin stages are namespaced, e.g. /research-kit:research.proposal …; update later with /plugin marketplace update.

GitHub Copilot CLI — plugin (no script):

copilot plugin marketplace add jiancui-research/research-kit
copilot plugin install research-kit@research-kit

Copilot reads the same .claude-plugin bundle directly, exposing the namespaced /research-kit:research.* stages; update later with copilot plugin update research-kit.

Any agent — script:

./install.sh            # Claude Code (default). Also: --codex, --copilot, --all

Then, in your paper repo:

/research.init                       # once per repo: copy templates into .research/
/research.constitution <focus>       # optional: set writing voice + venue
/research.proposal <your raw idea>   # pipeline entry
/research.relatedwork
/research.feasibility
/research.tasks                      # writes three plans: design, eval, paper
/research.design                     # build-papers only: implement the system into ./design/
/research.eval                 # evaluate the build; runs parallel with paper, synced via claims.md
/research.paper
/research.analyze                    # also a "sync" check: what drifted, what to re-run
/research.review

Each command writes its result into ./.research/ and suggests the next one. (Plugin installs prefix every command with research-kit:.)

Commands

Command	What it does
/research.init	Copy the bundled templates into this paper repo's .research/templates/ (run once per repo, after install.sh).
/research.constitution	Establish or update the research constitution: quality principles, writing voice, and venue norms.
/research.proposal	Pipeline entry: turn a raw idea into a sharp, falsifiable proposal — NABC, the gap, measurable contributions, testable RQs, venue and paper type.
/research.relatedwork	Survey prior work into related-work.md and sharpen the proposal's gap and positioning.
/research.feasibility	De-risk the central result with a quick probe and return a GO / NO-GO / PIVOT verdict before you invest in the full build.
/research.tasks	Produce three paper-type-aware plans: the design/build plan, the eval plan, and the paper task list (READY vs blocked-on-claim).
/research.design	Build lane (build-papers only): implement the system from tasks/design.md into actual code in the project's ./design/ folder. Skipped for measurement / SoK.
/research.eval	Run the eval tasks that evaluate the built system and keep the claim-evidence matrix current.
/research.paper	Human-led writing: outline a section or critique your draft, every claim traceable to the evidence matrix; System Design sourced from tasks/design.md.
/research.analyze	Read-only cross-artifact audit and the sync checker across the design/eval/paper lanes: flags drift and names the exact re-run. Outputs a prioritized gap report.
/research.review	Simulate a reviewer panel reading only the paper: mock reviews + scores, plus a suggested fix command per finding; you route them and loop until clean.
/research.rebuttal	Draft a prioritized, evidence-backed rebuttal to reviewer comments, fitted to the venue word limit.
/research.ae	Prepare an artifact-evaluation submission: reproducibility checklist, artifact README, badge plan, archival link.

Supported agents

The same pipeline installs for three agents; pick one or more (--all for every one; default is Claude Code).