research-workflow
End-to-end research workflow for Claude Code — an adversarial validation Gate across the
full research lifecycle. 面向 Claude Code 的端到端科研工作流 —— 覆盖全研究生命周期的对抗性校验 Gate。
English · 中文
English
research-workflow turns Claude Code into a research companion that walks a project from a
blank page to a submitted paper — literature exploration, problem formulation, method
design, evaluation design, experiments, writing, simulated peer review, and rebuttal — with
an adversarial validation Gate at every phase transition.
Core principle — finding a fatal flaw one phase earlier costs 10× less to fix.
What it is / who it's for
It is for anyone running a research project with Claude Code — students, researchers,
engineers writing systems / empirical / benchmark / survey / theory papers. Domain- and
venue-agnostic. It does not write your paper for you; it imposes a disciplined process —
typed artifacts, evidence-cited claims, pre-mortem reviews — so flaws surface early and
claims never outrun their evidence.
What it gives you:
- A phased lifecycle (Phase 0–9), each step with one clear artifact to produce.
- An Adversarial Gate at every transition — multi-persona reviewers hunt for fatal
flaws before they become unfixable.
- Anti-fabrication contracts — every citation is verified, every quantitative claim
carries a source.
- Typed project state — progress, risks, decisions, artifacts live in structured
files, not in Claude's conversation memory.
- Learning — the skill records what worked and what failed and gets sharper with use.
Install
/plugin marketplace add dianxiang-sun/research-workflow
/plugin install research-workflow@research-workflow
Restart Claude Code if prompted. The skill is then invocable as /research-workflow:research.
Quick start (2 minutes)
/research-workflow:research init # answer 5 questions: name, domain, paper type, venue, deadline
/research-workflow:research explore # systematic literature scan + gap analysis
/research-workflow:research status # phase, risks, next actions, time to deadline
init writes three files into your project's .claude/ directory (see Project files).
After each phase the skill offers an Adversarial Gate — accept it, address its findings,
then move on. That is the whole loop.
The research lifecycle
| Phase | Mode | Produces |
|---|
| 0 | explore | Literature landscape, gap analysis, a research direction |
| 1 | foundation | Deep reading, motivation, research questions |
| 2 | design | Method / framework design, novelty stress-test |
| 2.5 | pilot | Feasibility pilot on 5–10 examples |
| 3 | eval-design | Dataset, metrics, baselines, ablations — the highest-ROI gate |
| 4 | implement | Engineering implementation (systems / tool papers) |
| 5 | experiment | Experiment execution, monitoring, honest analysis |
| 6 | write | Paper writing with claim provenance |
| 7 | review | Multi-persona simulated peer review + submission prep |
| 8 | rebuttal | Reviewer response, camera-ready, or venue pivot |
| 9 | present | Talk / poster / demo preparation |
Note: Phases 0-8 are the tracked main research sequence in research-state.yaml. Phase 9 (present) is a post-acceptance add-on; run /research-workflow:research present after acceptance for talk, poster, or demo preparation.
Paper types (chosen at init) decide which phases are active:
| Type | Active phases |
|---|
systems | 0–8 (all) |
empirical | 0–3, 5–8 — study design replaces implementation |
benchmark | 0–3, 5–8 — evaluate the dataset / benchmark itself |
survey | 0–1, survey-methodology, 6–8 |
theory | 0–2, theory-proofs, 6–8 |
Phases are a guide, not a prison — work adjacent phases together, skip ahead to test an
idea, or use any mode at any time.
Modes
Invoke as /research-workflow:research <mode> [args].
Lifecycle: init · status · explore · foundation · design · pilot ·
eval-design · implement · experiment · write [§N] · review · rebuttal · present
Cross-cutting — usable at any time:
gate [N] — run the Adversarial Gate (pre-mortem) for phase N (default: current).rollback N — structured return to phase N with impact analysis.risk — view / add / update the Risk Registry.cost — bottom-up cost / time estimate with failure-retry multipliers.novelty-watch — scan recent literature for competing work.evolve — periodic self-improvement: promote learned rules, update protocols.mine-patterns — extract writing patterns from past papers for style consistency.advisor-prep — prepare an advisor meeting: progress, open decisions, risk highlights.
