AI Research Workflows Plugin

A skills-first Claude Code plugin for Research Software Engineers and researchers. It covers the full arc of research-software work — from understanding code and surveying prior art through planning, experimentation, implementation, validation, reproducibility, and handoff. The plugin is interactive when you want collaboration (ask questions, propose options, gate on approval) and direct when you don't (act on intent, narrate briefly, stop only when blocked). Skills auto-trigger from natural language — no slash command required.

Version: 0.2.0

Background & Motivation

Research software is written under pressures that general-purpose development rarely faces all at once: it is exploratory and fast-changing, its correctness is often the actual research result, it must be reproducible by others (and by your future self), and it is frequently handed between students, collaborators, and sessions. AI coding assistants are excellent at writing code — but used ad hoc, they tend to skip the very rigor research demands. They jump into edits before understanding the existing code or the prior art, lose the reasoning behind decisions when a session ends, and rarely stop to capture provenance or to verify that a result actually holds.

This plugin exists to put that structure back. It encodes the research-software lifecycle — understand, survey prior art, plan, experiment, implement, validate, reproduce, harden, hand off — as a set of composable skills that trigger from plain natural language. Each phase produces a durable, cross-linked Markdown document under docs/rse/specs/, so the context and the decisions survive beyond a single conversation and transfer cleanly to the next person or session. The aim is to keep an AI assistant honest about the parts of research engineering that are easiest to skip and most expensive to lose.

Principles

• Skills, not ceremony. Describe the task in natural language and the right skill loads; slash commands are just shortcuts. You are never forced through a rigid pipeline — invoke a single skill or chain the whole arc.
• Durable context. Every phase writes a versionable artifact to docs/rse/specs/, committed to version control alongside the code it describes. Plans cite the research that informed them; implementations cite their plan; handoffs cite everything. Context is a file in your repo, not a chat log.
• Adaptive interaction. Collaborative when you want to think together, Direct when you just want the work done — chosen per request (see Interaction modes).
• Strategy over mechanics. The skills decide what and why, and defer the how of environments, tests, packaging, and docs to specialist plugins (see Cross-plugin deferral) — so this plugin stays language-agnostic and avoids duplicating their guidance.
• Research-grade rigor is first-class. Reproducibility (provenance capture) and hardening (correctness, regression, and stability checks) are dedicated skills, not afterthoughts.

It is built for the needs of Research Software Engineers and researchers. Examples lean toward scientific Python, but the strategy-level guidance applies to any language or runtime.

Contents

This plugin ships with 1 agent, 10 skills, and 9 commands.

• 1 Agent: Research Workflow Orchestrator
• 10 Skills: 7 workflow skills + 2 research-software skills + 1 meta-skill
• 9 Commands: thin wrappers that invoke the corresponding skill

Skills

Workflow skills (7)

Research-software skills (2)

Meta-skill (1)

Commands

Commands are thin wrappers — each invokes the corresponding skill. Skills also auto-trigger from natural language: describe the task and the right skill loads automatically.

Interaction modes

Every skill begins by choosing one of two modes:

• Collaborative — asks one focused question at a time, proposes options with a recommendation, and gates on your approval before producing artifacts. Use for brainstorming, planning, and exploratory work where your input shapes the outcome.
• Direct — acts on stated intent, narrates progress briefly, and stops only when genuinely blocked or before an irreversible action. Use when the task is clear and you want results without back-and-forth.

How the mode is chosen:

1. Explicit user phrasing wins: "brainstorm / walk me through / help me think" → Collaborative; "just do it / don't ask / go ahead" → Direct.
2. Otherwise infer: vague or exploratory phrasing, or required inputs missing → Collaborative; a specific directive with enough context → Direct.
3. Else use the skill's default lean (shown in the tables above).
4. Hard stops regardless of mode: destructive, irreversible, or outward-facing actions always require confirmation.

Full protocol: skills/using-research-workflows/references/interaction-modes.md

docs/rse/specs/ — workflow artifacts

All workflow documents are saved to docs/rse/specs/ in the project root, and are meant to be committed to version control alongside the code they describe — they are the durable decision record, not scratch files. The directory is created automatically on first use.

Documents from an earlier version of this plugin may live in a legacy .agents/ directory. Skills still read from .agents/ as a fallback when a document isn't found under docs/rse/specs/, but always write new documents to docs/rse/specs/.

Documents cross-link using relative paths (## References sections), creating a navigable graph of technical decisions: a plan links to the research docs that informed it; an implementation summary links back to the plan; a handoff document cites all relevant artifacts. Legacy prior-art-<slug>.md documents from earlier versions are still read when present.

Cross-plugin deferral

Some activities are intentionally delegated to specialist plugins:

Workflow patterns

Pattern 1: Full workflow (complex changes)

For significant features or architectural changes requiring thorough documentation:

/research [topic]           → Understand current code and/or external prior art
/plan [feature]             → Create phased implementation plan
/experiment [comparison]    → (Optional) Prototype and compare approaches
/iterate-plan [adjustments] → Refine plan based on findings
/implement [plan-file]      → Execute plan with per-phase verification
/validate [plan-file]       → Verify against success criteria
/reproduce [result]         → Capture provenance for key results

Pattern 2: Simple feature (known pattern)

For straightforward additions that follow existing patterns:

/research [existing patterns]
/plan [new feature]
/implement [plan-file]
/validate [plan-file]

Pattern 3: Rapid iteration (approach already clear)

When you have enough context and just need to plan and build:

/plan [feature]
/iterate-plan [scope adjustment]
/implement [plan-file]

Pattern 4: Research only (building context)

For understanding a codebase without immediate implementation intent:

/research [system A]
/research [system B]
[Documents in docs/rse/specs/ available for future planning]

Pattern 5: Research-first (prior art informs design)

For work where design decisions depend on what already exists in the field:

/research [domain / question]    → Understand existing code and/or survey external approaches
/plan [feature]                  → planning-implementations picks up research doc automatically

Installation

This plugin is part of the uw-ssec/rse-plugins collection. To install from the marketplace:

# Add the rse-plugins marketplace
claude plugin marketplace add uw-ssec/rse-plugins

# Install the plugin
claude plugin install ai-research-workflows@rse-plugins

To use directly from a local clone:

git clone https://github.com/uw-ssec/rse-plugins.git

Verify installation:

/help

Contributing

We welcome contributions to this plugin. You can:

• Add or improve skills — extend workflow guidance in skills/
• Enhance commands — update command prompts in commands/
• Improve the agent — adjust orchestration in agents/
• Report issues — open an issue on GitHub with the label ai-research-workflows
• Share examples — contribute real-world usage patterns

See the main repository CONTRIBUTING.md for detailed guidelines.

License

This plugin is part of the RSE Plugins project. See LICENSE file for details.

Authors

SSEC Engineering Team — https://escience.washington.edu/software-engineering/ssec/